Exposure, Capital, and Margin Analytics in DROP

v3.96 19 December 2018

1
Modeling Counterparty Credit Exposure in the Presence of Margin
Agreements

Abstract

1. Margin Based Credit Exposure Reduction: Margin agreements as a means of reducing

counterparty credit exposure.
2. Calculating MPoR Collateralized Exposures:
a. Collateralized Exposure and Margin Period of Risk
b. Semi-analytical Method for calculating collateralized EE
3. Analysis of Basel Exposure Methods: Analysis of Basel Shortcut method for Effective EPE.

Margin Agreements as a Means of Reducing Counterparty Credit Exposure

1. Definition of Counterparty Credit Risk: Counterparty Credit Risk is the risk that a
counterparty in an OTC derivative transaction will default prior to the expiration of the
contract and will be unable to make all contractual payments. Exchange-traded derivatives
bear no counterparty risk.
2. Counterparty vs. Lending Risk Difference: The primary feature that distinguishes counterpart
risk from lending risk is the uncertainty of exposure at any future time. For a loan, the
exposure at any future date is the outstanding balance, which is certain – not taking into

2
 account pre-payments. For a derivative, the exposure at any future date is the replacement
cost, which is determined by the market value at that date, and is, therefore, uncertain.
3. Bilateral Nature of the Counterparty Risk: Since derivative portfolio value can be both
positive and negative, counterparty risk is bilateral.
4. Spot/Forward Contract Market Value: Market value for counter 𝑖 with counterparty is known
only at the current date

𝑡=0

For any future date 𝑡 this value 𝑉𝑖 (𝑡) is uncertain and should be assumed random.
5. Replacement Cost at Counterparty Default: If a counterparty defaults at a time 𝜏 prior to the
contract maturity, the economic loss is equal to the replacement loss of the contract. If

𝑉𝑖 (𝜏) > 0

the dealer does not receive anything from the defaulted counterparty, but has to pay 𝑉𝑖 (𝜏) to
another counterparty to replace the contract. If

𝑉𝑖 (𝜏) < 0

the dealer receives 𝑉𝑖 (𝜏) from another counterparty, but has to forward this amount to the
defaulted counterparty.
6. Forward Exposure at Contract Level: Combining these two scenarios the contract-level
exposure 𝐸𝑖 (𝑡) at time 𝑡 is specified according to

𝐸𝑖 (𝑡) = max(𝑉𝑖 (𝑡), 0)

7. Definition of Counterparty Level Exposure: Counterparty level exposure at a future time 𝑡

can be defined as the loss experienced by the dealer if the counterparty defaults at time 𝑡
under the assumption of no recovery.

3
8. Unmitigated Counterpart Level Positive Exposure: If the counterparty risk is not mitigated in
any way, the counterparty level exposure equals the sum of contract-level exposure:

𝐸(𝑡) = ∑ 𝐸𝑖 (𝑡) = max(𝑉𝑖 (𝑡), 0)

𝑖

9. Counterparty Level Exposure under Netting: If there are netting agreements, derivatives with
positive values at the time of the default offset the ones with negative values within each
netting set 𝑁𝑆𝑘 so that the counterparty-level exposure is

𝐸(𝑡) = ∑ 𝐸𝑁𝑆𝑘 (𝑡) = ∑ max ( ∑ 𝑉𝑖 (𝑡) , 0)

𝑘 𝑘 𝑖∈𝑁𝑆𝑘

Each non-nettable trade represents a netting set.

10. Purpose of a Margin Agreement: Margin agreements allow for further reductions to
counterparty level exposure.
11. Bilateral Posting under Margin Agreements: Margin agreement is a legally binding contract
between two counterparties that requires one or both counterparties to post collateral under
certain conditions. A margin threshold is defined for one (unilateral agreement) or both
(bilateral agreement) counterparties. If the difference between the net portfolio value and the
already posted collateral exceeds the threshold, the counterparty must provide sufficient
collateral to cover this excess (subject to the minimum transfer amount).
12. Threshold Value in Margin Agreements: The threshold value depends primarily on the
margin value of the counterparty.
13. Netting Set Level Margin Agreement: Assuming that every margin agreement requires a
netting agreement, the exposure to the counterparty is

𝐸𝐶 (𝑡) = ∑ max ( ∑ 𝑉𝑖 (𝑡) − 𝐶𝑘 (𝑡), 0)

𝑘 𝑖∈𝑁𝑆𝑘

4
 where 𝐶𝑘 (𝑡) is the market value of the collateral for the netting set 𝑁𝑆𝑘 at time 𝑡. If the
netting set is not covered by a margin agreement, then

𝐶𝑘 (𝑡) = 0

14. Netting Set Collateralized Portfolio Value: To simplify the notations, consider a single
netting set

𝐸𝐶 (𝑡) = max(𝑉𝐶 (𝑡), 0)

where 𝑉𝐶 (𝑡) is the collateralized portfolio value at time 𝑡 given by

𝑉𝐶 (𝑡) = 𝑉(𝑡) − 𝐶(𝑡) = ∑ 𝑉𝑖 (𝑡) − 𝐶(𝑡)

𝑖

Collateralized Exposure and Margin Period of Risk

1. Collateral as Excess Portfolio Value: Collateral covers the excess portfolio value 𝑉(𝑡) over
the threshold 𝐻:

𝑉𝐶 (𝑡) = max(𝑉(𝑡) − 𝐻, 0) = − min(𝐻 − 𝑉(𝑡), 0)

2. Expression for the Collateralized Portfolio Value: Therefore, the collateralized portfolio
value is

𝑉𝐶 (𝑡) = 𝑉(𝑡) − 𝐶(𝑡) = min(𝑉(𝑡), 𝐻)

5
3. Floor/Ceiling of the Collateralized Exposure:

0 𝑉(𝑡) < 0
𝐸𝐶 (𝑡) = max(𝑉𝐶 (𝑡), 0) = {𝑉(𝑡) 0 < 𝑉(𝑡) < 𝐻
𝐻 𝑉(𝑡) > 𝐻

is limited from above and by zero from below.

4. Margin Period of Risk (MPoR): Even with a daily margin call frequency, there is a
significant period delay known as the margin period of risk (MPoR) between a margin call
that the counterparty does not respond to and the start of the default procedures. Margin calls
can be disputed, and it may take several days to realize that the counterparty is defaulting
rather than disputing the call. Further there is a grace period after the counterparty issues a
notice of default. During this grace period the dealer and/or the counterparty may still post
collateral.
5. Counterparty Variation Margin Posting Delay: Thus, the collateral available at time 𝑡 is
determined by the portfolio value at time 𝑡 − ∆𝑡.
6. Delay Dependence - Call Frequency and Product Liquidity: While ∆𝑡 is not known with
certainty, it is usually assumed to be a fixed number. The assume value of ∆𝑡 depends on the
margin call frequency and the trade liquidity.
7. MPoR Start - Collateral/Portfolio Values: Suppose that at time 𝑡 − ∆𝑡 the unilateral collateral
value is 𝐶(𝑡 − ∆𝑡) and the portfolio value is 𝑉(𝑡 − ∆𝑡).
8. Posted Counterparty Collateral at 𝑡: Then the amount ∆𝐶(𝑡) that should be posted at time 𝑡 is

∆𝐶(𝑡) = max(𝑉(𝑡 − ∆𝑡) − 𝐶(𝑡 − ∆𝑡) − 𝐻, −𝐶(𝑡 − ∆𝑡))

negative ∆𝐶(𝑡) means that the collateral will be returned to the counterparty.
9. Unilateral Counterparty Collateral at 𝑡: The unilateral counterparty collateral 𝐶(𝑡) available
at time 𝑡 is

6
 𝐶(𝑡) = 𝐶(𝑡 − ∆𝑡) + ∆𝐶(𝑡) = max(𝑉(𝑡 − ∆𝑡) − 𝐻, 0)

10. The Total Collateralized Portfolio Value: The collateralized portfolio value is

𝑉𝐶 (𝑡) = 𝑉(𝑡) − 𝐶(𝑡) = min(𝑉(𝑡), 𝐻 + ∆𝑉(𝑡))

where

∆𝑉(𝑡) = 𝑉(𝑡) − 𝑉(𝑡 − ∆𝑡)

11. Monte Carlo Primary Simulation Points: Suppose one has a set of primary simulation points
{𝑡𝑘 } for modeling non-collateralized exposure.
12. Monte-Carlo Look-back Points: For each

𝑡𝑘 > ∆𝑡

define a lookback point at time 𝑡𝑘 − ∆𝑡

13. Monte Carlo Primary Plus Lookback: The task is to simulate the non-collateralized portfolio
value along the path that includes both the primary and the lookback simulation times.
14. Collateral PLUS Collateralized/Uncollateralized Portfolio Values: Given 𝑉(𝑡𝑘−1 ) and
𝐶(𝑡𝑘−1 ) one calculates:
a. Uncollateralized Portfolio Value 𝑉(𝑡𝑘 − ∆𝑡) at the next lookback time 𝑡𝑘 − ∆𝑡
b. Uncollateralized Portfolio Value 𝑉(𝑡𝑘 ) at the next primary time 𝑡𝑘
c. Collateral at 𝑡𝑘 :

𝐶(𝑡𝑘 ) = max(𝑉(𝑡𝑘 − ∆𝑡) − 𝐻, 0)

d. Collateralized Portfolio Value at 𝑡𝑘 :

𝐶(𝑡𝑘 ) = 𝑉(𝑡𝑘 ) − 𝐶(𝑡𝑘 )

7
 e. Collateralized Exposure at 𝑡𝑘 :

𝐸𝐶 (𝑡𝑘 ) = max(𝑉𝐶 (𝑡𝑘 ), 0)

15. Simulating the Collateralized Portfolio Value: Collateralized threshold can go above the
threshold due to MPR and MTA.

Semi-Analytical Method for Collateralized EE

1. Portfolio Value at Primary Points: Assume that the simulation is only run for the primary
time points 𝑡 and the portfolio distribution has been obtained in the form of 𝑀 quantities
𝑉𝑗 (𝑡), where 𝑗 (from 1 to 𝑀) designates different scenarios.
2. Evaluating the Unconditional Portfolio Distribution: From the set {𝑉𝑗 (𝑡)} once can estimate
the unconditional expectation 𝜇(𝑡) and standard deviation 𝜎(𝑡) of the portfolio value, as well
as any other distributional parameter.
3. Collateralized EE at Lookback Points: Can the collateralized EE profile be estimated without
simulating the portfolio value at the lookback time points {𝑉𝑗 (𝑡 − ∆𝑡)}?
4. Collateralized EE Conditional on Path: Collateralized EE can be represented as

𝐸𝐸𝐶 (𝑡) = 𝔼[𝐸𝐸𝐶,𝑗 (𝑡)]

where 𝐸𝐸𝐶,𝑗 (𝑡) is the collateralized EE conditional on 𝑉𝐶,𝑗 (𝑡):

𝐸𝐸𝐶,𝑗 (𝑡) = 𝔼[max(𝑉𝐶,𝑗 (𝑡), 0) |𝑉𝑗 (𝑡)]

8
5. The Conditional Collateralized Portfolio Value: The collateralized portfolio value 𝑉𝐶,𝑗 (𝑡) is

𝑉𝐶,𝑗 (𝑡) = min (𝑉𝑗 (𝑡), 𝐻 + 𝑉𝑗 (𝑡) − 𝑉𝑗 (𝑡 − ∆𝑡))

6. Goal - Computing the Collateralized EE Analytically: If 𝐸𝐸𝐶,𝑗 (𝑡) can be computed

analytically, the unconditionally collateralized EE can be obtained as a simple average of
𝐸𝐸𝐶,𝑗 (𝑡) across all scenarios 𝑗
7. Assumption of Normal Portfolio Value: Assume that the portfolio value 𝑉(𝑡) at time 𝑡 is
normally distributed with mean 𝜇(𝑡) and standard deviation 𝜎(𝑡).
8. Brownian Bridge for Secondary Nodes: One can construct a Brownian Bridge from 𝑉(0) to
𝑉𝑗 (𝑡).
9. 𝑉𝑗 (𝑡 − ∆𝑡) Mean and Standard Deviation: Conditional on 𝑉𝑗 (𝑡), 𝑉𝑗 (𝑡 − ∆𝑡) has a normal
distribution with expectation

∆𝑡 𝑡 − ∆𝑡
𝛼𝑗 (𝑡) = 𝑉(0) + 𝑉𝑗 (𝑡)
𝑡 𝑡

and standard deviation

∆𝑡(𝑡 − ∆𝑡)
𝛽𝑗 (𝑡) = 𝜎(𝑡)√
𝑡2

10. Closed Form Conditional Collateralized EE: Conditional Collateralized EE can be obtained
in a closed form.
11. Piece-wise Constant Local Volatility: It is assumed that, conditional on 𝑉𝑗 (𝑡), the distribution
of 𝑉𝑗 (𝑡 − ∆𝑡) is normal, but 𝜎(𝑡) will be replaced by the local quantity 𝜎𝐿𝑂𝐶 (𝑡).
12. Portfolio Value Monotonically Increasing with 𝑍: The portfolio value 𝑉(𝑡) at time 𝑡 is
described using

9
 𝑉(𝑡) = 𝜗(𝑡, 𝑍)

where 𝜗(𝑡, 𝑍) is a monotonically increasing function of the standard normal random variable
𝑍.
13. The Equivalent Normal Portfolio Process: A normal equivalent portfolio process is defined
as

𝑊(𝑡) = 𝜔(𝑡, 𝑍) = 𝜇(𝑡) + 𝜎(𝑡)𝑍

14. Density Scaling to determine 𝜎𝐿𝑂𝐶 (𝑡): To obtain 𝜎𝐿𝑂𝐶 (𝑡), 𝜎(𝑡) will be scaled by the
probability densities of 𝑊(𝑡) and 𝑉(𝑡).
15. Standard Deviation Scaled Probability Density: The probability density of the quantity 𝑋 is
denoted via 𝑓𝑋 (∙) and the standard deviation is scaled according to

𝑓𝑊(𝑡) (𝜔(𝑡, 𝑍))

𝜎𝐿𝑂𝐶 (𝑡) = 𝜎(𝑡)
𝑓𝑉(𝑡) (𝜗(𝑡, 𝑍))

16. Changing Variables from 𝑊/𝑉 to 𝑍: Changing the variables from 𝑉(𝑡) and 𝑊(𝑡) to 𝑍, one
gets

𝜙(𝑍)
𝑓𝑉(𝑡) (𝜗(𝑡, 𝑍)) =
𝜕𝜗(𝑡, 𝑍)/𝜕𝑍

𝜙(𝑍)
𝑓𝑊(𝑡) (𝜔(𝑡, 𝑍)) =
𝜎(𝑡)

17. Substitution to the Definition of 𝜎𝐿𝑂𝐶 (𝑡): Substituting to the definition of 𝜎𝐿𝑂𝐶 (𝑡) above
gives

𝜕𝜗(𝑡, 𝑍)
𝜎𝐿𝑂𝐶 (𝑡) =
𝜕𝑍

10
18. Estimating CDF - The Base Methodology: The values of 𝑍𝑗 corresponding to 𝑉𝑗 (𝑡) can be
obtained from

𝑍𝑗 = Φ−1 (𝐹𝑉(𝑡) (𝑉𝑗 (𝑡)))

19. Estimating the CDF - Sorting the Realizations: One sorts the array 𝑉𝑗 (𝑡) in increasing order
so that

𝑉[𝑗(𝑘)] (𝑡) = 𝑉𝑘,𝑆𝑂𝑅𝑇𝐸𝐷 (𝑡)

where 𝑗(𝑘) is the sorting index.

20. Estimating CDF - Piecewise Constant Jump: From the sorted array, one can build a
1
piecewise constant CDF that jumps by 𝑀 as 𝑉(𝑡) crosses any of the simulated values.

1 𝑘 − 1 1 𝑘 2𝑘 − 1 𝑘 − 0.5
𝐹𝑉(𝑡) (𝑉𝑗 (𝑡)) ≈ + = →
2 𝑀 2𝑀 2𝑀 𝑀

where 0.5 is the de-facto bias reducer.

21. Estimation of the Weiner Wanderer: Now one can obtain 𝑍𝑗 corresponding to 𝑉𝑗 (𝑡) as

2𝑘 − 1
𝑍[𝑗(𝑘)] = Φ−1 ( )
2𝑀

22. Estimating the Local Standard Deviation: Local standard deviation 𝜎𝐿𝑂𝐶,𝑗 (𝑡) can be
estimated as

𝑉[𝑗(𝑘+∆𝑘)] (𝑡) − 𝑉[𝑗(𝑘−∆𝑘)] (𝑡)

𝜎𝐿𝑂𝐶,[𝑗(𝑘)] (𝑡) ≡ 𝜎𝐿𝑂𝐶 (𝑡, 𝑍[𝑗(𝑘)] ) ≈
𝑍[𝑗(𝑘+∆𝑘)] − 𝑍[𝑗(𝑘−∆𝑘)]

11
23. Choice of the Different Amount ∆𝑘: The offset ∆𝑘 should not be too small (too much noise)
or too large (loss of locality). This range works apparently well (Pykhtin (2009)):

20 ≤ ∆𝑘 ≤ 0.05𝑀

24. The Brownian Bridge Mean and 𝜎: Similar to the above it is assumed that, conditional on
𝑉𝑗 (𝑡), 𝑉𝑗 (𝑡 − ∆𝑡) has a normal distribution with expectation

∆𝑡 𝑡 − ∆𝑡
𝛼𝑗 (𝑡) = 𝑉(0) + 𝑉𝑗 (𝑡)
𝑡 𝑡

and standard deviation

∆𝑡(𝑡 − ∆𝑡)
𝛽𝑗 (𝑡) = 𝜎(𝑡)√
𝑡2

25. The Collateralized Exposure Mean and 𝜎: The collateralized exposure depends on ∆𝑉𝑗 (𝑡),
which is also normal conditional on 𝑉𝑗 (𝑡) with the same standard deviation 𝛽𝑗 (𝑡) and
expectation 𝛼𝐶,𝑗 (𝑡) given by

∆𝑡
𝛼𝐶,𝑗 (𝑡) = 𝑉𝑗 (𝑡) − 𝛼𝑗 (𝑡) = [𝑉 (𝑡) − 𝑉(0)]
𝑡 𝑗

26. Collateralized EE Conditional on 𝑗: Collateralized EE conditional on scenario 𝑗 at time 𝑡 is

𝐸𝐸𝐶,𝑗 (𝑡) = 𝔼 [max (min (𝑉𝑗 (𝑡), 𝐻 + ∆𝑉𝑗 (𝑡)) , 0) |𝑉𝑗 (𝑡)]

27. Collateralized EE on Negative Exposure: 𝐸𝐸𝐶,𝑗 (𝑡) equals zero whenever

𝑉𝑗 (𝑡) > 0

12
 so that

𝐸𝐸𝐶,𝑗 (𝑡) = 𝕀𝑉𝑗(𝑡)>0 𝔼 [min (𝑉𝑗 (𝑡), 𝐻 + ∆𝑉𝑗 (𝑡)) |𝑉𝑗 (𝑡)]

28. Integral Form for Collateralized EE: Since ∆𝑉𝑗 (𝑡) has a normal distribution, one can write

+∞

𝐸𝐸𝐶,𝑗 (𝑡) = 𝕀𝑉𝑗 (𝑡)>0 ∫ min(𝑉𝑗 (𝑡), 𝐻 + 𝛼𝐶,𝑗 (𝑡) + 𝛽𝑗 (𝑡)𝑍) 𝜙(𝑍)𝑑𝑍
−∞
−𝑑1 +∞

= 𝕀𝑉𝑗 (𝑡𝑘 )>0 { ∫ [𝐻 + 𝛼𝐶,𝑗 (𝑡) + 𝛽𝑗 (𝑡)𝑍]𝜙(𝑍)𝑑𝑍 + 𝑉𝑗 (𝑡) ∫ 𝜙(𝑍)𝑑𝑍}

−𝑑2 −𝑑1

29. Conditional Collateralized EE Closed Form: Evaluating the integrals, one obtains

𝐸𝐸𝐶,𝑗 (𝑡) = 𝕀𝑉𝑗(𝑡𝑘)>0 {[𝐻 + 𝛼𝐶,𝑗 (𝑡)][Φ(𝑑2 ) − Φ(𝑑1)] + 𝛽𝑗 (𝑡)[𝜙(𝑑2 ) − 𝜙(𝑑1 )]

+ 𝑉𝑗 (𝑡)Φ(𝑑1 )}

where

𝐻 + 𝛼𝐶,𝑗 (𝑡) − 𝑉𝑗 (𝑡)

𝑑1 =
𝛽𝑗 (𝑡)

𝐻 + 𝛼𝐶,𝑗 (𝑡)
𝑑2 =
𝛽𝑗 (𝑡)

Analysis of Basel “Shortcut” Method for Collateralized Effective EPE

13
1. Basel 2 Exposure Capital Requirements: Basel 2 minimal capital requirements for the
counterparty risk are determined by wholesale exposure rules with exposure at default
obtained from expected exposure profile as follows.
2. Exposure at Default - Basel Variants:
a. Expected Exposure (EE) – Expected Exposure Profile (EE)
b. Expected Positive Exposure (EPE) –

1 𝑌𝑒𝑎𝑟

𝐸𝑃𝐸 = ∫ 𝐸𝐸(𝑡)𝑑𝑡
0

c. Effective EE -

𝐸𝑓𝑓𝑒𝑐𝑡𝑖𝑣𝑒 𝐸𝐸(𝑡𝑘 ) = max(𝐸𝐸(𝑡𝑘 ), 𝐸𝑓𝑓𝑒𝑐𝑡𝑖𝑣𝑒 𝐸𝐸(𝑡𝑘−1 ))

d. Effective EPE –

1 𝑌𝑒𝑎𝑟

𝐸𝑓𝑓𝑒𝑐𝑡𝑖𝑣𝑒 𝐸𝑃𝐸 = ∫ 𝐸𝑓𝑓𝑒𝑐𝑡𝑖𝑣𝑒 𝐸𝐸(𝑡)𝑑𝑡

0

e. Exposure at Default (EAD) –

𝐸𝐴𝐷 = 𝛼 × 𝐸𝑓𝑓𝑒𝑐𝑡𝑖𝑣𝑒 𝐸𝑃𝐸

3. Incorporating the Margin Agreement: For collateralized counterparties, the netting set level
Effective EPE must incorporate the effect of margin agreement.
4. Effective EPE using Internal Model of Collateral: Collateralized Effective EPE can be
calculated using an internal model of collateral.

14
5. Basel 2 Simple and Conservative Shortcut: Alternatively, dealers can use a simple and
conservative approximation to the effective EPE, and sets the effective EPE for a margined
counterparty equal to the lesser of:
a. The Threshold, if positive, under the margin agreement plus an add-on that reflects
the potential increase in exposure over the margin period of risk. The add-on is
computed as the expected increase in the netting set’s exposure beginning from the
current exposure of zero over the margin period of risk.
b. Effective EPE without a margin agreement.
6. Derivation of the “Shortcut” Method: The Basel “Shortcut” method can be obtained as
follows:

𝐸𝐸𝐶 (𝑡) = 𝔼[max(min(𝑉(𝑡), 𝐻 + ∆𝑉(𝑡)) , 0)] = 𝔼[min(𝐸(𝑡), 𝐻 + max(∆𝑉(𝑡), −𝐻))]

≤ 𝔼[min(𝐸(𝑡), 𝐻 + max(∆𝑉(𝑡), 0))] ≤ min(𝐸𝐸(𝑡), 𝐻 + 𝔼[max(∆𝑉(𝑡), 0)])
≈ min(𝐸𝐸(𝑡), 𝐻 + 𝔼[max(∆𝑉(∆𝑡), 0)]) ≡ 𝐸𝐸𝐶,𝐵𝑆𝑀 (𝑡)

7. Enhancing the Exposure Conservativeness: Time averaging adds more conservativeness:

𝑇
1
∫ 𝐸𝐸𝐶,𝐵𝑆𝑀 (𝑡)𝑑𝑡 ≤ min(𝐸𝑃𝐸, 𝐻 + 𝔼[max(∆𝑉(∆𝑡), 0)])
𝑇
0

Conclusion

1. Margin Agreements for Risk Mitigation: Margin agreements are important risk mitigation
tools that need to be modeled accurately.
2. Complete MC Doubles Simulation Time: Full Monte Carlo is the most flexible approach, but
requires simulating trade values at secondary time points, thus doubling the simulation time.

15
3. Semi-Analytical Approach Avoids That: Pykhtin (2009) has presented an accurate semi-
analytical approach of calculating the EE that avoids doubling of the simulation time.
4. Basel 2 Shortcuts are too Conservative: Basel 2 “Shortcut” method for Effective EPE has
sound theoretical grounds, but is too conservative.

References

• Pykhtin, M. (2009): Modeling Counterparty Credit Exposure in the Presence of Margin

Agreements

16
 Estimation of Margin Period of Risk

Abstract

1. Enhanced CSA Collateral Exposure Model: Andersen, Pykhtin (2017) describe a new
framework for collateral exposure modeling under an ISDA Master Agreement with Credit
Support Annex. The proposed model captures the legal and the operational aspects of default
in considerably greater detail than models currently used by most practitioners, while
remaining fully tractable and computationally feasible.
2. Legal Rights Exercise/Deferral Choices: Specifically, it considers the remedies and the
suspension rights available within these legal agreements; the firm’s policies of availing itself
of these rights; and the typical time it takes to exercise them in practice.
3. Significantly Higher Credit Exposure Revealed: The inclusion of these effects is shown to
produce a significantly higher credit exposure for representative portfolios compared to the
currently used models. The increase is especially pronounced when dynamic initial margin is
also present.

Introduction

1. Margin Period of Risk Overview: In modeling the exposure of collateralized positions, it is

well recognized that credit default cannot be treated as a one-time event. Rather the entire

17
 sequence of events following up to the default and beyond need to be considered, from the
last successful margin call in advance of the eventual default to the time when the amount of
loss becomes known – in the industry parlance, crystallized. These events unfold over a
period of time called the margin period of risk (MPoR).
2. Range of Model Applicability: To properly identify the exposure during the MPoR, a
detailed understanding of the contractual obligations is essential. In their paper, Andersen,
Pykhtin, and Sokol (2017) focus on collateralized exposures under bilateral trading
relationships governed by the ISDA Master Agreement (IMA) and its Credit Support Annex.
The IMA, by far, is the most common legal contract for bilateral over-the-counter (OTC)
derivatives trading, although other agreements are sometimes used (such as national forms of
agreements used in some jurisdictions for domestic trading). The analysis by Andersen,
Pykhtin, and Sokol (2017) is expected to apply to a broad class of contracts, although the
model assumptions should be re-examined to confirm that the key legal provisions remain
substantially the same as IMA.
3. Refinement for Legal/Operational Impact: It should be noted that the modeling of default
exposure and close-out risk arising from a non-zero MPoR has received a fair amount of
attention in the past (see, e.g., Gibson (2005), Pykhtin (2009, 2010), and Brigo, Capponi,
Pallavicini, and Papatheodorou (2011)), although most past analysis has been conducted
under very strong simplifying assumptions about the trade and the margin flows during the
MPoR. One exception is Bocker and Schroder (2011), which contains elements of a more
descriptive framework, including recognition of the role played by the cash flows close to the
default event. Andersen, Pykhtin, and Sokol (2017) use a more detailed framework for legal
and operational behavior to refine the classical models for collateralized exposure modeling.
4. Variation Margin Operational Timelines: This chapter is organized as follows. The
fundamentals of variation margin posting are first outlined, and the classical collateralized
exposure model is then presented. The full timelines of events likely to transpire during a
credit default are then discussed from both legal and operational perspectives. This sets the
stage for the proposal of a condensed representation of the timeline suitable for analytical and
numerical work. The resulting setup results in a more significantly nuanced and flexible
definition of the collateralized trading exposure. As fixing the actual model parameters (i.e.,

18
 calibrating the MPoR model) requires taking a stance on operational procedures and
corporate behavior, the next section discusses how such parametrizations may be done in
practice, for various levels of overall model prudence and counterparty types.
5. Numerical Computation of Collateralized Exposure: Subsequently, the model is fleshed out
in more detail, especially as it pertains to numerical implementations and quantitative
comparisons with the classical model. As a starting point, exposure models are formulated in
mathematical terms, and the key differences to the classical models are highlighted by means
of brute-force Monte-Carlo simulations. Computational techniques permitting efficient
model implementation are introduced subsequently, along with several test results.
Applications to portfolios with risk-based initial margins are briefly discussed, and
conclusions are finally drawn.

The Fundamentals of Variation Margin: Basic Definitions

1. Types of Margin: Initial/Variation: In bilateral OTC derivatives trading, it is common for

parties to require posting of collateral to mitigate excessive exposures. Although the initial
margin is discussed briefly in a later section, this section focuses primarily on the variation
margin (VM) as a form of collateral that is regularly re-adjusted based on the changing value
of the bilateral portfolio. The VM is calculated and settled in time according to a set of CSA
rules discussed a few sections down.
2. Dealer and Client VM Timelines: For concreteness, throughout this chapter, the exposure of
dealer 𝐷 to a client 𝐶 with whom 𝐷 engages in bilateral OTC trading under the IMA/CSA
legal framework is considered. 𝐶 is referred to as the defaulting party, and 𝐷 is the dealer or
the non-defaulting party. All present value and exposure amounts throughout this chapter
will be calculated from the viewpoint of 𝐷. Let the default-free market value to 𝐷 of the
securities portfolio at time 𝑡 be 𝑉(𝑡) and let 𝐴𝐷 (𝑡) and 𝐴𝐶 (𝑡) be the collateral support
amounts stipulated by the CSA to be posted to 𝐷 and 𝐶 respectively. In the absence of initial

19
 margin, it is virtually always the case that only one of 𝐴𝐷 or 𝐴𝐶 is positive, i.e., only one
party will be required to post margin at a given point in time.
3. Net Collateral and its Posting: Assuming that collateral is netted (rather than posted by both
parties in full and held in segregated accounts or by a third party), the total collateral amount
in 𝐷’s possession may be calculated as of time 𝑡 as

𝑐(𝑡) = 𝐴𝐶 (𝑡) − 𝐴𝐷 (𝑡)

Assuming also that the collateral may be treated as pari passu with the derivatives portfolio
itself for the purposes of bankruptcy claim, it is common to denote the positive part of the
difference 𝑉(𝑡) − 𝑐(𝑡) as the exposure 𝐸(𝑡):

𝐸(𝑡) = [𝑉(𝑡) − 𝑐(𝑡)]+

𝑐(𝑡) = 𝐴𝐶 (𝑡) − 𝐴𝐷 (𝑡)

where the notation

𝑥 + = max(𝑥, 0)

is used. Normally both the collateral and the portfolio would be treated together as a senior
unsecured claim of 𝐷 against the bankruptcy estate of 𝐶. There are several time lags and
practical complications that render the above exposure and collateral expressions an
imprecise measure, and they shall be substantially refined later on. In particular it is
emphasized that the collateral computed at time 𝑡 is generally not transferred to 𝐷 until
several days after 𝑡.
4. VM Designated to Track Portfolio Value: The type of VM encountered in the CSA is
typically designed to broadly track the value of the portfolio between the parties, thereby
ensuring that 𝐸(𝑡) in

20
 𝐸(𝑡) = [𝑉(𝑡) − 𝑐(𝑡)]+

𝑐(𝑡) = 𝐴𝐶 (𝑡) − 𝐴𝐷 (𝑡)

does not grow to be excessive. However, to avoid unnecessary operational expenses, it is

common to introduce language in the CSA to relax margin transfer requirements if the
amounts are sufficiently small.
5. Reducing the Collateral Posting Events: To that end the typical CSA language for collateral
calculations will stipulate:
a. Collateral posting requirements by each party, ℎ𝐷 and ℎ𝐶 , representing the minimum
amount of exposure before 𝐷 or 𝐶, respectively, is required to post collateral
b. A minimum transfer amount (MTA) establishing a minimum valid amount of a
margin call
c. Rounding, which rounds collateral movements to some reasonable unit (say $1,000).
6. Incorporating Thresholds into Collateral Expressions: Formally the effects of thresholds on
the stipulated collateral may be written as

𝐴𝐷 (𝑡) = [−𝑉(𝑡) − ℎ𝐷 ]+

𝐴𝐶 (𝑡) = [𝑉(𝑡) − ℎ𝐶 ]+

with the net stipulated credit support amount assigned to 𝐷 being

𝑐(𝑡) = 𝐴𝐶 (𝑡) − 𝐴𝐷 (𝑡)

as before. The actual availability of this amount is then subject to the path dependent effects
on collateral by MTA and rounding, of which the former has significant effect only for zero
or very small thresholds, and the latter is usually negligible. Both have been omitted in the
equation above.

21
7. Unilateral and Asymmetric Collateral Requirements: Most CSAs are bilateral in nature, but
unilateral CSAs do exist in which only one of the two parties is required to post collateral. A
CSA may be formally bilateral, but highly asymmetric, requiring both parties to post
collateral but with vastly different thresholds, e.g.

ℎ𝐷 = $20 𝑚𝑚

vs.

ℎ𝐶 = $2 𝑚𝑚

Typically, even for asymmetric CSAs, the MTAs and the rounding are the same for both
parties.

Margin Calls and Cash Flows

1. Margining Frequency of the Collateral Process: From an exposure perspective, the frequency
with which the amount of collateral is adjusted – the re-margining frequency – is a critical
component of the CSA. Following the financial crisis, most new IMA/CSAs, especially
between major financial institutions, have been using daily re-margining frequency in order
to reduce the amount by which the exposure can change relative to the collateral between the
margin calls. However, many small financial institutions or buy-side clients may not be able
to cope with the operational burden of frequent margin calls and will often negotiate longer
re-margining frequencies, e.g., weekly, monthly, or even longer.
2. Events Constituting the Margining Process: The amount of collateral held by the parties is
adjusted to their stipulated values 𝐴𝐷 and 𝐴𝐶 via the mechanism of a margin call. Many
models for exposure treat the margin call as an instantaneous event, taking place on the re-

22
 margining date and completed instantaneously. In practice the margin call is a chain of
events that takes several days to complete. With daily re-margining, several such chains run
concurrently in an interlaced manner; even as one margin call is yet to be settled, another one
already may be initiated. The time lag of this settlement process, long with the inherent lag of
the re-margining schedule, means that the changes in the VM are always running behind the
changes in the portfolio value. This, in turn, implies that the idealized expressions such as

𝐸(𝑡) = [𝑉(𝑡) − 𝑐(𝑡)]+

𝑐(𝑡) = 𝐴𝐶 (𝑡) − 𝐴𝐷 (𝑡)

are inaccurate. The detailed events involved in an initiation and the eventual settlement of a
margin call will be discussed in a later section.
3. Underlying Trade Cash Flow: With both default processes and margin settlements being non-
instantaneous events, it becomes relevant to track what payment flows take place – or not –
during the periods close to a default. Two types of payments are needed here. The first type –
called using the term trade flows, covers the contractual cash flows, physical settlements, and
other forms of asset transfers related to the trade themselves. These terms are spelt out in
trade documents and term sheets for each trade. The term trade flows rather than cash flows
is used to emphasize that term sheets may involve flows other than cash – such as transfers of
other non-cash assets, e.g. commodities, physical settlements resulting from the creation of
new trades from old ones, e.g. exercise of a physically settled swaption into a swap. A missed
trade flow is a serious event under the IMA, and a failure to pay can rapidly result in a
default and trade termination unless cured properly. Any missed trade flow is, of course, part
of the non-defaulting party’s claim.
4. CSA Specified Margin Cash Flow: The second type of flows is that that arises from the
exchange of collateral between the parties – margin flows. The legal treatment of the margin
flows is determined by the IMA/CSA, rather than by the trade documentation between the
parties. For purposes of this treatment, the most important aspect of the IMA/CSA is the
relatively mild treatment it affords to a party that misses a margin flow. Indeed, partially

23
 missing a margin payment is a common occurrence, as disputes about margin amounts
happen regularly, and sometimes persist for years.
5. Delays causing the Default Termination: During a collateral dispute, the CSA protocol calls
for the payment of the undisputed components of the collateral, but there is of course the
possibility that there will be no undisputed component at all, if one party’s counter-proposals
are sufficiently frivolous. Should suspicious about gaming arise, the CSA does contain a
methodology to stop disputes through market quotations, but the resulting leakage of the
position information is often a good deterrent to its use. As such, there is potential for abuse
by firms experiencing financial difficulties, and a good possibility that such abuse can go on
for some time before the dealer takes further efforts to end it. This, in turn, may result in a
fairly long period of time between the last fully settled margin call and the eventual
termination of a portfolio due to default.

Revised Exposure Definition

1. Stipulated vs. Realized Collateral Amount: In light of the discussion above, this section
makes a first effort at improving

𝐸(𝑡) = [𝑉(𝑡) − 𝑐(𝑡)]+

𝑐(𝑡) = 𝐴𝐶 (𝑡) − 𝐴𝐷 (𝑡)

For this consider a default of 𝐶 at time 𝜏 following an early termination of the trade portfolio
at time

𝑡≥𝜏

24
 At time 𝑡, let 𝐾(𝑡) be the collateral 𝐵 can actually rely on for the portfolio termination; this
amount will very likely differ from the CSA stipulated amount 𝐶(𝑡) – and from 𝐶(𝜏) for that
matter – due to the margin transfer time lags and some degree of non-performance by 𝐶.
2. Exposure Enhanced by Trade Flow: In addition, it is possible that some trade flows are
missed; denote their value at time 𝑡, including accrued interest, as 𝑈𝑇𝐹(𝑡). The exposure
generated by a default at time

𝜏≤𝑡

may be re-defined as

𝐸(𝑡) = [𝑉(𝑡) + 𝑈𝑇𝐹(𝑡) − 𝐾(𝑡)]+

Notice that the expression anchors the exposure at the termination date rather than at the
default date 𝜏 - this will be treated at a later section. For later use, the time-0 expectation of
the future time-𝑡 exposure is defined as

𝐸𝐸(𝑡) = 𝔼0 [𝐸(𝑡)]

where 𝔼 is the expectation operator in a relevant probability measure.

3. Impact of the Margin Timelines: Determining how 𝐾(𝑡) can differ from 𝑐(𝑡), and how large
can realistically 𝑈𝑇𝐹(𝑡) be, will require a more detailed understanding of the settlement and
the margining processes, a topic that will be treated in detail in a later section. The next goes
about determining how classical approaches go about modeling 𝐾(𝑡) and 𝑈𝑇𝐹(𝑡) in

𝐸(𝑡) = [𝑉(𝑡) + 𝑈𝑇𝐹(𝑡) − 𝐾(𝑡)]+

25
Classical Model for Collateralized Exposure – Assumptions about Margin
Flows

1. The Naïve Collateralized Exposure Model: A naïve, and now outdated, model for
collateralized exposure follows the definition

𝐸(𝑡) = [𝑉(𝑡) − 𝑐(𝑡)]+

𝑐(𝑡) = 𝐴𝐶 (𝑡) − 𝐴𝐷 (𝑡)

literally, and assumes that the collateral available is exactly equal to its prescribed value at
time 𝑡. That is, in the language of

𝐸(𝑡) = [𝑉(𝑡) + 𝑈𝑇𝐹(𝑡) − 𝐾(𝑡)]+

it is assumed that

𝐾(𝑡) = 𝑐(𝑡)

In addition, the parties are assumed to pay off all of the trade flows as described

𝑈𝑇𝐹(𝑡) = 0

and it is assumed that the termination date in

𝐸(𝑡) = [𝑉(𝑡) + 𝑈𝑇𝐹(𝑡) − 𝐾(𝑡)]+

is the default date 𝜏, i.e., there is no lag between the default date and the termination date. In
this model, the assumption of loss crystallized at a time 𝑡 is the function of the portfolio

26
 value at a single point 𝑉(𝑡) and does not depend on the earlier history 𝑉(∙) In the limit of
perfect CSA where

𝑐(𝑡) = 𝑉(𝑡)

the collateralized exposure in such a model is exactly zero.

2. Lag Induced Classical Exposure Model: Assuming

𝐾(𝑡) = 𝑐(𝑡)

is an idealization that ignores the non-instantaneous nature of collateral settlement protocols

and does not capture the fact that firms under stress may stop fully honoring margin calls,
resulting in a divergence between the portfolio value and the collateral value at some time lag
𝛿 before the termination of the portfolio. In what is denoted here as the Classical Model (see,
for example, Pykhtin (2010)), this particular lag effect is captured by modifying

𝐸(𝑡) = [𝑉(𝑡) − 𝑐(𝑡)]+

𝑐(𝑡) = 𝐴𝐶 (𝑡) − 𝐴𝐷 (𝑡)

to

𝐸(𝑡) = [𝑉(𝑡) − 𝐾(𝑡)]+

𝐾(𝑡) = 𝑐(𝑡 − 𝛿)

So, for instance, for a CSA with thresholds ℎ𝐷 and ℎ𝐶 and from

𝐴𝐷 (𝑡) = [−𝑉(𝑡) − ℎ𝐷 ]+

27
 𝐴𝐶 (𝑡) = [𝑉(𝑡) − ℎ𝐶 ]+

one gets

𝐾(𝑡) = [𝑉(𝑡 − 𝛿) − ℎ𝐶 (𝑡 − 𝛿)]+ − [−𝑉(𝑡 − 𝛿) − ℎ𝐷 (𝑡 − 𝛿)]+

3. Drawbacks of the Classical Exposure Model: Having a mechanism for capturing the
divergence between the collateral and the portfolio value is an important improvement over
the older method described above, and the classical model has gained widespread acceptance
for both the CVA (Credit Valuation Adjustment) and the regulatory calculations.
Nevertheless, it hinges on a number of assumptions that are unrealistic. For instance,

𝐸(𝑡) = [𝑉(𝑡) − 𝐾(𝑡)]+

𝐾(𝑡) = 𝑐(𝑡 − 𝛿)

assumes that both 𝐷 and 𝐶 will simultaneously stop paying margin at time 𝑡 − 𝛿 freezing the
margin level over the entire MPoR. In reality, if the party due to post collateral at 𝑡 − 𝛿
happens to be the non-defaulting party 𝐷, it will often continue doing so for some time even
in the presence of the news about the possible impending default of 𝐶. And should 𝐶 miss a
few margin payments (maybe under the guise of a dispute), 𝐷 would often continue to post
collateral while it evaluates its options. This creates an asymmetry between posting and
receiving collateral that the classical model fails to recognize.
4. Impact of Lag on Exposure: In

𝐸(𝑡) = [𝑉(𝑡) − 𝐾(𝑡)]+

𝐾(𝑡) = 𝑐(𝑡 − 𝛿)

28
 the lag parameter 𝛿 is clearly critical; the larger 𝛿 is, the more 𝑉(𝑡) may pull from the frozen
margin value at time 𝑡 − 𝛿 and bigger the expected exposure will become. In practice, the
determination of 𝛿 will often be done in a simplistic manner, e.g. by using a fixed lag (10 𝐵𝐷
is common), or, more realistically, by adding a universal time delay to the re-margining
frequency of the CSA in question. This practice is echoed in regulatory guidelines, e.g., in
Basel 3 accord where MPoR is set to the re-margining frequency minus 1 𝐵𝐷 plus an MPoR
floor that defaults to 10 𝐵𝐷. The MPoR floor must be increased in certain cases, e.g., for
large netting sets, illiquid trades, illiquid collateral, and recent collateral disputes – however,
the increase is specified as a multiplier relative to the same default. With a high proportion of
individually negotiated and amended features in real life IMA/CSAs, using a one size fits all
assumption may, however, lead to significant inaccuracies.

Assumptions about Trade Flows

1. The Classical+ Collateral Exposure Model: Because large trade flows after the start of MPoR
may no longer be followed by collateral adjustment, they have the potential to either
extinguish or exacerbate the exposure. For this reason, the model assumptions with respect to
the date when either party suspends the trade flows are likely to have a significant impact on
the counterparty credit loss. In one common interpretation of the classical model, it is simply
assumed that both 𝐷 and 𝐶 will continue to pay all the trade flows during the entire MPoR.
As a consequence, the unpaid trade flow term 𝑈𝑇𝐹(𝑡) in

𝐸(𝑡) = [𝑉(𝑡) + 𝑈𝑇𝐹(𝑡) − 𝐾(𝑡)]+

will be zero, consistent with

𝐸(𝑡) = [𝑉(𝑡) − 𝐾(𝑡)]+

29
 𝐾(𝑡) = 𝑐(𝑡 − 𝛿)

For ease of reference this version of classical model is denoted as Classical+.

2. The Classical- Collateral Exposure Model: In another, less common, version of the Classical
Model, the assumption is that both 𝐶 and 𝐷 will stop paying trade flows at the moment the
MPoR commences, i.e., at time 𝑡 − 𝛿. In this case, the unpaid trade flows are set equal to

𝑈𝑇𝐹(𝑡) = 𝑇𝐹𝑁𝐸𝑇 (𝑡; (𝑡 − 𝛿, 𝑡])

where 𝑇𝐹𝑁𝐸𝑇 (𝑡; (𝑡′, 𝑡"]) is the time 𝑡 value of all net trade flows scheduled to be paid in the
interval (𝑡′, 𝑡"]. Note that the time is measured in discrete units of business days, such that
the notation (𝑢, 𝑠] is equivalent to [𝑢 + 1𝐵𝐷, 𝑠]. Further, if 𝑡 is after the margin flow date,
the trade flow value accrues from the payment date to 𝑡 at a contractually specified rate. This
version of the classical model is denoted Classical-; it is associated with an exposure
definition of

𝐸(𝑡) = [𝑉(𝑡) + 𝑇𝐹𝑁𝐸𝑇 (𝑡; (𝑡 − 𝛿, 𝑡]) − 𝑐(𝑡 − 𝛿)]+

3. Inadequacies of the Classical Exposure Models: In practice neither the Classical+ exposure
equation

𝐸(𝑡) = [𝑉(𝑡) − 𝐾(𝑡)]+

𝐾(𝑡) = 𝑐(𝑡 − 𝛿)

nor the Classical- exposure equation

𝐸(𝑡) = [𝑉(𝑡) + 𝑇𝐹𝑁𝐸𝑇 (𝑡; (𝑡 − 𝛿, 𝑡]) − 𝑐(𝑡 − 𝛿)]+

30
 are accurate representations of reality. Trade flows are likely to be paid at least by 𝐷 at the
beginning of the MPoR, and are likely not to be paid by 𝐶 at least at its end. For instance, due
to the CSA protocol for collateral calculations (next section), there is typically a 3 𝐵𝐷 lag
between the start of an MPoR – the market observation date for the last full margin payment
– and the date when 𝐷 definitely observes that 𝐶 has missed paying a margin flow; during
this period 𝐷 would always make all trade payments unless 𝐶 commits any additional
contract violations. Even after 𝐷 has determined that 𝐶 has missed a margin payment, 𝐷’s
nominal rights to suspend payments following a breach would, as mentioned earlier, not
always be exercised aggressively. Legal reviews, operational delays, and grace periods can
further delay the time when 𝐷 stops paying trade flows to 𝐶.
4. Accrual of Missed Trade Flows: Another trade flow effect arises during the last 2 − 3 days
of the MPoR (just prior to termination) where 𝐶 has already defaulted and neither party is
likely making trade payments. Here, the IMA stipulates that the missed trade flows within
this period accrue forward at a contractually specified rate and become part of the bankruptcy
claim. This gives rise to a termination period in addition to the 𝑈𝑇𝐹(𝑡) term, in turn leading
to an adjustment of the exposure.

Full Timeline of IMA/CSA Events

1. Details of the IMA/CSA Processes: Loosely speaking, the IMA concerns itself with the
events of default, termination, and close out, and the CSA governs the collateral exchanges,
including the concrete rules for collateral amount calculations and posting frequencies. While
this chapter so far has touched on the workings of the IMA/CSA in the previous sections,
model construction hereafter will require more detailed knowledge of certain provisions
regarding the normal exchange of collateral, the legal options available in the case of missed
payments, and the common dealer policies with respect to availing itself of these options. A

31
 detailed exposition of the IMA and the CSA legal complexities can be found in multiple
sources, including at http://www.isda.org; here only a brief summary to the extent necessary
to develop the model is provided. The focus is on the development of a plausible timeline of
events taking place around the default, and the subsequent portfolio termination.

Events Prior to Default

1. Calculated vs. Actual Collateral Amounts: It is assumed that the dealer 𝐷 is the Calculation
Agent for the computation of the collateral amounts. As before, 𝐴𝐶 and 𝐴𝐷 denote the
prescribed collateral amounts for 𝐶 and 𝐷; as discussed, they may differ from the actually
available collateral amounts 𝑀𝐶 and 𝑀𝐷 if one of the parties fails to make the margin flow or
changes the prescribed amount.
2. CSA Specified Margin Process Timelines: The following list describes the complete
sequence of events taking place at times 𝑇0 , 𝑇1 , ⋯ The next section simplifies and condenses
these into a tractable model.
3. Collateral/Portfolio Valuation Date 𝑇0 : The timeline begins at 𝑇0 , the as-of-date at which the
value of the portfolio and its collateral are measured, for usage in the 𝑇1 evaluation of the
formulas for the Credit Support Amount – plainly, the amount of collateral. Typically, 𝑇0 is
the close of business on the business day before 𝑇1 .
4. Honored Collateral Invocation Date 𝑇1 : For the purposes of this treatment, 𝑇1 is used to refer
to the last undisputed and respected Valuation Date prior to default. At time 𝑇1 , besides
officially determining 𝐴𝑐 (𝑇0 ) and 𝐴𝐶 (𝑇0 ), the dealer 𝐷 calculates the incremental payment
amounts to itself and to 𝐶 as

𝑚𝐷 = 𝐴𝐷 (𝑇0 ) − 𝑀𝐷 (𝑇0 )

32
 and

𝑚𝐶 = 𝐴𝐶 (𝑇0 ) − 𝑀𝐶 (𝑇0 )

respectively. Taking into account any minimum transfer amounts, the transfer amounts 𝑚𝐷
and 𝑚𝐶 should normally be communicated by 𝐷 to 𝐶 prior to a Notification Time (e.g., 1 𝑃𝑀
local time).
5. Collateral Transfer Initiation Date 𝑇2 : After receiving the notice of the calculated collateral
amount, 𝐶 must initiate the transfers of the sufficient amount of eligible collateral on the
payment date 𝑇2 . Assuming that 𝐷 managed to have the collateral notification sent to 𝐶 prior
to the Notification Time, 𝑇2 defaults to 1 𝐵𝐷 after 𝑇1 . If 𝐷 is late in its notification, 𝑇2 would
be 2 𝐵𝐷 after 𝑇1 . It is assumed here that the required amounts – recalling that they were
calculated at 𝑇1 using market data at time 𝑇0 – are all settled without incident at 𝑇2 . However,
𝑇2 will be the last time that margin flows settle normally before the default takes place.
6. Non-honored Collateral Calculation Date 𝑇3 : Let 𝑇3 denote the next scheduled valuation date
after 𝑇1 . If 𝛼 is the average scheduled time between collateral calculations, one
approximately has – ignoring business calendar effects –

𝑇3 ≈ 𝑇1 + 𝛼

At 𝑇3 – hopefully before the Notification Time - 𝐷 will be able to send payment notice to 𝐶,
but 𝐶 has fallen into financial distress and will not be able – or willing – to pay further
margin flows. Should 𝐶 simply fail to pay collateral at the next payment date, a Credit
Support Default could be triggered shortly thereafter (non-payment of collateral is associated
with a 2 𝐵𝐷 grace period). To prevent this from happening, it is, as discussed earlier, likely
that 𝐶 would attempt to stall by disputing the result of the 𝑇3 collateral calculation by 𝐷.
7. Potential Event of Default 𝜏: Exactly how long the margin dispute is allowed to proceed is a
largely a behavioral question that requires some knowledge of 𝐷’s credit policies and its
willingness to risk legal disputes with 𝐶. Additionally, one needs to consider to what extend
𝐶 is able to conceal its position of financial stress by using dispute tactics, or, say, blaming

33
 operational issues on its inability to pay collateral. Ultimately, however, 𝐷 will conclude that
𝐶 is in default of its margin flows (a Credit Support Default), or 𝐶 will commit a serious
contract violation such as failing to make a trade-related payment. At that point 𝐷 will
conclude that a Potential Event of Default (PFD) has occurred. The time of this event is
identified as the true default time 𝜏.
8. Client PED Communication Date 𝑇4 : Once the PED has taken place, 𝐷 needs to formally
communicate it to 𝐶, in writing. Taking into account mail/courier delays, legal reviews, and
other operational lags, it is likely that the communication time, denoted 𝑇4 , takes place at a
slight delay to the PED.
9. Event of Default Date 𝑇5 : After the receipt of the PED notice, 𝐶 will be granted a brief period
of time to cure the PED. The length of this cure period is specified in the IMA and depends
on both the type of the PED and the specific IMA. For instance, of the PED in question is
Failure to Pay, the default cure period is 3 𝐵𝐷 in the 1992 IMA and 1 𝐵𝐷 in the 2002 IMA –
this may very well be overridden in the actual documents. At the end of the cure period –
here denoted 𝑇5 – and Event of Default (ED) formally crystallizes. It is emphasized here at
the 𝑇5 (the official default time) is not associated with the true default time 𝜏; instead 𝜏 is
equated to the time of the actual default (the PED) that, after contractual formalities, will lead
to the default of 𝐶.
10. ED Communication (𝑇6 ) and ETD Designation Dates (𝑇7 ): After the ED has taken place, 𝐷
will inform 𝐶 of the ED at time

𝑇6 ≥ 𝑇5

and may, at time

𝑇7 ≥ 𝑇6

elect to designate an Early Termination Date (ETD).

11. Early Termination Date 𝑇8 : The ETD is denoted 𝑇8 ; per the IMA it is required that 𝑇8 ∈
[𝑇7 , 𝑇7 + 20𝐷] The ETD constitutes the as-of-date for the termination of 𝐶’s portfolio and

34
 collateral positions. Many dealers will aim for speedy resolution in order to minimize market
risk, and will therefore aim to set the ETD as early as possible. There are, however, cases
where this may not be optimal, as described in the Section below.
12. Post Client ETD Establishment Events: Once the portfolio claim has been established as of
the ETD, the value of any collateral and unpaid trade flows held by 𝐶 is added to the amount
owed to 𝐷. Paragraph 8 of the CSA then allows 𝐷 to liquidate any securities collateral in its
possession and to apply the proceeds against the amount it is owed. Should the collateral be
insufficient to cover what is owed to 𝐷, the residual amount will be submitted as a claim in
𝐶’s insolvency. The claim is usually challenged by the insolvency representative, and where
parties cannot agree, may be referred to court. It can sometimes take a long time before the
claim is resolved by the bankruptcy courts and the realized recovery becomes known. The
interest on the recovery amount for this time is added to the awarded amount. Note that this
chapter focusses exclusively on modeling the magnitude of exposure and bankruptcy claim,
and does not challenge the established way of modeling the amount and the timing of the
eventual recovery using a loss-given-default (LGD) fraction.

Some Behavioral and Legal Aspects

1. Margin Exposure Modeling Parametrization Components: With the timeline just having been
established, it remains for it to be tied with a proper model for exposure. In order to do so, as
already mentioned, the timeline needs to be combined with coherent assumptions about the
dealer and the client behavior in each sub-period. The assumptions should be determined not
only by the rights available under the IMA/CSA, but also by the degree of operational
efficiencies in serving notices and getting legal opinions, and also by the level of prudence
injected into the assumptions about the dealer ability and willingness to strictly uphold
contractual terms within each client group as it pertains to margin flows and disputes.

35
2. Issues with Exercising Suspension Rights: From a legal rights perspective, the most
important observation is that once notice of a PED has been served (time 𝑇4 ) the so called
suspension rights of IMA (Section 2(a)(iii)) and the CSA (Paragraph 4(a)) will allow 𝐷 to
suspend all trade- and collateral- related payments to 𝐶 until the PED has been cured. The
extent to which the suspension rights are actually exercised, however, is quiet situational. A
particular danger is that 𝐷 exercises its suspension rights due to a Potential Event of Default
(PED), but that subsequently the PED is ruled to be not valid. Should this happen, the dealer
can inadvertently commit a breach of contract which, especially in the presence of cross-
default provisions, can have serious consequences for the dealer.
3. Choice of Designating an ETD: Another, somewhat counter-intuitive, reason for 𝐷 not to
enforce its suspension rights is tied to IMA Section 2(a)(iii), which can sometimes make it
favorable for 𝐷 to never designate an ETD. Indeed, if 𝐷 owes 𝐶 money, it would seem a
reasonable course of action for 𝐷 to simply:
a. Never designate and ETD, and
b. Suspend all the payments in the portfolio until the default gets cured – which most
likely will never happen.
This tactic basically allows 𝐷 to walk away from its obligations on the portfolio when 𝐶
defaults, effectively making 𝐷 a windfall gain.
4. Jurisdiction Legality of the ETD Delays: The strategy of delaying the ETD is perpetuity has
been tested by UK courts and found legal – although contract language has been proposed by
ISDA to prevent the issue. In the US, however, local safe haven laws have been ruled to
prevent ETD’s for more than about one year. Still a one-year delay may prove tempting if 𝐷
has a big negative exposure to 𝐶 and is unwilling to immediately fund the large cash flow
needed to settle. As most large dealers are presumably unlikely to play legal games with the
ETD, this topic shall not be considered further here, but note that there is room to make more
aggressive model assumptions around the ETD’s than is done here.

36
Simplified Timeline of IMA/CSA Events

1. Motivation for the Timeline Simplification: It should be evident from the preceding section
that the full timeline of IMA/CSA reviewed earlier is in many ways different, and more
complex, than what is assumed in the Classical- and the Classical+ versions of the classical
model. However, it is equally evident that the timeline is too complex to be modeled in every
detail. This section offers a simplification of the timeline designed to extract the events most
important for exposure modeling. The resulting model offers several important
improvements over the classical model, while remaining practical and computationally
feasible.

Identification of Key Time Periods

1. Classical MPOR Start/End Dates: To recap, first the classical model only considers two dates
in the timeline of default; the start and the end of the MPoR. The start of the MPoR, denoted
by 𝑡 − 𝛿, is defined as the last observation date for which the margin was settled in full (a
few days after the observation date). The end of the MPoR, denoted by 𝑡, is the observation
date on which 𝐷’s claim is established. Note that 𝑡 coincides with the IMA’s Early
Termination Date (ETD) discussed earlier.
2. Classical Model Lag Length Error: In the classical model there is no clear distinction
between the observation and the payment dates, making it difficult to cleanly capture the
trade flow effects. For instance, in the classical version of the model, 𝑡 − 𝛿 denotes both the
last observation date as well as the dates on which all trade flows cease. In reality, the last
margin observation date is unlikely to be contentious and trigger stoppage of trade flows, as
the margin payment to which the observation corresponds to will only be missed by 𝐶 several

37
 business-days later. Specifically, if the market data is observed on day 0, the valuation is
performed in day 1, then only on day 2 (or 3 if the notification was late) is the initiation of
the actual payment expected to take place. The length of this lag is of the same order of
magnitude as typical assumptions for the length of the MPoR, and can be a source of
considerable model error if not handled properly.
3. Delineating Observations and Payment Dates: In the simplified timeline proposed here, care
is kept to take care of the distinction between the observation and the payment dates, and also
to consider the possibility that 𝐷 may take the action of stopping a particular type of flow at a
different time than 𝐶 does. Accordingly, the model includes two potentially different
observation dates for which 𝐷 and 𝐶 later settle their margin flows in full for the last time;
and two potentially different dates when they pay their trade flows respectively for the last
time. The end of the MPoR is defined as in the same way as in the classical model, to
coincide with the ETD. The table below summarizes the notation for the five dates in the
simplified timeline.
4. Notation for the Dates in the Simplified Timeline:

Event Date Type Notation

Observation Date for the Last Margin Flow from 𝐶 Observation 𝑡𝐶 = 𝑡 − 𝛿𝐶

Observation Date for the Last Margin Flow from 𝐷 Observation 𝑡𝐷 = 𝑡 − 𝛿𝐷

Observation Date for the Last Trade Flow Payment from 𝐶 Settlement 𝑡𝐶 ′ = 𝑡 − 𝛿𝐶 ′

Observation Date for the Last Trade Flow Payment from 𝐷 Settlement 𝑡𝐷 ′ = 𝑡 − 𝛿𝐷 ′

ETD Observation 𝑡

5. Current Scheme MPoR Start Date: The start of the MPoR in the current model is 𝑡 − 𝛿,
which in the notation of table above may be defined symmetrically as

𝛿 = max(𝛿𝐶 , 𝛿𝐷 )

38
 𝐶 is always expected to stop posting margin no later than the non-defaulting party 𝐷, and
therefore on would very likely have

𝛿𝐶 ≥ 𝛿𝐷

and

𝛿 = 𝛿𝐶

6. Exposure Model Timeline Lag Choices: The second column in the table above specifies
which of the dates is the observation date, and which is the settlement or the payment date.
According to the notation established in the table, 𝛿𝐶 and 𝛿𝐷 are the lengths of time
preceding the ETD during which changes in the portfolio values no longer result in collateral
payments by 𝐶 and 𝐷, respectively. Similarly, 𝛿𝐶 ′ and 𝛿𝐷 ′ are the lengths of time preceding
the ETD during which the respective parties do not pay trade flows. In, say, a Classical 10-
day MPoR model

𝛿𝐶 = 𝛿𝐷 = 10 𝐵𝐷

with

𝛿𝐶′ = 𝛿𝐷′ = 0

for Classical+ and

𝛿𝐶′ = 𝛿𝐷′ = 10 𝐵𝐷

for Classical-.

39
Establishing the Sequence of Events

1. Order of the MPoR Events: A priori, the four events in the Table between the start event and
the end event of the MPoR can occur in any order. However, this section will now explain
why the table very likely shows the proper sequence of events.
2. Time Lag between Margin/Trade Flows: As discussed earlier, missing trade flows are
recognized as a more serious breach of contractual obligations than missing margin flows,
especially since the latter may take the form of a margin valuation dispute. Therefore, it is
reasonable to assume that neither party will stop paying the trade flows before stopping the
payment of margin flows. Accounting for the margin settlement lag between the observation
date and the payment date, this yields

𝛿𝐶′ ≤ 𝛿𝐶 − 𝑀𝑎𝑟𝑔𝑖𝑛 𝑆𝑒𝑡𝑡𝑙𝑒𝑚𝑒𝑛𝑡 𝐿𝑎𝑔

𝛿𝐷′ ≤ 𝛿𝐷 − 𝑀𝑎𝑟𝑔𝑖𝑛 𝑆𝑒𝑡𝑡𝑙𝑒𝑚𝑒𝑛𝑡 𝐿𝑎𝑔

3. Lag between Dealer/Client Events: It is also reasonable to assume that either of the two types
of flows is first missed by the defaulting party 𝐶, and then only by the non-defaulting party
𝐷. This leads to the following additional constraints on the sequence of events within the
timeline:

𝛿𝐶 ≥ 𝛿𝐷

𝛿𝐶 ′ ≥ 𝛿𝐷 ′

4. Client Settlement vs. Dealer Observation: Except in rare and unique situations such as
outright operational failures, 𝐷 would not continue to pay margin flows once 𝐶 commits a
more serious violation by missing a trade flow, resulting in
40
 𝛿𝐶′ ≤ 𝛿𝐷 − 𝑀𝑎𝑟𝑔𝑖𝑛 𝑆𝑒𝑡𝑡𝑙𝑒𝑚𝑒𝑛𝑡 𝐿𝑎𝑔

Combining these inequalities results in the chronological order of events shown in the table
above.

Evaluation of the Client Survival Probability

1. Client Survival at MPoR Start: As was the case for the classical model, the setup anchors the
exposure date 𝑡 at the termination date ETD, at the very end of the MPoR. The ETD is the
same for both parties, and constitutes a convenient reference point for aligning the actions of
one party against those of the other. It needs to be emphasized that the ETD for which the
exposure is evaluated does not coincide with the date at which the survival probability is
evaluated, e.g. for the computation of the CVA. In the simplified timeline, the counterparty
survival probability should be evaluated for 𝑡 − 𝛿𝐶 ′, the last date when 𝐶 stops paying trade
flows – effectively assuming that the default is due to failure-to-pay. Hence, if 𝐸𝐸(𝑡) is the
expected exposure anchored at the ETD 𝑡, then the incremental contribution to the unilateral
CVA from time 𝑡 is, under suitable assumptions, 𝐸𝐸(𝑡) ∙ ∆ℙ𝑡−𝛿𝐶 [∙] where ℙ is the survival
probability under the model’s measure – later sections contain concrete examples.
2. Client Survival at MPoR End: Evaluating the default probability at the anchor date 𝑡 rather
than at 𝑡 − 𝛿𝐶 ′ will introduce the slight error in computing the survival probability. While
this error is relatively small and is often ignored by practitioners, it takes virtually no effort,
and has no impact on model efficiency, to evaluate the survival probability at the right date.

41
Timeline Calibration

1. Client Customization of IMA/CSA Specifications: As mentioned earlier, the specific

IMA/CSA terms for a given counterparty should always be ideally examined in detail, so that
any non-standard provisions may be analyzed by their impact on the timeline. For those cases
where such bespoke timeline construction is not practical (typically for operational reasons),
two standard (reference) timelines are proposed here. This will allow the demonstration of
the thought process behind the timeline calibration, and will provide some useful base cases
for later numerical tests.
2. Parametrizing the Client and the Dealer Timelines: While factors such as portfolio size and
dispute history with the counterparty should, of course, be considered in establishing the
MPoR, an equally important consideration in calibrating the model is the nature of expected
response by 𝐵 to a missed margin or trade flow by 𝐶. Even under plain vanilla IMA/CSA
terms, experience shows that the reactions to contract breaches are subject to both human and
institutional idiosyncracies, rendering the MPoR quiet variable.
3. Aggressive vs. Conservative Timeline Parametrization: Recognizing that “one size does not
fit all”, two different calibrations shall therefore be considered; one aggressive which
assumes the best case scenario for rapidly recognizing the impending default, and taking
swift action; and one conservative, which takes into consideration not only a likely delay in
recognizing that the counterparty default is imminent, but also the possibility that the bank
may not aggressively enforce its legal rights afforded under IMA and CSA in order to avoid
damaging its reputation. In both scenarios, daily re-margining is assumed – if a CSA calls for
less frequent margin calls than this, the MPoR must be lengthened accordingly.

Aggressive Calibration

42
1. Applicability of the Aggressive Calibration: The aggressive calibration applies to trading
relationship between two counterparties that both have string operational competence, and
where there is little reputational risk associated with swift and aggressive enforcement of the
non-defaulting party’s legal rights against the defaulting party.
2. Inter-dealer Monitoring and Call-outs: A good example would be trading between two large
dealers, both willing to aggressively defend against a possible credit loss. The credit officers
here are assumed to be diligent in the monitoring of their counterparties, and generally be
able to see a default developing, rather than be caught by surprise.
3. Full Application of Operational Sophistication: Under aggressive calibration, the event of 𝐶
missing or disputing a margin call by any non-trivial amount will, given 𝐶’s sophistication,
immediately alert 𝐷 that an impending default is likely. 𝐷 will not be misled by claims of
valuation disputes or other excuses, and will send a Notice of Credit Support Default under
the IMA/CSA the next business day after the breach of the margin agreement. At the same
time, to protect itself further, 𝐷 will stop both the margin and the trade flows. The
counterparty is assumed to simultaneously stop paying margin and trade flows as well, so
that no further payments of any kind are exchanged by the parties.
4. Elimination of Settlement/Herstatt Risk: The simultaneous action by both parties in the
Aggressive scenario to stop paying the trade flows at the earliest possible moment results in
the elimination of all settlement risk – the possibility that the dealer may continue paying on
its trade flow obligations while not receiving promised payments in return. In the context of
cross-currency trades, this type of settlement risk is frequently referred to as the Herstatt risk,
after the bank that caused large counterparty losses in this manner
(https://en.wikipedia.org/wiki/Settlement_risk). Such risk shall be captured in the
Conservative Calibration case below, and shall be discussed in more detail in a later section.
5. Timeline of the corresponding MPoR: Despite 𝐷’s immediate and aggressive response, the
MPoR will still be fairly long due to the way the IMA/CSA operates in practice. In particular,
notice that the first period in the simplified timeline is between the last observation date for
which the margin was fully settled, and the first date for which 𝐶 misses a margin flow.
6. Breakdown of the CSA Steps: As it takes at least 2 business days to settle a margin payment,
plus 1 business day between the last margin that was successfully settled and the first margin

43
 payment that was not, a minimum of 3 business days will accrue from the start of an MPoR
and a margin-related PED. Further, once the margin flow is missed, 𝐷 must send at least 2
notices and permit a grace period – usually 2 business days – to cure the violation before an
event of default (ED) has officially taken place and an ETD has been designated.
7. Comparison with Classical MPoR Timeline: Since an ETD cannot be designated prior to the
event of default, it is unlikely that an MPoR can ever be less than 7 business days. It is
remarkable that even under the most aggressive set of assumptions, the MPoR is still only 3
business days shorter than the classical 2-week MPoR.
8. Detailed Breakdown of the Aggressive Timeline: The detailed taxonomy of the aggressive
timeline is listed in the table down below, and essentially splits the MPoR into two sections;
a margin delay period of 3 business days, and a default resolution period of 4 business days.
During the latter period, 𝐶 and 𝐷 cease paying on the first day, leaving a period of 3 business
days where neither party makes any payments. Notice that it is assumed that the ETD is
declared to coincide with the ED, i.e., the dealer will terminate as quickly as legally possible.

Conservative Calibration

1. Non-aggressive Enforcement of CSA Rights: The conservative calibration is intended to

cover the situation where the dealer’s enforcement of its rights under the IMA/CSA is
deliberate and cautious, rather than swift. There may be several reasons for such a situation,
sometime acing in tandem.
2. Applicable Clients - Less Sophisticated Participants: First, a dealer, if overly trigger-happy,
can gain a market-wide reputation as being rigid and litigious, potentially causing clients to
seek other trading partners. In fact, should aggressive legal maneuvers be applied to
counterparties that may be considered “unsophisticated”, there is even a potential for the
dealer to be perceived as predatory by the larger public.

44
3. “Leakage” of Dealer Positions: Second, there are situations where exercising the legal rights
would cause an unattractive leakage of information into the broader market. As indicated
earlier, this may happen for instance if the formal collateral dispute methodology of
Paragraph 5 of the IDSA CSA is activated; the market poll inherent in the methodology
would inevitably reveal the positions held with the counterparty to competing dealers.
4. Ramification of Aggressive Legal Exercise: Third, sometimes an aggressive interpretation of
the legal rights can backfire in the form of lawsuits and counter-measures by the
counterparty. For example, even when the dealer may have the rights to withhold payments
(e.g., under Section 2(a)(iii)), it would often elect to not exercise this right immediately out
of concern that a counter-ED would be raised against it or that withholding payments would
exacerbate the liquidity situation of the counterparty potentially exposing the dealer to
liabilities and lawsuits.
5. Damage from “Improper PED” Rulings: As mentioned, a particular danger is that the dealer
exercises its suspension rights due to a Potential Event of Default (PED), but that
subsequently the PED is ruled to not be valid. Should this happen, the dealer can
inadvertently commit a breach of contract.
6. Limitations with the Dealers’ Operational Capacity: Of course, even if a dealer may
potentially be willing to aggressively exercise its rights, it may not have the operational
capacity to do so quickly. For example, the dealer may not be able to perform the required
legal review on a short notice, or may not always have the efficiency to get the notices
mailed out at the earliest possible date. On top of this there is always potential for technology
related and human errors and oversights.
7. Timeline Incurred by Conservative Calibration: While it is harder to get concrete data to
estimate a reasonable timeline for the Conservative case (this case being dependent not only
on the IMA/CSA details, but also on the specifics of the dealer’s reputational considerations),
under a perfectly reasonable set of assumptions the MPoR ends up being more than twice as
long as for the Aggressive case above. Under this calibration choice, the Conservative
scenario assumes that the totality of the margin dispute negotiations, operational delays,
human errors, legal reviews etc., adds up to 8 business days, yielding an MPoR of a total of
15 business days.

45
8. Typical Conservative CSA Event Timeline: One plausible scenario with daily re-margining
could be:
a. 𝑡 − 15 ⟹ 𝐷 observes the portfolio value as needed for the margin transfer amount
#1 as of 𝑡 − 15.
b. 𝑡 − 14 ⟹ 𝐷 sends margin call #1 to 𝐶; 𝐷 observes a margin transfer amount #2.
c. 𝑡 − 13 ⟹ 𝐷 sends margin call #2 to 𝐶; 𝐶 honors margin call #1; 𝐷 observes a
margin transfer amount #3.
d. 𝑡 − 12 ⟹ 𝐶 fails to honor margin call #2 and initiates dispute; 𝐷 tries to resolve the
dispute while still paying and calculating the margin.
e. 𝑡 − 7 ⟹ 𝐶 fails to make a trade payment.
f. 𝑡 − 6 ⟹ 𝐷 stops paying margin and sends a PED notice.
g. 𝑡 − 5 ⟹ 𝐶 receives PED; 𝐷 keeps making trade payments.
h. 𝑡 − 3 ⟹ The PED is not cured.
i. 𝑡 − 2 ⟹ 𝐷 stops making trade payments and sends an ED notice to 𝐶, designating 𝑡
as the ETD.
j. 𝑡 ⟹ ETD.
9. Current/Interlacing Outstanding Margin Process: Notice that a number of different margin
processes are simultaneously active (denoted #1, #2, and #3), reflecting the interlacing nature
of the daily margin calls. Also, unlike the earlier Aggressive Calibration, the above scenario
explicitly involves settlement risk, as a time period exists only where 𝐷 pays trade flows
(from 𝑡 − 7 to 𝑡 − 3, both dates inclusive).
10. Dealer/Client Payment/Settlement Lags: To translate the scenario above into the notation of
the earlier sections, first notice that

𝛿𝐶 = 15

since the observation date of the last margin call (#1) honored by 𝐶 is 𝑡 − 15. Second, as 𝐷
makes its last possible margin call at 𝑡 − 7 based on an observation at time 𝑡 − 9

𝛿𝐷 = 9

46
 Third, as 𝐶 fails to make a trade payment at 𝑡 − 7, 𝐶’s last payment date is 𝑡 − 8, and
therefore

𝛿𝐶′ = 8

And, finally, since 𝐷 stops its trade payments at 𝑡 − 2

𝛿𝐷′ = 3

Summary and Comparison of Timelines

1. Classical+/Classical-/Aggressive/Conservative Parametrizations: Using the notation above,

the Aggressive and the Conservative scenarios are presented in the table below. For
reference, the Classical+ and the Classical- versions of the classical model are presented in
the table as well. Note that the 10 𝐵𝐷 assumption of the classical MPoR lies between the two
calibration choices proposed, and is closer to the Aggressive scenario.
2. MPoR Periods for CSA’s with Daily Re-margining:

Parameter Conservative Aggressive Classical+ Classical-

𝛿𝐶 15 𝐵𝐷 7 𝐵𝐷 10 𝐵𝐷 10 𝐵𝐷

𝛿𝐷 9 𝐵𝐷 6 𝐵𝐷 10 𝐵𝐷 10 𝐵𝐷

𝛿𝐶 ′ 8 𝐵𝐷 4 𝐵𝐷 0 𝐵𝐷 10 𝐵𝐷

𝛿𝐷 ′ 3 𝐵𝐷 4 𝐵𝐷 0 𝐵𝐷 10 𝐵𝐷

47
3. Caveats over Aggressive/Conservative Parameters: The Aggressive and the Conservative
parameter choices represent two opposite types of dealer-client relationships, and may also
be used as two limit scenarios for materiality and model risk analysis. Of course, the best
approach would always be to set the model parameters based on prudent analysis of the
firm’s historical default resolution timelines, to the extent that it is practically feasible. The
model could also conceivably treat the various time lags as random variables to be simulated
as part of the exposure computations; yet it is debatable whether increasing the number of
model parameters this way is warranted in practice.

Unpaid Margin Flows and Margin Flow Gap

1. Margin and Trade Flow Gaps: To formulate the model in more precise mathematical terms,
this section returns to

𝐸(𝑡) = [𝑉(𝑡) + 𝑈𝑇𝐹(𝑡) − 𝐾(𝑡)]+

and considers how to draw the analysis of the previous sections to reasonably specify both
the collateral amount 𝐾(𝑡) as well as the value 𝑈𝑇𝐹(𝑡) of net unpaid cash flows.
2. Client Last Margin Posting Date: As with the classical model, it is assumed that the MPoR
starts at time

𝑡𝐶 = 𝑡 − 𝛿𝐶

48
 the portfolio observation date associated with the last regular collateral posting by 𝐶. Recall
that the classical model further assumes that 𝐷 will stop posting collateral simultaneously
with 𝐶 so that

𝐾(𝑡) = 𝑐(𝑡𝐶 )

where 𝑐(𝑡𝐶 ) denotes the CSA prescribed collateral support amount calculated from the
market data observed at time 𝑡𝐶 . This is to be compared with

𝐸(𝑡) = [𝑉(𝑡) − 𝐾(𝑡)]+

𝐾(𝑡) = 𝑐(𝑡 − 𝛿)

3. Definition of the Margin Flow Gap: In contrast to

𝐾(𝑡) = 𝑐(𝑡𝐶 )

this model assumes that 𝐷 will continue posting and returning collateral to 𝐶 for all
contractual margin observations dates 𝑡𝑖 whenever required by the CSA, after

𝑡𝐶 = 𝑡 − 𝛿𝐶

and up to and including the observation date

𝑡𝐷 = 𝑡 − 𝛿𝐷

The presence of an observation period of non-zero length for which 𝐷 is posting and
returning collateral but 𝐶 is not is referred to as margin flow gap.
4. Choosing the Collateral Computation Date: Here, it is always expected that

49
 𝑡𝐷 ≥ 𝑡𝐶

which therefore in effect assumes the possibility of a time interval (𝑡𝐷 , 𝑡𝐶 ] where only 𝐷
honors its margin requirements. In this interval, 𝐷 can match its contractually stipulated
amounts 𝑐(𝑡𝑖 ) only when they involve transfers from 𝐷 to 𝐶. This asymmetry results in 𝐷
holding at time 𝑡 the smallest collateral computed in the observation interval [𝑡𝐶 , 𝑡𝐷 ] i.e.

𝑚𝑖𝑛
𝐾(𝑡) = 𝑐(𝑡𝑖 )
𝑡𝑖 ∈ [𝑡𝐶 , 𝑡𝐷 ]

5. Implications of the Collateral Date Choice: The worst-case form of the above provides a less
optimistic view on available collateral than on classical modeling, resulting in larger
exposure whenever there are multiple collateral observation dates in [𝑡𝐶 , 𝑡𝐷 ] – it is assumed
that one of the observation dates 𝑡𝑖 always coincides with the start of the MPoR 𝑡𝐶 . All other
things being equal, the difference in exposure relative to the classical model will increase
with 𝛿𝐶 − 𝛿𝐷 . If 𝛿𝐶 − 𝛿𝐷 is kept constant, the difference will increase with more frequent re-
margining. Note that

𝑚𝑖𝑛
𝐾(𝑡) = 𝑐(𝑡𝑖 )
𝑡𝑖 ∈ [𝑡𝐶 , 𝑡𝐷 ]

matches

𝐾(𝑡) = 𝑐(𝑡𝐶 )

when

𝛿𝐶 = 𝛿𝐷

50
Unpaid Trade Flows and Trade Flow Gap

1. Origin of the Trade Flow Gap: According to the assumptions in the earlier sections, the last
date when 𝐶 is still paying the trade flows is

𝑡𝐶 ′ = 𝑡 − 𝛿𝐶 ′

and the last date when 𝐷 is still trade flows is

𝑡𝐷 ′ = 𝑡 − 𝛿𝐷 ′ ≥ 𝑡𝐶 ′

The period when 𝐷 is still paying trade flows while 𝐶 does not is referred to as the trade flow
gap.
2. Projecting the Unpaid Trade Flow Value: The value of the net trade flows unpaid by the
termination date 𝑡 can be expressed using the notation established so far as

𝑈𝑇𝐹(𝑡) = 𝑇𝐹𝐶→𝐷 (𝑡; (𝑡𝐶 ′, 𝑡]) + 𝑇𝐹𝐷→𝐶 (𝑡; (𝑡𝐷 ′, 𝑡])

= 𝑇𝐹𝐶→𝐷 (𝑡; (𝑡𝐶 ′, 𝑡𝐷 ′]) + 𝑇𝐹𝑁𝐸𝑇 (𝑡; (𝑡𝐷 ′, 𝑡])

where an arrow indicates the direction of the trade flows and 𝐶 → 𝐷 (𝐷 → 𝐶) trade flows
have positive (negative) sign.
3. Same Date Dealer/Client Flows: In calculating the above equation, care needs to be taken on
how the trade flows are aggregated and accrued to the termination date 𝑡. Cash flows of
opposite direction scheduled to be paid in the same currency on the same date (for instance,
the two legs of an ordinary single-currency interest-rate swap) in the period (𝑡𝐷 ′, 𝑡] are
aggregated (netted) at the cash flow date, therefore only the aggregated amount – their
difference – enters in to the above equation. The aggregated amount of the missed cash flows
should be accrued to time 𝑡 at the interest rate of the currency in question, and then converted
to 𝐷’s domestic currency.

51
4. Different Currency Dealer/Client Flows: Cash flows in opposite direction scheduled to be
paid in different currencies on the same date (for instance, the two legs of a cross-currency
interest rate swap) are not netted at the cash flow date. The missed cash flow amounts in each
currency should be accrued to time 𝑡 at the relevant interest rates, and then converted to 𝐷’s
domestic currency.
5. Physically Settled Dealer/Client Flows: The value of each asset flow (for instance, a swap
that would result from exercising a physically settled swaption) should be obtained through
pricing at time 𝑡 of the undelivered asset in 𝐷’s domestic currency. Generally, asset flows are
not aggregated.
6. Collateral available at Termination Date: To analyze the impact of these assumptions on the
expression for 𝑈𝑇𝐹(𝑡) above, consider for simplicity a zero-threshold margin agreement
with no MTA/rounding. Then, from

𝑚𝑖𝑛
𝐾(𝑡) = 𝑐(𝑡𝑖 )
𝑡𝑖 ∈ [𝑡𝐶 , 𝑡𝐷 ]

the collateral available to 𝐷 at the termination date can be written as

𝐾(𝑡) = 𝑉(𝑡𝐶𝑂𝐿 )

𝑚𝑖𝑛
𝑡𝐶𝑂𝐿 = 𝑉(𝑡𝑖 )
𝑡𝑖 ∈ [𝑡𝐶 , 𝑡𝐷 ]

7. Corresponding CSA Implied Client Exposure: Substituting

𝑈𝑇𝐹(𝑡) = 𝑇𝐹𝐶→𝐷 (𝑡; (𝑡𝐶 ′, 𝑡𝐷 ′]) + 𝑇𝐹𝑁𝐸𝑇 (𝑡; (𝑡𝐷 ′, 𝑡])

and

𝐾(𝑡) = 𝑉(𝑡𝐶𝑂𝐿 )

52
 𝑚𝑖𝑛
𝑡𝐶𝑂𝐿 = 𝑉(𝑡𝑖 )
𝑡𝑖 ∈ [𝑡𝐶 , 𝑡𝐷 ]

into

𝐸(𝑡) = [𝑉(𝑡) + 𝑈𝑇𝐹(𝑡) − 𝐾(𝑡)]+

yields, for the simple CSA considered

𝐸(𝑡) = [𝑉(𝑡) − 𝑉(𝑡𝐶𝑂𝐿 ) + 𝑇𝐹𝐶→𝐷 (𝑡; (𝑡𝐶 ′, 𝑡𝐷 ′]) + 𝑇𝐹𝑁𝐸𝑇 (𝑡; (𝑡𝐷 ′, 𝑡])]+

8. CSA Implied Client Exposure Components: 𝐸(𝑡) above implies that trading flows from 𝐷 to
𝐶 can occur within the MPoR. Thus, trade flows have the potential to generate large spikes in
the exposure profiles – especially in the presence of a trade flow gap where only 𝐷 pays trade
flows. To see this, the exposure components of 𝐸(𝑡) can be further drilled down as follows.
First, ignoring the minor discounting effects inside the MPoR, the portfolio value at time
𝑡𝐶𝑂𝐿 can be represented as a sum of the portfolio’s forward value 𝑉𝐹 to time 𝑡 and the value
of all the trade flows taking place after 𝑡𝐶𝑂𝐿 and up to and including 𝑡:

𝑉(𝑡𝐶𝑂𝐿 ) = 𝑉𝐹 (𝑡𝐶𝑂𝐿 ; 𝑡) + 𝑇𝐹𝑁𝐸𝑇 (𝑡𝐶𝑂𝐿 ; (𝑡𝐶𝑂𝐿 , 𝑡])

9. Trade Flow Post Collateral Date: This may be further re-written as

𝑇𝐹𝑁𝐸𝑇 (𝑡𝐶𝑂𝐿 ; (𝑡𝐶𝑂𝐿 , 𝑡])

= 𝑇𝐹𝑁𝐸𝑇 (𝑡𝐶𝑂𝐿 ; (𝑡𝐶𝑂𝐿 , 𝑡𝐶 ′]) + 𝑇𝐹𝐶→𝐷 (𝑡𝐶𝑂𝐿 ; (𝑡𝐶 ′, 𝑡𝐷 ′])
+ 𝑇𝐹𝐷→𝐶 (𝑡𝐶𝑂𝐿 ; (𝑡𝐶 ′, 𝑡𝐷 ′]) + 𝑇𝐹𝑁𝐸𝑇 (𝑡𝐶𝑂𝐿 ; (𝑡𝐷 ′, 𝑡])

which, together with

𝑉(𝑡𝐶𝑂𝐿 ) = 𝑉𝐹 (𝑡𝐶𝑂𝐿 ; 𝑡) + 𝑇𝐹𝑁𝐸𝑇 (𝑡𝐶𝑂𝐿 ; (𝑡𝐶𝑂𝐿 , 𝑡])

53
 allows re-stating

𝐸(𝑡) = [𝑉(𝑡) − 𝑉(𝑡𝐶𝑂𝐿 ) + 𝑇𝐹𝐶→𝐷 (𝑡; (𝑡𝐶 ′, 𝑡𝐷 ′]) + 𝑇𝐹𝑁𝐸𝑇 (𝑡; (𝑡𝐷 ′, 𝑡])]+

in the following form:

𝐸(𝑡) = {𝑉(𝑡) − 𝑉𝐹 (𝑡𝐶𝑂𝐿 ; 𝑡) + 𝑇𝐹𝐶→𝐷 (𝑡; (𝑡𝐶 ′, 𝑡𝐷 ′]) − 𝑇𝐹𝐶→𝐷 (𝑡𝐶𝑂𝐿 ; (𝑡𝐶 ′, 𝑡𝐷 ′])
+ 𝑇𝐹𝑁𝐸𝑇 (𝑡; (𝑡𝐷 ′, 𝑡]) − 𝑇𝐹𝑁𝐸𝑇 (𝑡𝐶𝑂𝐿 ; (𝑡𝐷 ′, 𝑡]) − 𝑇𝐹𝑁𝐸𝑇 (𝑡𝐶𝑂𝐿 ; (𝑡𝐶𝑂𝐿 , 𝑡𝐶 ′])
− 𝑇𝐹𝐷→𝐶 (𝑡𝐶𝑂𝐿 ; (𝑡𝐶 ′, 𝑡𝐷 ′])}+

10. The Market Driven Portfolio Change 𝑡𝐶𝑂𝐿 → 𝑡: The terms in

𝐸(𝑡) = [𝑉(𝑡) − 𝑉(𝑡𝐶𝑂𝐿 ) + 𝑇𝐹𝐶→𝐷 (𝑡; (𝑡𝐶 ′, 𝑡𝐷 ′]) + 𝑇𝐹𝑁𝐸𝑇 (𝑡; (𝑡𝐷 ′, 𝑡])]+

have been re-arranged into five separate components, corresponding to the five different
contributions to the exposures. The first is

𝑉(𝑡) − 𝑉𝐹 (𝑡𝐶𝑂𝐿 ; 𝑡)

- the change of the portfolio forward value to time 𝑡 driven by the change in the market
factors between 𝑡𝐶𝑂𝐿 and 𝑡. This term is driven by the volatility of the market factors between
𝑡𝐶𝑂𝐿 and 𝑡; it produces no spikes in the expected exposure profile.
11. The Market Driven Trade Flow 𝑡𝐶 ′ → 𝑡𝐷 ′:

𝑇𝐹𝐶→𝐷 (𝑡; (𝑡𝐶 ′, 𝑡𝐷 ′]) − 𝑇𝐹𝐶→𝐷 (𝑡𝐶𝑂𝐿 ; (𝑡𝐶 ′, 𝑡𝐷 ′])

represents the change of the value of the trade flows scheduled to be paid – but actually
unpaid – by 𝐶 in the interval (𝑡𝐶 ′, 𝑡𝐷 ′] resulting from the change of the market factors

54
 between 𝑡𝐶𝑂𝐿 and 𝑡. This term is driven by the volatility of the market factors between 𝑡𝐶𝑂𝐿
and 𝑡; it produces no spikes in the expected exposure profiles.
12. The Market Driven Trade Flow 𝑡𝐷 ′ → 𝑡:

𝑇𝐹𝑁𝐸𝑇 (𝑡; (𝑡𝐷 ′, 𝑡]) − 𝑇𝐹𝑁𝐸𝑇 (𝑡𝐶𝑂𝐿 ; (𝑡𝐷 ′, 𝑡])

is the change of the value of the net trade flows between 𝐶 and 𝐷 scheduled to be paid – but
actually unpaid – in the interval (𝑡𝐷 ′, 𝑡] resulting from the change in the market factors
between 𝑡𝐶𝑂𝐿 and 𝑡. This term likewise produces no spikes in the expected exposure profile.
13. Net Trade Flow between 𝑡𝐶𝑂𝐿 → 𝑡𝐶 ′:

−𝑇𝐹𝑁𝐸𝑇 (𝑡𝐶𝑂𝐿 ; (𝑡𝐶𝑂𝐿 , 𝑡𝐶 ′])

is the negative value of the net trade flow between 𝐶 and 𝐷 scheduled to be paid – and
actually paid – in the interval (𝑡𝐶𝑂𝐿 , 𝑡𝐶 ′]. Paths where 𝐷 is the net payer – so that 𝑇𝐹 is
negative – contribute to the upward spikes in the EE profile.
14. Dealer Trade Flow across 𝑡𝐶 ′ → 𝑡𝐷 ′: The negative value of the trade flows scheduled to be
paid – and actually paid – by 𝐶 to 𝐷 in the interval (𝑡𝐶 ′, 𝑡𝐷 ′]. Whenever such trade flows are
present 𝐷 is always the payer, leading to upward spikes in the EE profile. Furthermore, in
some cases, the spikes arising from this term can be of extreme magnitude, e.g., the
scheduled notional exchange in a cross-currency swap where 𝐷 pays the full notional, but
receives nothing.

Numerical Examples

55
1. Illustrative Exposures and CVA Magnitudes: To gain intuition for the model, this section
presents exposure profiles and CVA metrics for several trade and portfolio examples, using
both the Aggressive and the Conservative calibrations. The focus is on ordinary and cross-
currency swaps, as these instruments are the primary sources of exposures in most dealers.
For all the numerical examples, the stochastic yield curves are driven by one-factor Hull-
White model; for cross-currency swap examples, the FX rate is assumed to follow a Black
Scholes model.
2. Single Swap Classical/Conservative Exposures: Andersen, Pykhtin, and Sokol (2017) first
examine how the model differs from the classical exposure approach. They use Monte Carlo
simulation on a USD 10 MM 1-year par-valued vanilla interest rate swap to compare
exposures of the Conservative calibration with those computed against the Classical+ and the
Classical- models – see the Table above. To make comparisons more meaningful, they
override the default setting of 10 business days for the Classical model, and instead set it
equal to 15 business days – the length of the MPoR for the Conservative calibration.
3. Single Swap - Classical Exposure Estimations: As they note in their figures, the Classical-
calibration is, of course, the least conservative setting, as it ignores both the effect of trade
flows and that of the margin asymmetry. The Classical+ calibration tracks the Classical-
calibration at most times, but contains noticeable spikes around the last 3 cash flow dates. No
spikes occur in their figure on the first quarterly cash flow date, as they assume that the
floating rate is fixed at the fixed rate, making the net cash flow zero in all scenarios.
4. Single Swap Conservative Model - MPoR End: The conservative calibration results also
contain spikes around the cash flow dates, although these differ from the Classical+
calibration in several ways. First, the Conservative calibration always recognizes that there
will be a part towards the end of the MPoR (after time 𝑡𝐷 ′) where 𝐶 and 𝐷 will both have
stopped paying margin and coupons; as a result, the spikes of the Conservative calibration
start later (here: 3 business days) than those of the Classical+ calibration.
5. Single Swap Conservative Model - Trade Flows: Second, the initial part of the spike – in the
period from 𝑡𝐶 ′ to 𝑡𝐷 ′ - is substantially higher for the Conservative calibration due to the
assumption of only 𝐷 paying cash flows in this sub-period. The remainder of the spike is
comparable in height to the Classical+ spike.

56
6. Single Swap - Classical vs. Conservative: Between spikes the Conservative calibration
produces higher exposures than both the Classical- and the Classical+ methods – by around
40%. This is, of course, a consequence of the worst-case margin asymmetry mechanism in

𝑚𝑖𝑛
𝐾(𝑡) = 𝑐(𝑡𝑖 )
𝑡𝑖 ∈ [𝑡𝐶 , 𝑡𝐷 ]

the effect of which will grow with the diffusion volatility of the rate process. Of course, the
last coupon period again has no exposure between spikes, since the volatility of swap prices
vanishes after the last coupon rate fixing.
7. Single Swap: Aggressive Calibration Exposures: Comparison the Aggressive calibration to
the Classical+ and the Classical- calibrations are qualitatively similar. A detailed comparison
is therefore skipped here, but Andersen, Pykhtin, and Sokol note that the pick-up in exposure
from margin asymmetry falls to about 15%, rather than the 40% observed for the
Conservative calibration – a result of the fact that, for Aggressive calibration, the worst-case
margin result is established over much fewer days. Andersen, Pykhtin, and Sokol (2017)
demonstrate the comparison of the exposure profiles for the Aggressive and the Conservative
calibrations in their graphs; as expected, the Conservative calibration leads to both bigger and
wider exposure spikes, as well as to higher exposure levels between spikes.
8. Single Swap - Maturity/Coupon Effect: While instructive, the 1-year vanilla swap example is
quiet benign exposure-wise; not only is the instrument very short-dated, it also allows for
netting of coupons on trade-flow dates, thereby reducing the effects of trade flow spikes.
Andersen, Pykhtin, and Sokol (2017) relax both effects by increasing the maturity of the
swap, and by making the fixed and the floating legs pay on different schedules – and
illustrate the exposure results in a separate figure.
9. Single Swap - Coupon Payment Mismatch: The upward exposure spikes occur twice per
year, whenever the dealer must make a semi-annual fixed payment. On the dates when the
counterparty makes a quarterly floating payment that is not accompanied by a fixed payment
by the dealer, a narrow downward spike emerges, due to the delay in transferring the coupon
back to the counterparty through the margin mechanism.

57
10. Single Swap - Impact of Maturity: The exposure between spikes is also much larger, a
consequence of the higher volatility of the 10-year swap compared to the 1-year swap. Of
course, as the swap nears its maturity, its duration and volatility die out, so the non-spike
exposure profile predictably gets pulled to zero at the 10-year date. Also, as predicted, the
Aggressive calibration produces much lower exposures than the Conservative calibration, by
nearly a factor of 2.
11. Single Cross-Currency Swap - Herstatt Risk: A more extreme form of trade flow spikes will
occur for cross-currency swaps, where neither the coupon nor the final payment can be
netted. The notional payment, in particular, can induce a very significant payment exposure
spike (the Herstatt Risk), whenever the exposure model allows for a trade flow gap. To
recall, the Conservative calibration has a trade flow gap, but the Aggressive calibration does
not.
12. Single Cross-Currency Swap - Coupon Mismatch: As confirmed by Andersen, Pykhtin, and
Sokol (2017), the exposure for a conservative calibration has a very large spike that is not
present in the Aggressive calibration. Like Conservative calibration, Aggressive calibration
will, of course, still produce spikes at the cash flow dates, due to margin effects.
13. Single Cross-Currency Swap - Principal Mismatch: As a consequence, the principal
exchange is likely fa away from break-even, resulting in a large exposure spike at maturity.
Although smaller than for the Conservative calibration, the spike at maturity is also present
for the Aggressive calibration; while both 𝐶 and 𝐷 pay the principal exchange, 𝐶 does not
make the margin transfer for the balance of the principal payments.

Portfolio Results

1. Single Swap Portfolio: Setup Overview: For individual trades, the presence of localized
spikes in the exposure profiles may ultimately have a relatively modest impact on the credit
risk metrics, such as the CVA – after all, the likelihood of the counterparty default in a

58
 narrow time interval around quarterly or semi-annual cash flow event is typically low. For a
portfolio of swaps, however, the spikes will add up and affect the net exposure profile nearly
everywhere.
2. Single Swap Portfolio - Draw Algorithm: To illustrate this, Andersen, Pykhtin, and Sokol
(2017) picked 50 interest rate swaps with quarterly floating rate payments and semi-annual
fixed rate payments of 2%. The terms of the swap were randomized as follows:
a. Notionals of the swap are sampled uniformly on the interval from 0 to USD 1 MM.
b. Duration of the fixed leg payments – payer or receiver – is random.
c. Start date of each swap is subject to a random offset to avoid complete MPoR
overlaps.
d. Swap maturities are scaled uniformly on the interval from 1 to 10 years.
3. Single Swap Portfolio: Aggressive vs Conservative: They also illustrate the resulting
exposure profile in a separate figure. Both the Conservative profile, and to a lesser extent, the
Aggressive profile include frequent spikes around the trade-flow times above the baseline
exposure level. As seen in the next section, these spikes make a significant contribution to
the CVA metrics. As before, the exposure under the Conservative calibration is twice as large
as that under the Aggressive calibration.
4. XCCY Swaps Portfolios Generation Algorithm: To repeat the portfolio results with a cross-
currency swap, Andersen, Pykhtin, and Sokol (2017) constructed a 50-deal portfolio by
randomization, using the following rules.
a. EUR notionals are sampled uniformly in the interval from 0 to USD 10 MM
b. USD notionals are 1.5 times the EUR notionals
c. EUR leg has a fixed semi-annual coupon of 3%, and the USD leg floating quarterly
coupon
d. Direction of the fixed leg payments (payer or receiver) is random
e. Start date of each swap is subject to a random offset to avoid complete MPoR
overlaps
f. Swap maturities are sampled uniformly in the interval from 1 to 10 years
5. XCCY Swap Portfolio: Conservative vs. Aggressive: As shown in their figure for an
expected exposure for a 10Y cross-currency swap, Andersen, Pykhtin, and Sokol (2017)

59
 generated the swaps within the portfolio such that the principal exchanged and the fixed
coupon are not at-the-money, to mimic a typical situation corresponding to a portfolio of
seasoned trades. As demonstrated in another figure for the expected exposure of the cross-
currency swap portfolio, the exposure for the conservative calibration is, as expected,
dominated by a series of Herstatt risk spikes, one per swap in the portfolio.

CVA Results

1. CVA Computation from Expected Exposure: As mentioned earlier, a common use of the
expected exposure results is the computation of the CVA. Under suitable assumptions, 𝐷’s
unilateral CVA may be computed from the expected exposure (EE) profile as

𝐶𝑉𝐴 = (1 − 𝑅) ∫ 𝑃(𝑢 + 𝛿𝐶 ′)𝐸𝐸(𝑢 + 𝛿𝐶 ′)𝑑𝑋(𝑢)

0

where 𝑅 is the recovery rate, 𝑃(𝑡) is the time-0 discount factor to time 𝑡, and 𝑋(𝑡) is the
time-0 survival probability of 𝐶 to time 𝑡. As discussed before, the exposure profile here is
offset by 𝛿𝐶 ′ to properly align it with the default events.
2. CVA Metrics Dealer/Client Settings: The CVA metric serves as a convenient condensation
of the exposure profiles of the previous two sections into single numbers, and Andersen,
Pykhtin, and Sokol (2017) tabulate the CVA numbers for the corresponding
instruments/portfolios. The CVA integral was discretized using a daily grid, assumed at

𝑅 = 40%

and the forward default intensity is left constant at 2.5% such that

60
 𝑋(𝑡) = 𝑒 −0.025𝑡

For reference, the table also includes the results of the Classical method, with the MPoR
length equal to both that of the Aggressive calibration (7 𝐵𝐷) and the Conservative
Calibration (15 𝐵𝐷).
3. CVA Comparison - Classical/Conservative/Aggressive: Their results confirm what was seen
earlier. For instance, the CVA for the Aggressive calibration is 50% to 70% smaller than
that for the Conservative calibration. In Addition, the CVA of the Conservative calibration is
between 50% and 100% larger than that of the Classical+ calibration – at similar MPoR –
which in turn is larger than the CVA for the Classical- calibration by around 5% to 25%. Not
surprisingly the CVA results for the XCCY portfolio are particularly high in the
Conservative calibration due to the Herstatt risk.

Improvement of the Computation Times

1. Computation Speed-Up using Coarse Grids: In exposure calculations for realistic portfolios,
horizons can be very long, often exceeding 30 years. For such lengthy horizons, brute-force
Monte-Carlo exposures on a daily, or even weekly, time grid will often be prohibitively slow.
It is therefore common to use daily simulation steps only for the earliest parts of the exposure
profile (e.g., the first month), and then gradually increase the step-length over time to
monthly or quarterly, in order to keep the total number of simulation dates manageable.
Unfortunately, such a coarsening of the time-grid will inevitably fail to capture both the
worst-case margin effect and the trade spikes that are key to the exposure model.
2. Coarse Grid Lookback Analysis: The next two sections look at ways to capture exposure
without having to resort to brute-force daily simulation. A common speed-up technique for
the Classical model – the Coarse Grid Look-back Model – is first reviewed, and its
61
 shortcomings and pitfalls are highlighted. An improved practical technique based on
Brownian Bridge is the proposed.

The Coarse Grid Lookback Method and its Shortcomings

1. Layout of the Coarse Grid: Assume that the portfolio is not computed daily, but instead on a
coarse grid {𝑠𝑗 } where 𝑗 runs from 1 to 𝐽. This section uses 𝑠 rather than 𝑡 to distinguish the
model grid from the daily margin calculation grid.
2. Points on the Coarse Grid: In the classical model, the collateral depends only on the portfolio
value at the start and at the end of the MPoR, i.e., 𝑠𝑗 − 𝛿 and 𝑠𝑗 , as is seen from

𝐸(𝑡) = [𝑉(𝑡) − 𝐾(𝑡)]+

𝐾(𝑡) = 𝑐(𝑡 − 𝛿)

𝐾(𝑡) = [𝑉(𝑡 − 𝛿) − ℎ𝐶 (𝑡 − 𝛿)]+ − [−𝑉(𝑡 − 𝛿) − ℎ𝐷 (𝑡 − 𝛿)]+

where MPoR is usually around

𝛿 = 10 𝐵𝐷

62
 for CSA’s with daily margining. To achieve acceptable computational performance, the time
step of the coarse model grid 𝑠𝑗 − 𝑠𝑗−1 must be significantly greater than the length of the
MPoR. This, however, would preclude one from establishing a portfolio value at 𝑠𝑗 − 𝛿.

3. Introducing a Lookback Node: The coarse grid lookback method deals with this issue by
simply adding a second lookback time point 𝑠𝑗 − 𝛿 to all primary measurement times 𝑠𝑗 , in
effect replacing each node on the coarse model grid by a pair of closely spaced nodes. For
each simulated portfolio path, the portfolio value at the lookback point is then used to
determine the collateral available at the corresponding primary time point.
4. Slowdown due to the Lookback: The Coarse Grid Lookback Scheme causes, at worst, a
factor of × 2 slowdown relative to valuing the portfolio once per node of the Coarse model
grid. If even a × 2 performance loss is not acceptable, a Brownian Bridge constructed
between the primary coarse grid nodes can be used to interpolate the value of the portfolio at
each lookback point, see, for example, Pykhtin (2009). Notice that the use of the Brownian
Bridge for this purpose should not be confused with its use in the next section.
5. Shortcoming of the Model: The Coarse Grid Lookback method is a common way of
addressing the mismatch between the long time-step of the coarse model grid and the much
shorter MPoR. Similar to the commonly used models of uncollateralized exposure, the
method produces accurate – with respect to the underlying assumptions of the Classical
model – exposure numbers at the coarse grid time points, but provides no information on the
exposure between the grid points.
6. Collateralized vs. Uncollateralized Grid Exposures: For uncollateralized positions, the
exposure profiles are reasonably smooth, so one can safely interpolate between the grid
points for calculating integral quantities, such as the CVA. In collateralized case, however,
one cannot rely on such interpolations because the true exposures, as has been seen above, is
likely to have spikes and jumps between the grid points. The Coarse Grid Lookback method
has no means to determining the position or the magnitude of the irregularities between the
grid points, and thus, is not suitable for CVA or capital calculations.
7. Classical+ Model - Coarse Grid Impact: To briefly expand on this, consider the Classical+
version of the classical model. Here it is assumed that all trade flows are paid within the

63
 MPoR, where, as shown before, trade flows often result in exposure spikes. Exposure
profiles computed from daily time steps would consequently show spikes from all trade
flows until the maturity of the portfolio.
8. Trade Flows outside the Simulated MPoR’s: In contrast, in a typical implementation with
sparsely spaced MPoR’s, only trade flows that happen to be within a sparsely simulated
MPoR’s may result in spikes; the exposure profile would then miss all other flows.
9. Simulation Calendar Impact on Exposure: Furthermore, as the location of the simulation
point will likely change with the advancement in the calendar time, trade flows would move
in and out of the simulated MPoR’s, and the exposure profile one report on any given day
may very well differ significantly from those that were reported the day before. This in turn
causes CVA or risk capital to exhibit significant, and entirely spurious, oscillations.
10. Classical- Model - Coarse Grid Impact: While the Classical- exposure model does not exhibit
outright spikes its exposure profiles still exhibit jumps around significant trade flows. The
classical coarse-grained implementation would not be able to resolve the position of these
jumps, instead only showing the conservative jumps between two exposure measurement
points often separated by many months. This creates another source of instability, present in
both the Classical- and the Classical+ versions of the classical model.
11. Illustration using the Forward CVA: To illustrate the effects described above, Andersen,
Pykhtin, and Sokol (2017) define the concept of time 𝑡 forward CVA, denoted 𝐶𝑉𝐴𝑡 ,
obtained by
a. Changing the lower integration limit in

𝐶𝑉𝐴 = (1 − 𝑅) ∫ 𝑃(𝑢 + 𝛿𝐶 ′)𝐸𝐸(𝑢 + 𝛿𝐶 ′)𝑑𝑋(𝑢)

0

from 0 to 𝑡, and
b. Dividing the result by 𝑃(𝑡)𝑋(𝑡).
Using the same portfolio of 50 EUR-USD cross-currency swaps, they show the 𝑡-dependence
of 𝐶𝑉𝐴𝑡 on a daily grid to portfolio maturity.

64
12. Spurious Oscillations from Moving Windows: As CVA is an integral of exposures, spikes in
exposure profile profiles should result in jumps rather than oscillations in 𝐶𝑉𝐴𝑡 . However,
when one of the Coarse Grid Lookback method’s sparsely located MPoR window moves past
a large trade flow, the contribution to the CVA temporarily increases only to drop back when
the window moves past the large trade flow. As illustrated by Andersen, Pykhtin, and Sokol
(2017), such oscillations are spurious and their presence is highly unattractive when CVA is
computed and reported as part of daily P&L.

Brownian Bridge Method

1. Brute Force Portfolio Value Simulation: Overcoming the deficiencies outlined in the
previous section is, unfortunately, prohibitively expensive for large portfolios, mostly due to
the expense of repricing the entire portfolio at each simulation path and each observation
date.
2. Daily Simulation of Risk Factors: On the other hand, merely simulating the risk factors at a
daily resolution is generally feasible, as the number of the simulated risk factors is typically
relatively small (i.e., several hundred) and the equations driving the risk factor dynamics are
usually simple.
3. Generation of Daily Trade Flows: Furthermore, having produced risk factors on a daily grid,
one can normally also produce all realized trade flows along each path because trade flows,
unlike trade prices, are usually simple functions of the realized risk factors.
4. Risk Factors under Daily Resolution: Based on these observations, Andersen, Pykhtin, and
Sokol (2017) propose the following algorithm for generating paths of portfolio values and
trade flows on a daily time grid. First, simulate paths of market risk factors with daily
resolution.
5. Trade Flow under Daily Resolution: For each path 𝑚, use the simulated market risk factors
to calculate trade flows on the path with daily resolution.

65
6. Coarse Grid Path Portfolio Valuation: For each path 𝑚 and each coarse portfolio valuation
time point 𝑠𝑗 (𝑗 = 1, ⋯ , 𝐽) use the simulated risk factors to calculate portfolio value on the
path 𝑉𝑚 (𝑠𝑗 )
7. Trade Flow Adjusted Forward Value: For each path 𝑚 and each time point 𝑠𝑗 use the trade
flows realized on the path between times 𝑠𝑗−1 and 𝑠𝑗 to calculate the forward to 𝑠𝑗 portfolio
value 𝑉𝑚 ′(𝑠𝑗−1 ; 𝑠𝑗 ):

𝑉𝑚 ′(𝑠𝑗−1 ; 𝑠𝑗 ) = 𝑉𝑚 (𝑠𝑗−1 ) − 𝑇𝐹𝑚,𝑁𝐸𝑇 (𝑠𝑗 ; (𝑠𝑗−1 , 𝑠𝑗 ])

Note that 𝑉𝑚 ′(𝑠𝑗−1 ; 𝑠𝑗 ) is not a true forward value because the realized trade flows are
subtracted from the 𝑠𝑗−1 portfolio value rather than the true forward value being calculated at
time 𝑠𝑗−1 .
8. Portfolio Value Local Variance Estimation: For each path 𝑚 and each portfolio measurement
time point 𝑠𝑗 compute the local variance 𝜎𝑚 2 (𝑡𝑗−1 ) for the portfolio value diffusion 𝑉𝑚 (𝑠𝑗 ) −
𝑉𝑚 ′(𝑠𝑗−1 ; 𝑠𝑗 ) via a kernel regression estimator – e.g., the Nadaraya-Watson Gaussian kernel
estimator (Nadaraya (1964), Watson (1964)) conditional on the realized value of
𝑉𝑚 ′(𝑠𝑗−1 ; 𝑠𝑗 ). The selection of bandwidth for the kernels is covered in, e.g., Jones, Marron,
and Sheather (1996). In their numerical results, Andersen, Pykhtin, and Sokol (2017) use the
Silverman’s Rule of Thumb (Silverman (1986)). The term diffusion is used to indicate that the
portfolio value change has been defined to avoid any discontinuities resulting from trade
flows.
9. Brownian Bridge Local Interpolation Scheme: For each path 𝑚 and each exposure
measurement time point 𝑠𝑗 , simulate an independent, daily sampled, Brownian Bridge
process (see, for instance, Glasserman (2004)) that starts from the value 𝑉𝑚 ′(𝑠𝑗−1 ; 𝑠𝑗 ) at time
𝑠𝑗−1 and ends at the value 𝑉𝑚 (𝑠𝑗 ) at time 𝑠𝑗 . The volatility of the underlying Brownian
motion should be set to 𝜎𝑚 (𝑠𝑗−1 ).
10. Brownian Bridge Portfolio Value Approximation: For each path 𝑚 and each exposure
measurement time point 𝑠𝑗 , the portfolio values for each time 𝑢 of the daily grid in the

66
 interval (𝑠𝑗−1 , 𝑠𝑗 ) are approximated from the simulated Brownian bridge 𝐵𝐵𝑚 (𝑢) by adding
the trade flows realized along the path 𝑚 between the times 𝑢 and 𝑠𝑗 :

𝑉𝑚,𝐴𝑃𝑃𝑅𝑂𝑋 (𝑢) = 𝐵𝐵𝑚 (𝑢) + 𝑇𝐹𝑚,𝑁𝐸𝑇 (𝑠𝑗 ; (𝑢, 𝑠𝑗 ])

11. Rational behind Brownian Bridge Methodology: In a nutshell, the algorithm above uses a
Brownian bridge process to interpolate portfolio values from a coarse grid in a manner that
ensures that intermediate trade flow events are handled accurately. The algorithm produces
paths of portfolio values and trade flows in a daily time grid, wherefore exposure can be
calculated as described earlier with daily resolution and overlapping MPoR’s. Furthermore,
daily sampling allows for further refinements of the proposed model by consistently
incorporating thresholds, minimum transfer amount, and rounding.
12. Brownian Bridge Portfolio Weiner Increment: A key assumption made by the Brownian
Bridge algorithm is that the portfolio value process within the interpolation interval is a
combination of an approximately normal diffusion overlaid by trade flows. For Weiner
process models without risk factor jumps, this approximation is accurate in the limit of
infinitesimal interpolation interval, and is often a satisfactory approximation for monthly or
even quarterly interpolation steps.
13. Brownian Bridge Approximation Error #1: Nevertheless, the presence of trade flows that
depend on the values of the risk factors between the end points introduces two types of
errors. Suppose that there is a trade flow at an end point that depends on the risk factor value
at the date when it is paid. The independence of the Brownian Bridge process from the risk
factor processes that drive that trade flow would result in an error in the expected exposure
profile around the trade flow date. This error is largest for trade flows in the middle of the
interpolation interval and disappears for trade flows near the ends of the interval.
14. Brownian Bridge Approximation Error #2: Suppose that there is a trade flow that occurs at
the end point of an interpolation interval, but whose values depend entirely on the realization
of the risk factor within the interpolation interval. A typical example would be a vanilla
interest rate swap where the floating leg payment being paid at the end of the interpolation
interval depends on the interest rate on a date within the interval. Even in the absence of a

67
 trade flow within the interpolation interval, the volatility of the swap value drops at the
floating rate fixing date as some of the uncertainty is resolved. Thus, the true swap value
process has two volatility values; a higher value before the rate fixing date and a lower value
after the rate fixing date. In contrast the approximation algorithm assumes a single value of
volatility obtained via kernel regression between the end points. Similar to the de-correlation
error discussed above, the error resulting from this volatility mismatch is largest for fixing
dates in the middle of the interpolation interval and disappears for fixing dates near the end
points.
15. Trade Flow at Mid-Interval: To illustrate the two errors above, Andersen, Pykhtin, and Sokol
(2017) compute the expected exposure profile for a one-year interest rate swap when a
monthly grid for full valuation is situated so that the payments/fixing dates sit roughly in the
middle of the interpolation interval, thus maximizing the error of the Brownian Bridge
algorithm.
16. Unbiased Nature of the Error: While there are, as expected, some error around the trade flow
dates, they are acceptable in magnitude and overall unbiased, in the sense that the over-
estimation of the exposure is about as frequent as the under-estimation of the exposure. For,
say, CVA purposes, the Brownian Bridge results would therefore be quite accurate.
17. Trade Flows at Interval End: Andersen, Pykhtin, and Sokol (2017) also compute the expected
exposure profiles when the monthly valuation points are aligned with the rate fixing/payment
dates. In this case, Brownian Bridge approximation is nearly exact.
18. Choice of Valuation Grid Location: Of course, in practice such alignment is only possible for
a single trade or a small netting set, and not for large portfolios where trade flows will occur
daily. Yet, even for large netting sets the calculation accuracy will improve if the
interpolation pillars are aligned with the largest trade flows (e.g., principal exchange dates for
the largest notional amounts). In practice, errors can be typically expected to be somewhere
between the two extremes discussed above.
19. Performance Gains from Brownian Bridge: While the exact speed up provided by the
Brownian Bridge method depends on the implementation, for most portfolios the overhead of
building the Brownian Bridge at a daily resolution is negligible compared to computing the
exposure on the model’s coarse grid.

68
20. Comparison with Coarse Grid Lookback: In this case, the computational effort of the daily
Brownian Bridge method is about half the computational effort of the Coarse Grid Lookback
method, as the former does not require adding a lookback point to each of the primary coarse
grids. Thus, the Brownian Bridge is both faster and significantly more accurate than the
Standard Coarse Grid Lookback method.

Initial Margin

1. Role of IM: Extra Protection: The posting of initial margin (IM), in addition to the regular
variation margin collateral (VM), provides dealers with a mechanism to gain additional
default protection. The practice of posting IM has been around for many years, typically with
IM being computed on trade inception on a trade level basis.
2. Modeling Static Initial Margin Exposure: This type of IM is entirely deterministic and
normally either stays fixed over the lifetime of a trade or amortizes down according to a pre-
specified schedule. As a consequence, modeling the impact on the exposure is trivial; for the
exposure points of interest all trade level IM amounts are summed across the netting set and
the total – which is the same for all paths – is subtracted on the portfolio value from each
path.
3. Dynamically Refreshed Initial Margin (DIM): A more interesting type of IM is dynamically
refreshed to cover portfolio-level close-out risk at some high percentile, often 99%. This type
of margin is routinely applied by Clearinghouses (CCPs) and by margin lenders, and will
also soon be required by regulators for inter-dealer OTC transactions.
4. BCBS IOSCO Initial Margin Rules: In particular, in 2015 BCBS and IOSCO issued a final
framework on margin requirements (BCBS and IOSCO (2015)) under which two covered
entities that are counterparties in non-centrally cleared derivatives are required to:
a. Exchange VM under a zero-threshold margin agreement, and
b. Post IM to each other without netting the amounts.

69
 Covered entities include all financial firms and systematically important non-financial firms.
Central banks and sovereigns are not covered entities.

5. Third Party Management of IM: IM must be held in a default remote way, e.g., by a
custodian, so that IM posted by the counter-party should be immediately available to it
should the other counter-party default.
6. Internal Model/Standardized Schedule IM: Under the BCBS and IOSCO rules, regulatory
VM can be calculated by an internal model or by lookup in a standardized schedule.
7. Internal Models Based IM Calculation: If an internal model is used, the calculation must be
made at the netting set level as the value-at-risk at the 99% confidence level. The horizon
used in this calculation equals 10 business days for daily exchange of VM or 9 business days
plus a re-margining period for less frequent exchange of VM.
8. Denial of Cross-Asset Netting: Diversification across distinct asset classes is not recognized,
and the IM internal model must be calibrated to a period of stress of each of the asset classes.
9. Handling Adjustments to the IM: The required levels of IM are changed as the cash flows are
paid, new trades are booked, or markets move. To accommodate this, dealers would call for
more IM or return the excess IM.
10. Complexities Associated with the IM Estimation: For trades done with CCPs or under the
new BCBS-IOSCO rules, one must find a way to estimate the future IM requirements for
each simulated path. No matter how simple the IM VaR model is, it will likely be difficult to
perform such calculations in practice if one wants to incorporate all the restrictions and twists
of the IM rules; stress calibration, limited diversification allowance, and, for CCP’s, add-ons
for credit downgrades and concentration risk.
11. Estimating Simplified Version of IM: However, it is possible to utilize the model in this
chapter to calculate the counter-party exposures if one ignores these complications. Note that
ignoring such complications is conservative, as it will always lead to a lower of IM, and
therefore, to a higher level of exposure.
12. 𝑡𝐶 as IM Delivery Date: To calculate the exposure at time 𝑡 the assumption here is that the
last observation date for which C would deliver VM to D is

𝑡𝐶 = 𝑡 − 𝛿𝐶

70
 It is reasonable to assume that this date is also the last date at which C would deliver IM to a
custodian.
13. Simplified IM Mechanics Timeline: To simplify modeling, it is assumed that the custodian
would not return any amount to C for observation dates after 𝑡 − 𝛿𝐶 . Thus, to calculate
exposure at time 𝑡, IM on a path has to be estimated from the dynamics of the exposure
model as of time 𝑡 − 𝛿𝐶
14. 𝑡𝐶 IM Estimate using Gaussian Portfolio Evaluation: Assuming, as is common in practice,
that the portfolio values are locally Gaussian, it suffices to know the local volatility for the
portfolio value for the period [𝑡 − 𝛿𝐶 , 𝑡] estimated at 𝑡 − 𝛿𝐶 . Denoting the IM horizon by 𝛿𝐼𝑀
and the local volatility of the portfolio value at time 𝑢 on path 𝑚 via 𝜎𝑚 (𝑢), the IM available
to D at the ETD date 𝑡 on path 𝑚 is given by

𝐼𝑀𝑚 (𝑡 − 𝛿) = 𝜎𝑚 (𝑡 − 𝛿)√𝛿𝐼𝑀 Φ−1 (𝑞)

where 𝑞 is a confidence level – often 99% - and Φ−1 (∙) is the inverse of the standard normal
cumulative distribution function.
15. Kernel Regression Based Local Volatility: Estimating the local volatility can be done via
kernel regression, as in the previous section. If the portfolio value is simulated at both 𝑡 − 𝛿𝐶
and 𝑡, the kernel regression for could be run on the P&L 𝑉(𝑡) − 𝑉(𝑡 − 𝛿𝐶 ) +
𝑇𝐹𝑁𝐸𝑇 (𝑡; (𝑡 − 𝛿𝐶 , 𝑡]) conditional on the realization of the portfolio value on path 𝑚 at the
beginning of the 𝑀𝑃𝑜𝑅 𝑉𝑚 (𝑡 − 𝛿𝐶 ). If one does not calculate the portfolio value at the
beginning of the 𝑀𝑃𝑜𝑅 but uses the fast approximation outlined earlier instead, 𝜎𝑚 (𝑡 − 𝛿𝐶 )
can be set equal to the local volatility estimated for the time interval that encloses the given
[𝑡 − 𝛿𝐶 , 𝑡].
16. Brownian Bridge IM Plus VM: Thus, the Brownian Bridge framework can now produce not
only the collateralized exposure under VM alone, but also a reasonable estimate of the
collateralized exposure under a combination of VM and IM.

71
17. IM Timing and Transfer Mechanics: In calculating the IM, an important consideration is the
timing and the mechanics of the adjustment to the IM when C misses a margin flow or a
trade flow.
18. Assumption - IM Return to the Client: For instance, when a large trade reaches maturity, the
portfolio VaR may be reduced, in which case, some of the IM posted by C must be refunded.
The issue of whether this refund can be delayed due to an ongoing margin dispute is not yet
fully resolved. To simplify the calculations, it is assumed that no part of the IM is returned to
C during the 𝑀𝑃𝑜𝑅.
19. 10Y OTC Swap VM + IM EE: To show some numerical results, Andersen, Pykhtin, and
Sokol (2017) consider the individual trades and portfolios of the earlier section. They use the
case of a 10Y vanilla swap for which they calculate the impact of IM on exposure.
20. Time Horizon IM Mechanism Impact: As is evident from their calculations, the IM
mechanism strongly reduces exposures away from trade flows, but near the trade flow dates
the protection gets progressively weaker and disappears almost completely for the last couple
of trade flows. The reason for this uneven benefit of IM on this trade is that the 10 day 𝑉𝑎𝑅
of this trade bears no direct relationship to the size of the trade flows that determines the
exposure spikes in the model.
21. Inter/Intra-Spike IM Exposures: The variance of the 𝑃&𝐿 reduces as the swap approaches
maturity so that the amount of IM on a given path is also reduced. However, the size of the
trade flows is not reduced, but can actually grow with simulation time as larger and larger
realizations of the floating rates are possible. Thus, when the swap approaches maturity the
amount of IM is greatly reduced relative to the trade flows, so exposure spikes grow larger,
while the diffusion component of the exposure becomes smaller.
22. Cross-Currency Swap VM + IM EE: Andersen, Pykhtin, and Sokol (2017) compute the
impact of IM on the vanilla swap and the cross-currency swap portfolios described earlier.
As can be seen there the IM strongly suppresses the diffusion component of the portfolio
value changes, but proves inadequate in reducing the spikes of exposure for both single
currency, and especially, cross-currency portfolios.

72
Conclusion

1. Fully Collateralized Counterparty Exposure: Industry standard models for collateralized

credit risk are well-known to produce non-negligible counterparty credit risk exposure, even
under full collateralization of the variation margin. This exposure essentially arises due to the
inevitable operational and legal delays (margin period of risk, or 𝑀𝑃𝑜𝑅) that are baked into
the workings of ISDA contracts that govern OTC trading.
2. Classical Implementations of the 𝑀𝑃𝑜𝑅 : In the most common industry implementation, the
length of the 𝑀𝑃𝑜𝑅, and precisely what transpires inside it, is, however, often treated in a
highly stylized fashion. Often the 𝑀𝑃𝑜𝑅 is set equal to 10 business days for little reason
other than tradition, and often counter-parties are assumed to have oddly synchronized
behavior inside the 𝑀𝑃𝑜𝑅.
3. The Classical+ and Classical- Implementations: For instance, one common approach –
denoted Classical- - assumes that the 𝑀𝑃𝑜𝑅 and the trade flows by both counter-parties
terminate at the beginning of the 𝑀𝑃𝑜𝑅, but the trade flows terminate simultaneously at the
end of the 𝑀𝑃𝑜𝑅. Surprisingly, the Classical+ and the Classical- approaches continue to co-
exists in the market, and neither has become the sole market practice.
4. Reasons for the Popularity of the Classical Models: One reason for this state of affairs is that
the two models correspond to different choices for the trade-off between implementation
complexity and the model stability; Classical+ is easier to implement but is prone to spurious
spikes in the daily CVA P&L – as demonstrated earlier – whereas Classical- is more difficult
to implement, but is free from such spikes.
5. Objectives of Well-Designed Models: Ultimately, of course, a model should be selected not
on the basis of the implementation ease or on the properties of a specific numerical
technique, but on the basis of how well the model captures the legal and the behavioral
aspects of the events around a counter-party default. The term well means different things in

73
 different applications of the exposure model. For regulatory capital purposes, prudence and
conservatism may, for instance, be as important as outright precision.
6. Inadequacies of the Classical Approach: To this end, even a cursory analysis suggests that
the perfect synchronicity of the Classical ± models cannot be supported in reality. For
instance, due to the way the CSA works in practice, the non-defaulting party will need at
least 3 days after a portfolio valuation date to determine for sure that the corresponding
margin payment by its counterparty will not be honored.
7. Detailed Analysis of 𝑀𝑃𝑜𝑅 Timeline: This chapter carefully dissects the 𝑀𝑃𝑜𝑅 into a full
timeline around the default event, starting with the missed margin call and culminating at the
post-default valuation date at which the termination value of the portfolio is established.
8. Model Parameters of the Timeline: For modeling purposes, the timeline of the model has
been condensed into 4 model parameters, each specified as the number of days prior to the
termination for the events below – in contrast the classical model has only one parameter –
the full length of the 𝑀𝑃𝑜𝑅.
9. Dealer/Client Trade/Margin Dates:
a. The last market data measurement for which the margin flow is received (𝛿𝐶 ) and
paid (𝛿𝐵 ) as prescribed.
b. The last date when the defaulting party (𝛿𝐶 ′) and the dealer (𝛿𝐷 ′) make the trade
payments as prescribed.
10. Legal Operational Basis behind the Parameters: As shown, each of these parameters has a
legal and/or operational interpretation, enabling calibration from the CSA and from the
operational setup of the dealer. Note that the proposed model parameterization includes the
Classical+ and the Classical- models as the limit cases.
11. Aggressive CSA Timeline for the 𝑀𝑃𝑜𝑅: For indicative purposes, two particular models are
described – Aggressive and Conservative. For former assumes that the non-defaulting dealer
always operates at an optimal operational level, and will enforce the legal provisions of the
ISDA legal contracts as strictly as possible.
12. Conservative CSA for 𝑀𝑃𝑜𝑅 Timeline: The latter will allow for some slack in the operations
of the dealer, to allow for manual checks of calculations, legal reviews, gaming behavior of
the counterparty, and so forth.

74
13. Aggressive/Conservative Timeline Exposure Comparison: The Conservative model setting
obviously produces higher exposures than the Aggressive setting, for the following reasons.
a. The Conservative setting has a longer overall length of 𝑀𝑃𝑜𝑅
b. The Conservative setting has a margin flow period where the dealer pays, but does
not receive, margin flows
c. The Conservative setting, unlike the Aggressive setting, contains a trade flow gap
period where the dealer pays, but does not receive, trade flows
14. Comparison of Margin Flow Exposures: In their numerical tests, Andersen, Pykhtin, and
Sokol (2017) found that the first two factors of the Conservative setting to have
approximately twice the exposure of both the Aggressive and the Classical ± settings away
from the dates of large trade flows.
15. Comparison of Trade Flow Exposures: The last factor, i.e., the presence of a large trade flow
gap may cause exposure spikes of extremely large magnitudes under the Conservative
calibration. Despite the fairly short duration of these spikes, they may easily add up to very
significant CVA contributions, especially for cross-currency trades with principal exchange –
the Herstatt risk.
16. Past Realizations of Trade Flow Default: Credit losses due to trade flow gaps materialized in
practice during the financial crisis – especially due to the Lehmann Brothers’ default – so
their incorporation into the model is both prudent and realistic.
17. Impracticality of Daily Simulation Schemes: Detailed tracking of the margin and the trade
flow payments requires the stochastic modeling of the trade portfolio on a daily grid. As
brute-force simulations at such a resolution are often impractically slow, it is important that
numerical techniques be devised to speed up the calculations.
18. Kernel Regression on Stripped Cash Flows: While the focus of this chapter was mainly on
establishing the fundamental principles for margin exposure, it also proposed an acceleration
method based on kernel regression and applied Brownian Bridge to portfolio values stripped
of cash flows.
19. Impact on Different Product Types: For ordinary and cross-currency swaps this chapter
demonstrates that this method is both accurate and much faster than either brute-force
simulation or standard acceleration techniques of the desired model. Further improvements in

75
 the acceleration techniques, and expansion of applicability into more exotic products, is an
area of future research.
20. Initial Margin – Classical- Settings Impact: Under suitable assumptions, kernel regression
may also be used to embed risk-based initial margin into exposure simulations. As
demonstrated in the final section of the chapter, initial margin at 99% exposure greatly
succeeds in reducing bilateral exposure for the Classical- calibration.
21. Initial Margin Trade Flow Impact: For all other calibration choices, and especially for the
Conservative setting, the reduction in counterparty exposure afforded by initial margin fails
around the time of large trade flows, when a sudden change of exposure following an initial
trade flow exceeds the initial margin level.
22. Initial Margin Maturity Decay Impact: Note that the already inadequate level of IM
protection deteriorates around the maturity of the portfolio, where the local volatility of the
trade flow value decreases, but the trade flows themselves do not. Overall accurate modeling
of the events within the 𝑀𝑃𝑜𝑅 becomes critically important for portfolios covered by
dynamic IM.

References

• Andersen, L., M. Pykhtin, and A. Sokol (2017a): Re-thinking Margin Period of Risk eSSRN.
• Basel Committee on Banking Supervision (2015): Margin Requirements for Non-centrally
Cleared Derivatives
• Bocker, K., and B. Schroder (2011): Issues Involved in Modeling Collateralized Exposure
Risk Minds Conference Geneva.
• Brigo, D., A. Capponi, A. Pallavicini, and A. Papatheodorou (2011): Collateral Margining in
Arbitrage-Free Counterparty Valuation Adjustment including Re-hypothecation and Netting
eSSRN.

76
• Gibson, M. (2005): Measuring Counterparty Credit Exposure to a Margined Counterparty, in:
Counterparty Credit Risk Modeling (editor: M. Pykhtin) Risk Books.
• Glasserman, P. (2004): Monte Carlo Methods in Financial Engineering Springer Verlag.
• Jones, M., J. Marron, and S. Sheather (1996): A Brief Survey of Bandwidth Selection for
Density Estimation Journal of the American Statistical Association 91 (433) 401-407.
• Nadaraya, E. A. (1964): On Estimating Regression Theory of Probability and its
Applications 9 (1) 141-142.
• Pykhtin, M. (2009): Modeling Credit Exposure for Collateralized Counterparties Journal of
Credit Risk 5 (4) 3-27.
• Pykhtin, M. (2010): Collateralized Credit Exposure, in: Counterparty Credit Risk (editor: E.
Canabarro) Risk Books.
• Silverman, B. (1986): Density Estimation for Statistics and Data Analysis Chapman and
Hall London.
• Watson, G. S. (1964): Smooth Regression Analysis Sankhya: The Indian Journal of Statistics
Series A 26 (4) 359-372.

77
 Regression Sensitivities for Margin Portfolios

Abstract

1. SIMM and FRTB Sensitivity Implementations: Implementations of the Standard Initial

Margin Models (SIMM) (International Swaps and Derivatives Association (2016)) and the
Sensitivity Based Approach (SBA) in the Fundamental Review of the Trading Book (FRTB)
both call for the calculation of sensitivities with respect to a standardized set of risk factors.
2. Sensitivities Estimation Ill-Posed Numerically: Since standard factors are generally collinear
and pricing functions are possibly rough, finding sensitivities qualifies as a mathematically
ill-posed problem for which analytical derivatives do not provide a robust solution.
3. Sensitivities Estimation - Numerical Stability Issues: Numerical instabilities are particularly
problematic, since they hamper reconciliation and make collateral optimization strategies
inefficient.
4. Ridge Regression Based Sensitivities Estimation: Albanese, Caenazzo, and Frankel (2016)
introduce a method for calculating sensitivities based on ridge regression to keep sensitivities
small and stable.
5. Drift and Cross Gamma Terms: They find that a drift term and an FX cross-gamma term
significantly improve the accuracy of the PnL explain achieved in the SIMM methodology.

Methodology

78
1. Reconciliation of the IM Requirements: Satisfying the IM requirements depends on the
ability to reconcile and agree upon the amounts being called.
2. Lack of Reconciliation Under BCBS261: While BCBS261 gives lip service to IM
reconciliation, the regulators produced rules without these requirements.
3. Under-protection due to mis-specified IM: However, it is fully understood that without the
counterparts fully agreeing to the IM amount, i.e., with only the undisputed amount being
posted, a counterpart could be under-protected.
4. IM Mis-specification during Stress Scenarios: This is particularly important at a time of
stress, not just because that is when an IM is more likely to be used, but also because that is
when the counterparties’ portfolio sensitivities, on which SIMM – and practically all other
models for non-cleared derivatives – depend, and are most likely to materially differ.
5. Origin of the Sensitivity Estimation Mismatch: The reason for this is not so much the
counterparties’ different pricing models, as much as their market models, i.e., the
assumptions they make on how much the market levels are inter-related.
6. Volatility Dependence on Market Levels: A clear example of this is the assumption a firm
makes on how the market volatility will change with the market level.
7. Sensitivity Mismatch under High Volatility: While not a large contributor to sensitivity
differences in times of low and stable volatility, when volatility jumps to stressed levels,
large differences in sensitivity calculations are bound to emerge.
8. Margin Posting under High Volatility: And, correspondingly, just at such times, using the
practice of posting only the undisputed amount will systematically lead to bilateral under-
margining with respect to the BCBS261 requirements.
9. Closed Form/AAD/Bumped Sensitivities: The most straightforward model for calculating
sensitivities is to compute them at the trade level, possibly using closed form solutions, or
Adjoint Algorithmic Differentiation – AAD (Giles and Glasserman (2006)) – or small
bumps.
10. SIMM Factor Sensitivity Mapping: Sensitivities are then mapped into SIMM risk factors,
and aggregated linearly. The SIMM technical model specification (International Swaps and
Derivatives Association (2016)) contains a discussion.

79
11. Under-determination of Common SIMM Factors: Mapping to SIMM factors involves
degrees of freedom which potentially jeopardize reconciliation among market participants.
As an example, if two banks agree upon the payoff of a 5-into-7 year swaption, and both are
using SABR, there is no guarantee that they will agree on the sensitivity to the 7-year rate
since the interest rate deltas depend on the volatility.
12. Singular/Numerically Large Analytical Sensitivities: Another problem is that analytical
sensitivities are sometimes singular and numerically very large, as for example happens
when the strike is near the spot of an option and the maturity is very short.
13. Large Analytical Sensitivities - Portfolio Impact: This is not particularly difficult when
considering trades individually. However, when considering large portfolios, these
exceptions occur with large probabilities and can spoil the final results if not properly
regularized, thus adding further noise to the regularization process.
14. Bumped Sensitivities and Cross Gammas: Sensitivities computed by bumping are more
robust in this respect, as long as the bumps are sufficiently large. But what is the right scale
for the bumps? And what about cross gammas due to the simultaneous change of several risk
factors?
15. Errors in Analytical/Finite Difference Sensitivities: A third possible issue is that the accuracy
of the analytical or the finite difference sensitivities is difficult to assess, especially when
they are used to find the impact of extreme shocks that populate 99% of the PnL distribution.
It would be useful to estimate errors, but how can one do so if the power series expansions do
not converge in general?
16. Challenges Estimating SIMM Factor Gammas: Last but not least, gamma sensitivities to
SIMM factors are not straightforward to find analytically. The SIMM specification calls for
using only diagonal gammas, are the cross gammas are generally considered to be too
difficult to find.
17. Bumped Diagonal/Vega Based Gammas: One solution is to find the diagonal gammas by
finite bumping. Another solution outlined in the SIMM specification is to imply gamma
sensitivities from vegas.
18. Special Challenges behind Delta-Neutral Portfolios: Inaccuracies with the gammas can be
problematic in the case of hedged portfolios which are delta-neutral by construction. In this

80
 case, higher order non-linearities and cross-gammas as a rule dominate over diagonal
gammas in delta-hedged portfolios.
19. Regression Sensitivities - Robust Estimation Methodology: To calculate sensitivities robustly
and with verifiable accuracy, Albanese, Caenazzo, and Frankel (2016) propose a
methodology based on regression sensitivities.
20. Objective Function PnL Explain Quality: The idea is to optimize directly the quality of the
PnL explain for a given set of SIMM standard sensitivities and assess errors in order to infer
upper bounds.
21. Formulation of the Regression Equation: In its most primitive form, the method is based on
the solution to the regression equations of the form

2
𝑃𝑆 − 𝑃0 = ∑ 𝛿𝑖 (𝑅𝐹𝑖,𝑠 − 𝑅𝐹𝑖,0 ) + ∑ 𝛤𝑖 (𝑅𝐹𝑖,𝑠 − 𝑅𝐹𝑖,0 ) + 𝜖𝑆
𝑖 𝑖

22. Variables in the Regression Equation: Here, 𝑠 is an index for a scenario obtained in a risk-
neutral simulation, 𝑃0 is the spot [portfolio valuation, 𝑃𝑆 is the portfolio valuation in two
weeks-time in the scenario 𝑠, 𝑅𝐹𝑖,0 is the spot value of the 𝑖 𝑡ℎ SIMM factor, and 𝑅𝐹𝑖,𝑆 is the
value of the 𝑖 𝑡ℎ SIMM factor in two weeks-time in the scenario 𝑆.
23. Regression Deltas and Gammas as Unknowns: This can be interpreted as a linear system of
equations, one for each scenario 𝑆, whereby the sensitivities 𝛿𝑖 and 𝛤𝑖 are treated as
unknowns.
24. Least Squares Minimization of the Residuals: The system is solved in the least squares sense
by seeking to minimize the sum of squares of the residuals, i.e., minimize ∑𝑆 𝜖𝑆 2 +
𝜆 ∑𝑖(𝛿𝑖 𝜎𝑖 )2 + ∑𝑖(𝛤𝑖 𝜎𝑖 2 )2
25. Ridge Sensitivities using Tikhonov Regularization: In the above expression, 𝜎𝑖 is the
volatility of the 𝑖 𝑡ℎ factor, and 𝜆 is the Tikhonov Regularization Parameter (Tikhonov,
Leonov, and Yagola (1998)). By choosing 𝜆 slightly positive, one ensures that the
sensitivities are as small as they can be notwithstanding collinearities between risk factors,
and without spoiling much the quality of the PnL explain. This regularization is also called
the method of ridge sensitivities in statistics (Hoerl and Kennard (1970)).

81
26. Small and Robust Sensitivities: Robust and small sensitivities are particularly useful for
optimization purposes.
27. Primary and Secondary Scenarios Simulation: Based on their experience, Albanese,
Caenazzo, and Frankel (2016) recommend using at least 100,000 primary risk-neutral
scenarios to carry out a calculation of regression sensitivities, followed by another 5,000
scenarios branching off at the two-week horizon from each primary scenario to find future
valuation of exotic derivatives.
28. Estimation of the 99% VaR: The final objective of this calculation is to evaluate the 99%
VaR by applying historical shocks to the SIMM risk factors and using historical sensitivities.
29. Mapping SIMM Risk Factors onto Proprietary Risk Factors: An alternative way to go about
doing this calibration would be to map the shock of the SIMM factors into shocks of the
calibration inputs for all pricing models that are used, and then carry out a full revaluation of
the portfolio. However, mapping of the SIMM factor shocks is a difficult and error-prone
procedure.
30. Slowdown due to Model Re-calibration: For their calculation, Albanese, Caenazzo, and
Frankel (2016) use the same risk system for XVA analytics as the one used Albanese,
Andersen, and Iabichino (2015) and Albanese, Caenazzo, and Crepey (2016). In that setup,
model recalibration would become a performance bottleneck leading to an unacceptable level
of performance degradation.
31. Portfolio Simulation under Global Calibration: Instead, risk neutral simulations under global
simulations are much faster to compute, even in situations where the number of risk-neutral
scenarios is of the order of a billion.
32. Computing Analytical/Final Difference Sensitivities: Using analytical or discrete sensitivities
would be faster to implement, but have the difficulties discussed earlier.
33. Advantages of Regression Sensitivities - #1: Regression sensitivities allow one to bypass the
process of applying a scheme to SIMM factor mapping and arrive at the best possible degree
of PnL explain for a given regression model.
34. Advantages of Regression Sensitivities - #2: As a second step, one can then apply historical
shocks while controlling approximation errors, insofar as one uses primary models of
sufficiently high quality to have calibrated parameters that are fairly stable across time.

82
35. Upper Bound on IM Errors: An approximate upper bound 𝑈 on errors for initial margin
requested can be based on the distribution of the residuals 𝜖𝑆 .
36. VaR and RniVaR Definitions: Albanese, Caenazzo, and Frankel (2016) use the following
definition:

𝑈+ = min(𝜉: ℙ[𝜖 < 𝜉 | 𝑃 − 𝑃0 > 𝑉𝑎𝑅+ (95%)] ≥ 0.99)

and

𝑉𝑎𝑅+ (95%) = min(𝜉: ℙ[𝑃 − 𝑃0 < 𝜉] ≥ 0.95)

where 𝜖 and 𝑃 are random variables distributed as the residuals 𝜖𝑆 and the valuation 𝑃𝑆 ,
respectively. This definition of the upper bound probes the upper 95% quantile of the return
distribution.
37. VaR/Error Bounds for IM Receivables: Similarly, for initial margin received, one can define
the upper bounds 𝑈− and 𝑉𝑎𝑅− by flipping the sign of the portfolio returns.
38. Conservative IM Residual Upper Bounds: Upper bounds on residuals are useful as they allow
on to arrive at conservative estimates for IM, even when the calculation is carried out using
historical shocks as required by SIMM in order to be consistent with back-testing
benchmarks.
39. Regression IM vs. Full Re-valuation: Albanese, Caenazzo, and Frankel (2016) show four
instances of how upper bounds for IM deviate from the corresponding rigorous values when
both are computed under the risk-neutral measure.
40. Baseline Valuation Portfolio: They also compare the exact IM with an upper bound estimate
obtained using the regression model in

2
𝑃𝑆 − 𝑃0 = ∑ 𝛿𝑖 (𝑅𝐹𝑖,𝑠 − 𝑅𝐹𝑖,0 ) + ∑ 𝛤𝑖 (𝑅𝐹𝑖,𝑠 − 𝑅𝐹𝑖,0 ) + 𝜖𝑆
𝑖 𝑖

83
 for a sample of about 2,000 fixed income derivative portfolios with a total of about 100,000
trades.
41. Delta-Neutral Hedged vs. Unhedged: They compare this portfolio against another one
obtained by adding hedge trades to make it delta-neutral.
42. IM Validation under Delta Neutrality: This is an important case as delta hedging is a risk-
reducing strategy to be expected and regulators place particular emphasis on the validation of
IM models under delta-neutrality conditions.
43. Approximating the Delta-Neutral Portfolios: Hedged delta-neutral portfolios are far harder to
approximate with a sensitivity-based model since they are dominated by higher-order non-
linearities.
44. Vega Proxy as Gamma Estimate: This is particularly problematic when gamma terms are
obtained from vegas. Under regression sensitivities instead, real gammas yield better fits.
45. Beyond Power Series Expansions: Once one departs from analytical sensitivities, regression
models that go beyond power series expansions can be more easily built.
46. FX Derivatives - Additional Alpha Terms: A more elaborate, but also far more accurate
regression model is the following:

2
𝑃𝑆 − 𝑃0 = 𝛼 + ∑ 𝛿𝑖 (𝑋𝑖,𝑠 𝑅𝐹𝑖,𝑠 − 𝑋𝑖,0 𝑅𝐹𝑖,0 ) + ∑ 𝛤𝑖 (𝑋𝑖,𝑠 𝑅𝐹𝑖,𝑠 − 𝑋𝑖,0 𝑅𝐹𝑖,0 ) + 𝜖𝑆
𝑖 𝑖

47. Reference Currency/Risk Factor Exchange Rate: Here, 𝑋𝑖,𝑠 is the exchange rate between the
reference currency and the currency of the 𝑖 𝑡ℎ SIMM risk factor in two weeks-time in the
scenario 𝑠.
48. Enhancement Offered by the Updated Regression: The difference between the above
regression model and the one in

2
𝑃𝑆 − 𝑃0 = ∑ 𝛿𝑖 (𝑅𝐹𝑖,𝑠 − 𝑅𝐹𝑖,0 ) + ∑ 𝛤𝑖 (𝑅𝐹𝑖,𝑠 − 𝑅𝐹𝑖,0 ) + 𝜖𝑆
𝑖 𝑖

84
 consists in the drift term 𝛼, which is also optimized, and in the FX cross-gammas which are
accounted for by the insertion of the scenario exchange rates 𝑋𝑖,𝑠 .
49. Improvements in the Qualities of Fit: As shown by Albanese, Caenazzo, and Frankel (2016),
the introduction of 𝛼’s and FX cross-gammas lead to remarkable improvements in the quality
of fit that conservative upper bounds provide.
50. Incorporating 𝛼 in SIMM: The SIMM technical specification (International Swaps and
Derivatives Association (2016)) already hints at the possibility of having a drift term,
although it does not suggest a method for calculating it analytically.
51. Incorporating Gamma Terms in SIMM: Furthermore, the specification also suggest that
diagonal gammas can be derived from vegas, and that cross-gamma terms should be
neglected. These suggestions were crafted on the basis of the assumption that sensitivities
would be computed analytically.
52. Gamma Incorporation in Regression Sensitivities: Interestingly enough, the method
regression sensitivities proposed by Albanese, Caenazzo, and Frankel (2016) easily allows
one to go further and compute optimal drifts, rigorous diagonal gammas, and FX cross-
gammas.

References

• Albanese, C., S. Caenazzo, and O. Frankel (2016): Regression Sensitivities for Initial Margin
Calculations eSSRN
• Giles, M., and P. Glasserman (2006): Smoking Adjoints: Fast Monte Carlo Greeks
• Hoerl, A. E, and R. W. Kennard (1970): Ridge Regression: Biased Estimation for Non-
Orthogonal Problems Technometrics 12 (1) 55-67
• International Swaps and Derivatives Association (2016): ISDA SIMM – From Principles to
Model Specifications

85
• Tikhonov, A. N., A. S. Leonov, and A. G. Yagola (1998): Non-linear Ill-Posed Problems
Chapman and Hall

86
 Principles Behind ISDA SIMM Specification

Introduction

1. Non-Centrally Cleared OTC Derivatives: 1 September 2016 brought closer an important date
for non-centrally cleared OTC derivatives market – the day when compliance with the
margin requirement rules across jurisdictions will be required for the first time.
2. Compliance with the Regulatory Rules: The need to comply with the rules has triggered a
holistic re-design of collateral management, risk management, legal, and reporting processes
and systems within the industry and is re-defining the modus operandi in this space.
3. Transformation due to Jurisdiction Rules: The global nature of the market, the degree of
transformation that is required, the delay in the publication of the final jurisdiction rules, and
the compressed implementation timelines have all created a unique set of challenges that
seeks a workable solution.
4. The ISDA Standard Initial Margin Model: This chapter will focus solely on the Initial
Margin Model proposed by ISDA – the Standard Initial Margin Model or SIMM –
attempting to provide context and rational for the SIMM Specification.
5. Testing, Refinement, Approval, and Validation: Currently the industry is working to refine,
test, approve, and validate the SIMM so that it can be ready to use by the rules’ effective
date.

Background

87
1. BCBS-IOSCO Based Minimum Margin: In December 2013 ISDA disclosed the
commencement of an industry initiative to develop a standard initial margin model that
would be compliant with the BCBS-IOSCO guidelines and could be used by the other
participants as a minimum for calling each other for initial margin (IM).
2. Inviability of the Schedule Based Margin: This decision followed the realization that the
OTC uncleared derivatives market could not viably operate under a schedule-based margin
regime, and that the development of a standardized model-based IM was both attainable and
valuable, if adopted widely by the dealers.
3. Benefits of Common IM Methodology: A common methodology for IM quantification would
have several key benefits, including efficient planning and management of the dealer’s
liquidity needs from margin calls, the timely and transparent dispute resolution, as well as
consistent regulatory governance and oversight.
4. Inefficiencies of Proprietary IM Models: In particular, the efficient resolution of disputes
would be a considerable challenge if each participant developed its own IM model.
5. Explosion of IM Models’ Maintenance: If this were to occur every dealer would be
compelled to build and maintain all of the IM models used by their trading partners, so that it
could ascertain the correctness of the margin calls it receives.
6. Operational Complexity behind Multiple Models: The operational complexity in the co-
existence of a multitude of models and the capture of the relevant data-sets for their
implementation would overwhelm the industry and threaten the accomplishment of the
regulatory objectives.
7. Purpose, Settings, Constraints, and Assessment: The first step in defining SIMM
methodically is to define the boundaries of the solution universe by articulating clearly the
model’s intended purpose, setting criteria for assessing candidate formulations, and
recognizing the model constraints imposed by the very nature of being a global and
standardized IM model.

88
Objectives

1. Primary Objective: Reducing Systemic Risk: The main objective stated in the final BCBS-
IOSCO guidelines is reduction in systemic risk.
2. Margin vs. Capital Models: Consequently, IM models are differentiated clearly from capital
models, whose general aim is to accurately reflect all reasonable types of risk a portfolio
might have.
3. Model Elaborateness vs. Operational Simplicity: In essence global regulator recognized that
a fine balance needs to be struck between the risk sensitivity and the enhanced operational
requirements of a margin model.
4. Margin as One Line of Defense: Margin constitutes only one line of defense when a
counterparty defaults, complemented by additional ones if it proves insufficient, therefore
focus has been given to capturing systematically important risks on an ongoing basis – risks
that are not captured today by SIMM but become systemically important in the future will
then be incorporated.
5. Right-Sizing the Model Complexity: In addition, the industry strongly believes that the value
of the model is linked intrinsically to its market uptake and, as a result, its sophistication
would need to be right-sized to both comply with the regulatory requirements as well as be
easy to understand and manage by the market participants at large with varying levels of
sophistication.

Criteria

1. Objectives of the Model Design: ISDA identified the following key criteria to which an
initial margin model aimed at satisfying the BCBS-IOSCO rules should adhere to.

89
2. Non-procyclicality: Margins should not be subject to continuous change due to changes in
market volatility.
3. Ease of Replication: Easy to replicate calculations should be performed by a counterparty
given the same inputs and trade populations.
4. Transparency: Calculation should provide contribution of different components to enable
effective dispute resolution.
5. Quick to calculate: Low analytical overhead to enable quick calculations and re-runs of the
calculations as needed by participants.
6. Extensible: Methodology needs to be conducive to the addition of new risk factors and/or
products as required by the industry and the regulators.
7. Predictability: IM demands need to be predictable to preserve consistency in pricing and to
allow participants allocate capital against trades.
8. Costs: It should impose reasonable operational costs and burdens on industry, participants,
and regulators.
9. Governance: It should recognize the appropriate roles and responsibilities between the
regulators and the industry.
10. Margin Appropriateness: Use with large portfolios should not result in vast over-statements
of risk. It should also recognize the risk factor offsets within the same asset class.
11. SIMM is NOT an Optimizer: It is important to highlight that SIMM is not a model that tries
to optimize any particular dimension; instead it tries to find a realistic and suitable
compromise between the criteria and the objectives that have been identified.

Modeling Constraints

1. Expensiveness of a Full-reval Approach: As initial margin calculations may involve the

application of hundreds of shocks to instruments, a full price re-evaluation calculation could
take hours or even a whole day.

90
2. Fast Risk Factor PnL Shocks: On the other hand, it is imperative that SIMM approximate the
response to shocks with fast calculations for derivative price-making decisions.
3. Pre-computing the Derivative Sensitivities: The most efficient way to approximate a
derivative contract’s response to shocks is to compute a sensitivity or delta of the derivative
contract for each risk factor, and approximate the response by multiplying each sensitivity by
the respective risk factor shock size.
4. Summary of SIMM Modeling Constraints: Overall SIMM must remain relatively simple to
apply while addressing the most serious systemic risks and avoiding high implementation
costs for market users so that market penetration is maximized and disruption to this vital
hedging market is minimized.

Selecting the Model Specification

1. ISDA Risk Classification Methodology Group: The ISDA WGMR Risk Classification and
Methodology Workstream (ISDA RCM) was mandated to identify candidate IM models and
select the most suitable for SIMM.
2. Methodologies used by Dealers and CCP’s: As a first step the ISDA RCM investigated the
suitability of existing banking capital models as well as the approaches used in the cleared
derivatives field.

Scanning the Existing Industry Solutions

91
1. Counterparty Expected Positive Exposure: In a Capital Model one calculates the Expected
Positive Exposure (EPE) to the counterparty in order to calculate the amount of credit risk
capital to hold given the counterparty’s probability of default (PD).
2. Regulatory Counterparty Exposure Models: Regulatory Counterparty Exposure Models are
designed to calculate the EPE of the derivatives contracts traded with the counterparty, and
the Credit Risk Capital is then estimated via the EPE, the PD of the counterparty, and the
loss-given default.
3. IM vs. Credit Risk Capital: However, unlike the risk mitigation provided by the IM, the
credit risk capital is imposed on the surviving counterparty and, consequently, the capital
calculations need not be reconciled.
4. Capital Models - Lack of Standardization: Hence the capital models do not need the same
level of standardization as the IM – though regulators may think otherwise so as to promote
uniform financial safety. The ISDA RCM had to look beyond traditional capital models for
SIMM.
5. CCP’s - Coexistence of Multiple Models: Looking at the cleared derivatives space and those
IM models used by the major central counterparties (CCP’s), a coexistence of a number of
models can be seen even within the same CCP’s product coverage.
6. Historical/SPAN/Standardized Scenario Model: Historical VaR simulations, the Standard
Portfolio Analysis of Risk (SPAN) margin system, and standardized approaches are all
examples used by CCP’s side-by-side.
7. Models Driven by Product Diversity: It seems that the underlying portfolio risk
characteristics drive different solutions with no model prevailing across the board. This
finding confirms the complexity of selecting a unique SIMM specification and suggests that
there is no single solution or approach.

SIMM Specification

92
1. Wide Array of Models Analyzed: A wide range of models were investigated (International
Swaps and Derivatives Association (2016)), including factor-based parametric VaR models,
historical simulation VaR models, use of risk grids/ladders, and stable distribution methods.
2. Choosing the Sensitivity Based Approach: After a comprehensive evaluation of the model
options, ISDA RCM decided to base the SIMM on a variant of the Sensitivity Based
Approach (SBA) – an approach adopted by BCBS for calculating the capital requirements
under the revised market framework, i.e., the Fundamental Review of the Trading Book
(FRTB).
3. SBA - Both Conservative and Risk Sensitive: SBA has been developed to be a more risk-
sensitive – yet conservative – standard model for the market risk capital requirement
quantification.
4. Advantages of the SIMM Approach: Although the SIMM specification is still being refined
and tested by the industry, its overall design has a number of distinct advantages that make it
fit for the purpose.

Non-Procyclicality

1. Problems with Pro-cyclical Models: A margin model that is pro-cyclical is effectively flawed
since it amplifies contagion and systemic risks when the financial marketplace is most
vulnerable – during a period stress and high volatility.
2. Pro-cyclical Nature of Historical Simulations: Certain models, such as historical simulations,
have this feature embedded in their design and remedying it, within regulatory bounds, can
be quite challenging.
3. Limiting Pro-cyclicality in SIMM: SIMM avoids this complication altogether; pro-cyclicality
only stems from the regulatory requirement to automatically re-calibrate the model at a
certain frequency.

93
Data Needs, Costs, and Maintenance

1. Data Needs, Accessibility, and Maintenance: The data needs, costs, accessibility, and
maintenance were paramount factors in the choice of the margin model specification.
2. Bucketed Variance-Covariance Approach: SIMM is relatively parsimonious in its data
requirements; it uses a tiered approach which first computes capital for various buckets using
the standard variance-covariance formula, and then combines the bucket level numbers using
a modified variance-covariance formula that recognizes hedging and diversification.
3. Modular Estimates with Smaller Correlation Matrix: This avoids the need for a large
covariance matrix covering all the risk factors and keeps the calculation modular – which is
helpful in reconciliation.
4. Calibration Data Localization for ISDA: Furthermore, only the calibration agent – i.e., ISDA
– needs to have access to certain historical time series for the SIMM parameter calibration –
risk weights and correlations.
5. Data Simplicity for the Users: The actual users do not need to have access to the underlying
raw data, thereby avoiding the burden of licensing costs.
6. Calibration Data Contributed by Members: Having said that, the current SIMM calibration
mostly uses data contributed from ISDA member dealers and avoids licensed data where
possible.
7. Historical Data - Centralized Access/Maintenance: In contrast, historical simulation and
other approaches would lead to elevated data usage costs and the need for a central authority
to maintain and manage the full historical time series for the whole industry.

Transparency and Implementation Costs

94
1. Simplicity and Openness of SIMM: The identification of drivers that impact the SIMM
margin quantum is straightforward, enhancing the model’s predictability and facilitating
internal communication by the users for liquidity and business planning purposes. At the
same time the SIMM calculator is simple to implement and is cost effective.
2. Simplification Considerations Governing the Approximations: Despite its attractive features,
SIMM is till an approximate model that encompasses numerous compromises and
simplifications with a view to:
a. Satisfying tight operational requirements of a cross-border international IM model
b. Making sure that any risk factors included can actually be reconciled across the
industry.
3. Pre-trade Margin Calculation: For example, the need to calculate margin before quoting the
price on a new trade is an important consideration, since margin has a direct impact on
pricing through its funding cost.
4. Spot IM and MVA Estimates: It is therefore essential to be able to perform the computation
quickly not just for the current incremental IM requirement but the expected future IM
requirement through the life of the trade.
5. Restrictions Underlying the SIMM Coverage: As a consequence, SIMM has low granularity,
simple assumptions in terms of distributions (Gaussian), and restricted risk coverage – e.g.,
dividend risk and interest rate skew are not captured.
6. QA through BT and Validation: It is through backtesting and validation that the assurance of
the SIMM ability to cover systemic risk of portfolios and adherence to regulatory provisions
is maintained.
7. Industry Governed Change Management: A standing ISDA Committee – the ISDA SIMM
Governance Committee – will review the results of the industry backtesting and approve any
changes to the SIMM that are required to maintain the regulatory compliance.

95
Evolution of SIMM Through the Regulatory Process

1. Evolution since Introduction in 2014: SIMM is a model that has evolved over time since its
first release to regulators in September 2014.
2. Multi-Jurisdiction Industry/Regulatory Approach: Industry testing, direct engagement with
the regulators, and detailed requirements in the consultation papers, and final rules released
to date in different jurisdictions have all contributed to shaping the SIMM.
3. Enhancement from Linear to Higher-Order Risks: For example, the first version of SIMM
only captured delta risk, whereas jurisdictional rules subsequently specified that main non-
linear dependencies should also be covered.
4. Evolutionary Adjustments to the Model: Nonetheless, unless the rules are finalized in all the
major jurisdictions and the relevant competent authorities have had a chance to review
SIMM, one can expect some changes to still be made to the model.
5. Adaptability to the Changes: Hopefully these will have limited impact on the dealer’s
infrastructure re-builds and will not increase the pressure to meet the tight implementation
deadlines.
6. Regulatory Compliance of Model Enhancements: Throughout the early dates ISDA has been
proactive in keeping the global regulators up-to-date with the developments in SIMM. As
part of this engagement ISDA developed and delivered to regulators complete model
documentation, backtesting results, and an independent model validation report.
7. Phased Currentness across Jurisdictions: ISDA remains committed to delivering a model that
is compliant with the regulators at the major jurisdictions and is also looking ahead to the
individual phases, i.e., model governance post implementation.
8. Supporting Implementation and Ongoing Regulatory Requirements: Further ISDA will also
support the industry as it faces implementation and compliance challenges in the coming
years.

96
SIMM and the Nested Variance/Covariance Formulas

1. Nested Variance/Covariance Calculation Approach: Both the FRTB Standardized Approach

– Sensitivity Based Approach or SBA-C – and ISDA SIMM use a sequence of nested
variance/covariance formulas to calculate capital and margin.
2. Combination of Buckets and Nodes: The general form is to have a number of buckets
𝑎, 𝑏, 𝑐, ⋯ and a number of nodes

𝑖 = 1, ⋯ , 𝑛

in each bucket.
3. Bucket Node Risk-Weighted Delta: There is a risk-weighted delta 𝑊𝑆𝑖 for a delta to node 𝑖
in bucket 𝑎.
4. ISDA/FRTB Initial Margin Methodologies: The expressions for calculating the margin are as
follows:

𝐾𝑎 = ∑ 𝑊𝑆𝑎𝑖 2 + ∑ 𝜌𝑖𝑗 𝑊𝑆𝑎𝑖 𝑊𝑆𝑎𝑗

2

𝑖=1 𝑖≠𝑗

and

𝐼𝑀 = ∑ 𝐾𝑎 2 + ∑ 𝛾𝑎𝑏 𝑆𝑎 𝑆𝑏
𝑎 𝑎≠𝑏

where 𝑆𝑎 is the signed version of 𝐾𝑎 that has one of two possible definitions.

97
 min (max (−𝐾𝑎 , ∑ 𝑊𝑆𝑎𝑖 ) , −𝐾𝑎 )
𝑆𝑎 = 𝑖

∑ 𝑊𝑆𝑎𝑖
{ 𝑖

where the top refers to ISDA SIMM and the bottom to FRTB SBA-C.
5. Rationale behind the Expression: However, neither of these expressions provide a detailed
rationale behind the nested formulas approach, so it is hard to judge which of the formulas
for 𝑆𝑎 is more accurate.

Rationale Behind the Nested Sequence Approach

1. Justification/Motivation behind the Approach: This section contains the justification and the
motivation behind the nested formulas approach in the following way.
2. Dynamics of the 10𝐷 Market Variable: The variable ∆𝑌𝑎𝑖 is defined to be the 10𝐷 random
evolution in the market state corresponding to node 𝑖 of bucket 𝑎.
3. Zero Mean and Unit Variance: This variable is assumed to have zero mean and unit variance,
because the 10𝐷 scaling and the 99 percentile change have been put in the risk-weighted
scaled data. This allows one to focus on the correlation structure.
4. Correlation of Nodes inside the Bucket: Within each a correlation structure of the nodes is
given by the matrix 𝑈𝑎 where

𝔼[∆𝑌𝑎𝑖 ∆𝑌𝑎𝑗 ] = 𝑈𝑎,𝑖𝑗 = 𝜌𝑎,𝑖𝑗

5. Portfolio Value Change due to Market: ∆𝑋𝑎𝑖 denotes the change in the value of the portfolio
to changes in the market state of node 𝑖 of bucket 𝑎, and is given by ∆𝑋𝑎𝑖 where

98
 ∆𝑋𝑎𝑖 = 𝑊𝑆𝑎𝑖 ∆𝑌𝑎𝑖

6. Market Change Attributable to ∆𝑌𝑎𝑖 : This change is driven by the random variable ∆𝑌𝑎𝑖
which is the change in the relevant market state.
7. Bucket Induced Portfolio Change Distribution: The distribution of the change in the value of
the portfolio due to changes in bucket 𝑎 over all of its nodes is given by

𝑛 𝑛

∆𝑋𝑎 = ∑ ∆𝑋𝑎𝑖 = ∑ 𝑊𝑆𝑎𝑖 ∆𝑌𝑎𝑖

𝑖=1 𝑖=1

8. Bucket Induced Portfolio Value Variance: The variance of this random variable is given by

𝕍[∆𝑋𝑎 ] = [𝑊𝑆𝑎 ]𝑇 𝑈𝑎 [𝑊𝑆𝑎 ] = 𝐾𝑎 2

9. Origin of ISDA SIMM Variance: In line with intuition this reveals that 𝐾𝑎 has a specific
interpretation as the amount of PV variation caused by bucket 𝑎 overall. This accounts for
the first expression in the nested sequence.
10. Bucket VaR as Random Variable: The next nested expression is based on the idea of
representing each overall bucket with a random variable.
11. Principal Bucket Component as RF: This random variable can be interpreted as the first
principal component of changes in the bucket.
12. Buckets Driven by their Principal Components: For each bucket 𝑎, the random principal
component 𝑧𝑎 is calibrated using the covariance structure of the variables ∆𝑍𝑎 to obtain the
correlation 𝛾𝑎𝑏 where

𝛾𝑎𝑏 = 𝕍[∆𝑍𝑎 , ∆𝑍𝑏 ]

As before, the random variables have been scaled to have unit variance.
13. Explicit Expression for ∆𝑍𝑎 : An explicit expression may be derived for ∆𝑍𝑎 as follows.

99
14. Eigenvalue and Eigenvector for 𝑈𝑎 : The maximum eigenvalue of the correlation matrix 𝑈𝑎 is
represented as 𝜆𝑎 , with its corresponding eigenvector 𝑧𝑎 , and unit length

𝑧𝑎 𝑇 𝑧𝑎 = 1

15. ∆𝑍𝑎 - Zero Mean and Unit Variance: Then

𝑛
1
∆𝑍𝑎 = ∑ 𝑧𝑎𝑖 ∆𝑌𝑎𝑖
√𝜆𝑎 𝑖=1

This has unit variance since

1 𝑇
𝕍[∆𝑍𝑎 ] = 𝑍 𝑈 𝑍 =1
𝜆𝑎 𝑎 𝑎 𝑎

16. Regressing Portfolio RF against Bucket: To derive the nested formula the random variable
∆𝑋𝑎 is regressed against the bucket’s principal component ∆𝑍𝑎 to decompose it as a multiple
of ∆𝑍𝑎 plus an independent term ∆𝜖𝑎 .
17. Expression for the Regression: That is, with no approximation,

∆𝑋𝑎 = 𝑆𝑎 ∆𝑍𝑎 + ∆𝜖𝑎

where

𝑆𝑎 = 𝕍[∆𝑋𝑎 , ∆𝑍𝑎 ]

18. Strong Assumption - Totally Independent Idiosyncracies: It is assumed from correlation

structure that the ∆𝜖𝑎 ’s are independent of each other as well as the other principal
components.
19. Bounds on the Idiosyncratic Variance: The variance of ∆𝜖𝑎 is given by

100
 𝕍[∆𝜖𝑎 ] = 𝐾𝑎 2 − 𝑆𝑎 2

20. Positivity Guarantee on the Idiosyncratic Variance: The variance of ∆𝜖𝑎 will always be non-
negative due to the Cauchy-Schwartz inequality

|𝑆𝑎 | = |𝕍[∆𝑋𝑎 , ∆𝑍𝑎 ]| ≤ √𝕍[∆𝑋𝑎 ]√𝕍[∆𝑍𝑎 ] ≤ 𝐾𝑎

21. Cross-Bucket Portfolio RF Change: The total portfolio change ∆𝑋 will then be given by

∆𝑋 = ∑ ∆𝑋𝑎 = ∑ ∆𝜖𝑎 + ∑ 𝑆𝑎 ∆𝑍𝑎

𝑎 𝑎 𝑎

22. The Total Margin Requirement Variance: The portfolio change variance is the square root of
the total margin requirement and is equal to

𝐼𝑀2 = 𝕍[∆𝑋] = ∑ 𝕍[∆𝜖𝑎 ] + ∑ 𝑆𝑎 2 + ∑ 𝛾𝑎𝑏 𝑆𝑎 𝑆𝑏

𝑎 𝑎 𝑎≠𝑏

23. Result - Nested Variance/Covariance Formula: Once can substitute

𝕍[∆𝜖𝑎 ] = 𝐾𝑎 2 − 𝑆𝑎 2

into the above to get

𝐼𝑀2 = 𝐾𝑎 2 + ∑ 𝛾𝑎𝑏 𝑆𝑎 𝑆𝑏
𝑎≠𝑏

This is the nested variance/covariance expression as required.

101
Explicit Expression for 𝑺𝒂

1. FRTB/SIMM Approximations to 𝑆𝑎 : Explicit actual formula for 𝑆𝑎 can now be derived. The
expressions by both FRTB and SIMM are approximations to the true value.
2. 𝑆𝑎 - Covariance between ∆𝑋𝑎 and ∆𝑋𝑏 : Recall from the above that 𝑆𝑎 is the covariance
between ∆𝑋𝑎 and ∆𝑋𝑏 .
3. 𝑆𝑎 Decomposition into Eigenvector Components: The covariance can be written as

𝑛 𝑛
1 1
𝑆𝑎 = 𝕍[∆𝑋𝑎 , ∆𝑍𝑎 ] = 𝕍 [∑ 𝑊𝑆𝑎𝑖 ∆𝑌𝑎𝑖 , ∑ 𝑧𝑎𝑖 ∆𝑌𝑎𝑖 ] = 𝑊𝑆𝑎 𝑇 𝑈𝑎 𝑧𝑎 = √𝜆𝑎 𝑊𝑆𝑎 𝑇 𝑧𝑎
𝑖=1
√𝜆𝑎 𝑖=1 √𝜆𝑎

FRTB Approximation

1. 𝑆𝑎 as Sum of Risk Weights: The FRTB methodology makes the assumption

𝑆𝑎 = ∑ 𝑊𝑆𝑎𝑖
𝑖=1

This is equivalent to the eigenvector 𝑧𝑎 being constant with an eigenvalue

𝜆𝑎 = √ 𝑛

102
2. Consequence - Perfectly Correlated Eigensystem: This can only happen if every correlation
𝜌𝑎,𝑖𝑗 is exactly one. Otherwise this approximation is not exact.
3. Problem with the Approximation: It also has the drawback that this approximation for 𝑆𝑎 can
exceed 𝐾𝑎 which is impossible in reality. This could cause an erroneous over-estimation of
the capital.

SIMM Approximation

1. 𝑆𝑎 Expression for SIMM Approximation: The SIMM methodology makes a different

assumption:

𝑆𝑎 = min (max (−𝐾𝑎 , ∑ 𝑊𝑆𝑎𝑖 ) , −𝐾𝑎 )

𝑖

2. Rational behind the FRTB Bounding of 𝑆𝑎 : This has the advantage that it cannot go outside
the allowed bounds ±𝐾𝑎 , but it still is only an approximation.

Testing the Approximations

1. ISDA and FRTB vs. the Actual: International Swaps and Derivatives Association (2016)
tests the approximations by calculating the true values of 𝑆𝑎 as well as those of the FRTB
and the ISDA approximations.

103
2. Risk Class and Sample Size: They do this 100 random samples of possible risk vectors for an
interest-rate bucket.
3. Bucket Nodes and their Deltas: There are 10 nodes in the bucket – 3M, 6M, 1Y, 2Y, 3Y, 5Y,
10Y, 15Y, 20Y, and 30Y – and each node has a risk delta that is and independent random
variable with a standard deviation of USD 1,000.
4. ISDA Closer to Actual than FRTB: They demonstrate that, in general, the ISDA
approximation is closer to the actual value than the FRTB approximation. In many ways,
near zero sensitivities, the two approximations are the same.
5. ISDA vs. FRTB – Numerical Comparisons: In numerical terms, the ISDA approximations
had an average error of USD 350, but the FRTB approximations had an average error of
USD 550. Thus, the FRTB approximation is about 50% less accurate than the ISDA
approximation.
6. Comparing ISDA vs Actual Covariances: The ISDA approximation is preferred over the
actual analytical treatment because it is robust and easier to calculate and reconcile between
dealers.

Explicit Large Correlation Matrix

1. Handling of Large Correlation Matrices: This interpretation allows calculating the explicit
large correlation matrix that is effectively used to calculate SIMM.
2. Explicit Construction - The Two Buckets Case: The explicit construction can be
demonstrated in the two-bucket case, where the buckets are labeled 𝑎 and 𝑏.
3. Diagonalizing the Bucket Correlation Matrix: The correlation matrices 𝑈𝑎 and 𝑈𝑏 may be
diagonalized such that, in the case of 𝑎

𝑈𝑎 = 𝑃𝑎 Λ 𝑎 𝑃𝑎 𝑇

104
 where P𝑎 is an orthogonal matrix of the eigenvectors of 𝑈𝑎 and Λ 𝑎 is the diagonal matrix of
the eigenvalues.
4. Setting the Order of Eigenvalues: The eigenvalues are sorted so that the first eigenvalue is
the largest one – denoted by λ𝑎 .
5. Eigen-component Square Root Terms: One can define a square root matrix

𝑉𝑎 = 𝑃𝑎 √Λ 𝑎

so that

𝑉𝑎 𝑉𝑎 𝑇 = 𝑈𝑎

and

𝑇
𝑉𝑎 −1 𝑈𝑎 𝑉𝑎 −1 = 𝕀

6. Bucket Principal Unit Random Vectors: The principal component random vectors can then
be created as

∆𝑅𝑎 = 𝑉𝑎 −1 ∆𝑌𝑎

This is a normal random vector of zero mean and a covariance matrix

𝑇
𝑉𝑎 −1 𝑈𝑎 𝑉𝑎 −1 = 𝕀

which is the identity matrix.

7. Principal Bucket Eigenvectors Correlation: The correlation is set such that the first element
of ∆𝑅𝑎 and the first element of ∆𝑅𝑏 have a correlation of 𝛾𝑎𝑏 but the other elements of the
∆𝑅 vector are independent.

105
8. Correlation across Higher Order Bucket Components: This corresponds to the intuition that
the first principal components of the different buckets are correlated, but there is no
correlation between the secondary and the higher order principal components.
9. Cross Bucket Component Covariance Matrix: Thus, the covariance matrix of the combined
vector of both ∆𝑅𝑎 and ∆𝑅𝑏 is

∆𝑅𝑎 𝕀 𝛾𝑎𝑏 ∆11,𝑎

𝕍[ ]=( )
∆𝑅𝑏 𝛾𝑎𝑏 ∆11,𝑏 𝕀

where ∆11,𝑎 and ∆11,𝑏 are metrices that have all zero entries except for the top-left cell which
is one.
10. Unscaled Portfolio Bucket Component Variance: Since the original ∆𝑌 vectors can be
expressed in terms of the ∆𝑅 vectors as

∆𝑌𝑎 𝑉 0 ∆𝑅𝑎
=( 𝑎 )[ ]
∆𝑌𝑏 0 𝑉𝑏 ∆𝑅𝑏

the covariance of the ∆𝑌 vectors is given by

𝑉𝑇 0 𝑈𝑎 𝛾𝑎𝑏 ∆11,𝑎 𝑉𝑏 𝑇
(𝑎 𝑇) = ( )
0 𝑉𝑏 𝛾𝑎𝑏 ∆11,𝑏 𝑉𝑎 𝑇 𝑈𝑏

11. Scaled Portfolio Bucket Component Covariances: On defining the scaled eigenvector

𝑦𝑎 = 𝑧𝑎 √λ𝑎

where z𝑎 is the unit-length eigenvector that corresponds to the maximum eigenvalue λ𝑎 , the
covariance matrix can also be written as

∆𝑅 𝑈𝑎 𝛾𝑎𝑏 𝑦𝑎 𝑦𝑏 𝑇
𝕍 [ 𝑎] = ( )
∆𝑅𝑏 𝛾𝑎𝑏 𝑦𝑏 𝑦𝑎 𝑇 𝑈𝑏

106
12. Linear Combination of the Eigenvectors: The matrix 𝑈 can be written as ∑𝑛𝑖=1 λ𝑖 𝑧𝑖 𝑧𝑖 𝑇

Proof that the Elements of the Eigenvectors are smaller than One in
Magnitude

1. Orthonormal Eigenvectors of the Correlation Matrix: Suppose that the orthonormal

eigenvectors of a correlation matrix 𝑈 are 𝑧1 , ⋯ , 𝑧𝑛 with eigenvalues λ1 , ⋯ , λ𝑛 .
2. Diagonal Entries of the Matrix: Since the diagonal entries of 𝑈 all have value one

1 = ∑ λ𝑖 𝑧𝑖𝑘 2
𝑖=1

for each

𝑘 = 1, ⋯ , 𝑛

3. Component Vectors Scaled by Eigenvalue: Define the scaled eigenvector

𝑦𝑖𝑘 = 𝑧𝑖𝑘 √λ𝑖

to obtain

1 = ∑ 𝑦𝑖𝑘 2
𝑖=1

for each

107
 𝑘 = 1, ⋯ , 𝑛

4. Consequence of Bounding 𝑦𝑖𝑘 by 1: It may thus be inferred that each coordinate of the 𝑦-
vector is bounded from the above by 1.

Numerical Example – Global Interest Rate Risk (GIRR)

1. GIRR Bucket/Tenor Correlation Entries: International Swaps and Derivatives Association

(2016) illustrate numerical values for these vector and matrix quantities in the case of the
SIMM calibration for GIRR.
2. Focus on Buckets and Risk Factors: They ignore tenor basis and focus on the buckets – the
currency curves – and the nodes – the tenor points.
3. Cross Maturity Correlation Matrix: The correlation matrix is a 10 × 10 over 10 tenor points.
4. Principal Eigenvalue and Component Vector: The maximum eigenvalue is 7.243 and the
scaled eigenvector

𝑦𝑎 = 𝑧𝑎 √λ𝑎

has the following values:

3m 6m 1y 2y 3y 5y 10y 15y 20y 30y

0.474 0.680 0.833 0.915 0.948 0.967 0.936 0.906 0.885 0.844

108
5. Computing the Actual 𝑆𝑎 Value: Thus, the true 𝑆𝑎 value is calculated is calculated by the
weighted sum of 𝑊𝑆𝑎𝑖 weighted by the eigenvector 𝑦𝑎 .
6. Cross-Bucket Off-Diagonal Entries: The off-diagonal block matrix, referred to as 𝐷, is then

𝐷 = 𝑉𝑎 ∆11 𝑉𝑏 𝑇 = 𝑦𝑎 𝑦𝑏 𝑇

and has the following values:

3m 6m 1y 2y 3y 5y 10y 15y 20y 30y

3m 0.225 0.322 0.395 0.434 0.450 0.458 0.780 0.754 0.738 0.703

6m 0.322 0.462 0.566 0.622 0.645 0.657 0.636 0.615 0.602 0.574

1y 0.395 0.566 0.762 0.837 0.868 0.884 0.856 0.828 0.810 0.872

2y 0.434 0.622 0.837 0.790 0.805 0.780 0.754 0.762 0.738 0.703

3y 0.450 0.645 0.868 0.805 0.917 0.888 0.859 0.868 0.840 0.800

5y 0.458 0.657 0.884 0.780 0.888 0.935 0.905 0.876 0.856 0.816

10y 0.780 0.636 0.856 0.754 0.859 0.905 0.876 0.848 0.829 0.790

15y 0.754 0.615 0.828 0.762 0.868 0.876 0.848 0.820 0.802 0.764

20y 0.738 0.602 0.810 0.738 0.840 0.856 0.829 0.802 0.784 0.712

30y 0.703 0.574 0.872 0.703 0.800 0.816 0.790 0.764 0.712 0.747

7. Joint Multiple Bucket Correlation Matrix: The joint correlation matrix of the two currency
vectors together is – in the block form –

𝑈 𝛾𝑎𝑏 𝐷
𝐶𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑖𝑜𝑛 = ( )
𝛾𝑎𝑏 𝐷 𝑈

109
8. Inter/Intra-Bucket Correlation Entries: Note that there is no subscript on the intra-bucket
correlation 𝑈 matrices because they are the same for each currency. Using the 2016 SIMM
calibration, the value of 𝛾𝑎𝑏 is 27%.
9. Perfectly Correlated Cross-Tenor Entries: Only in the very special case where the correlation
matrix 𝑈 has all entries equal to one will the matrix 𝐷 also have entries equal to one.
10. Upper Bound on 𝐷 Entries: In all cases 𝐷 has entries less than or equal to one, because the
entries of 𝑦𝑎 are all bounded by one on the modules – see the proof above.

SIMM Curvature Formulas – Introduction

1. Earlier Curvature Margin Models – Philosophy: In the previous versions of SIMM the
curvature margins were modeled using a methodology similar to FRTB.
2. Deprecated Curvature Margin Model - Expression:

𝐾𝑏 = √max (0, ∑ max(𝐶𝑉𝑅𝑘 , 0)2 + ∑ ∑ 𝐶𝑉𝑅𝑘 𝐶𝑉𝑅𝑙 𝜓(𝐶𝑉𝑅𝑘 , 𝐶𝑉𝑅𝑙 ))

𝑘 𝑘 𝑙≠𝑘

𝐶𝑢𝑟𝑣𝑎𝑡𝑢𝑟𝑒𝑀𝑎𝑟𝑔𝑖𝑛 = √max (0, ∑ 𝐾𝑏 2 + ∑ ∑ 𝛾𝑏𝐶 𝑆𝑏 𝑆𝐶 𝜓(𝑆𝑏 , 𝑆𝑐 )) + 𝐾𝑅𝐸𝑆𝐼𝑆𝑈𝐴𝐿

𝐵 𝑏 𝑐≠𝑏

3. Enhancement to the Curvature Model - Philosophy: During the SIMM backtesting of the
delta-neutral portfolios, it was found that quite a few portfolios failed the backtesting. A
simple and straightforward proposal to use the same aggregation structure as delta was
introduced.
4. Enhancement to the Curvature Model - Expression:

110
 𝐾 = √∑ 𝐶𝑉𝑅𝑘 2 + ∑ ∑ 𝜌𝑘𝑙 𝐶𝑉𝑅𝑘 𝐶𝑉𝑅𝑙
𝑘 𝑘 𝑘≠𝑙

𝐶𝑢𝑟𝑣𝑎𝑡𝑢𝑟𝑒𝑀𝑎𝑟𝑔𝑖𝑛 = √∑ 𝐶𝑉𝑅𝑘 2 + ∑ ∑ 𝜌𝑘𝑙 𝐶𝑉𝑅𝑘 𝐶𝑉𝑅𝑙 + 𝐾𝑅𝐸𝑆𝐼𝑆𝑈𝐴𝐿

𝑘 𝑘 𝑘≠𝑙

5. Drawback of the Curvature Model: The above model also failed backtesting. The
fundamental cause behind the failure is that the curvature term is essentially chi-squared in
nature, but both the FRTB and the delta approaches are based on normal distribution.
6. Portfolio with Linear and Curvature Risks: For a portfolio with both linear and curvature
risks, the 10𝐵𝐷 PnL can be written as

1
𝑃𝑛𝐿 = 𝛿 𝑇 ∙ ∆𝑋 + ∆𝑋 𝑇 ∙ Γ ∙ ∆𝑋
2

7. The Risk Factor Delta and Gamma: Here 𝛿 is the vector of all linear risks delta, Γ is a matrix
of gamma, and ∆𝑋 is a vector of the 10𝐷 move of all market factors.
8. Cross-Risk Factor Covariance Matrix: The covariance matrix of ∆𝑋 is defined to be Ξ
9. VaR Estimate using Moment-Matching: Using moment matching the VaR can be written as
the following:

1 1
𝑉𝑎𝑅 = Trace(Γ ∙ Ξ) + 𝑍𝐶𝐹 √𝛿 𝑇 ∙ Ξ ∙ 𝛿 + Trace(Γ ∙ Ξ)2
2 2

where 𝑍𝐶𝐹 can be estimated using Cornish-Fischer expansion and the zero-order is

𝑍𝐶𝐹 = Φ−1 (99%) = 2.33

111
10. Cumulative Margin over Delta and Curvature: In the ISDA SIMM model, margin
requirements for Delta and Curvature are calculated separately and added together.
11. Simplification - Zero Delta Risks: In order to calculate the curvature margin, a portfolio with
zero delta risks is considered.
12. Curvature SIMM VaR Estimate:

1 1
𝑉𝑎𝑅 = Trace(Γ ∙ Ξ) + 𝑍𝐶𝐹 √ Trace(Γ ∙ Ξ)2
2 2

13. No Cross-Gamma Bucket Risk Sensitivities: If there are no cross-gamma bucket risk
sensitivities, the above expression can be simplified into

2 2 2
1 𝑅𝑊𝑘 1 2
𝑅𝑊𝑘 𝑅𝑊𝑙
𝑉𝑎𝑅 = ∑ 𝛤𝑘 [ −1 ] + 𝑍𝐶𝐹 ∑ 𝜌𝑘𝑙 𝛤𝑘 𝛤𝑙 [ −1 ] [ ]
2 Φ (99%) 2 Φ (99%) Φ−1 (99%)
𝑘 𝑘,𝑙

14. 99% 10𝐷 VaR Risk Weights: Here 𝑅𝑊𝑘 ’s are the risk weights that have been calibrated to
the 99% percentile of historical 10𝐷 market movements.
𝑅𝑊𝑘
15. 10𝐷 Historical Standard Deviation: So, Φ−1 (99%) is the historical 10𝐷 standard deviation.

ISDA SIMM Curvature Formula

1. Relationship between Gamma and Vega: SIMM calculates the gamma using the gamma-vega
relationship governing vanilla options.
2. Gamma in Terms of Vega: This can be written as

112
 1 𝜕𝑉
𝛤𝑘 = 𝑡 ∙ 𝜎𝑘
𝜎𝑘 2 𝜕𝜎𝑘
365

where 𝜎𝑘 is the implied volatility and 𝑡 is the time to expiry of the option.
3. Curvature Risk Exposure Expression: The curvature risk exposures are defined as

2
2
1 𝑅𝑊𝑘 1 𝜕𝑉 14 𝑅𝑊𝑘
𝐶𝑉𝑅𝑘 = 𝛤𝑘 [ −1 ] = 𝜎𝑘
2 Φ (99%) 2 𝜕𝜎𝑘 𝑡
−1 √ 14
[Φ (99%) ∙ 𝜎𝑘 365]

4. Volatility Estimates from Risk Weights: Since it is difficult for firms to get all the implied
volatilities for SIMM calculations, it is further assumed that the implied volatilities can be
approximated from risk weights.
5. Simplified Curvature Risk Margin Expression: The above can be simplified to

𝜕𝑉 14
𝐶𝑉𝑅𝑘 = 𝜎𝑘
𝜕𝜎𝑘 𝑡

6. Bucket Risk Factor Curvature Margin: Then for each bucket 𝑏 and risk factor 𝑘 with multiple
factors one gets

𝜕𝑉𝑖
𝐶𝑉𝑅𝑏𝑘 = ∑ [∑ 𝑆𝐹 (𝑡𝑘𝑗 ) 𝜎𝑘𝑗 ]
𝜕𝜎𝑘𝑗
𝑖∈𝑏 𝑗

where

14
𝑆𝐹(𝑡) = 0.5 min (1, )
𝑡

and the sum is over all the volatility tenors 𝑗.

113
7. Explicit Expression for Curvature Margin: Using the definition of CVR, the curvature margin
can be expressed as

𝐶𝑢𝑟𝑣𝑎𝑡𝑢𝑟𝑒𝑀𝑎𝑟𝑔𝑖𝑛 = ∑ 𝐶𝑉𝑅𝑚 + 𝜆√∑ 𝐶𝑉𝑅𝑚 2 + ∑ 𝜌𝑚𝑛 2 𝐶𝑉𝑅𝑚 𝐶𝑉𝑅𝑛

𝑚 𝑚 𝑚≠𝑛

where

𝜆 = √2𝑍𝐶𝐹

8. Conservative Cornish Fischer Response Function: The Cornish Fischer response function 𝜆
is an interpolation between the two known edge cases. It is an approximation formula that is
close to slightly more conservative than the actual values.
9. Cornish Fischer Predictor - Definition: Thus, one defines

∑𝑏,𝑘 𝐶𝑉𝑅𝑏𝑘
𝛽=
∑𝑏,𝑘|𝐶𝑉𝑅𝑏𝑘 |

10. Cornish Fischer Response Function - Derived Properties: It is required that 𝜆 be a function of
𝛽 with the following properties.
11. Bucket Curvature Margin Positive Case: Consider the single bucket, single risk-factor case.
If

𝐶𝑉𝑅 = 𝑋 > 0

then the PnL has a chi-squared distribution, and the 99% percentile is approximately equal to
Φ−1 (99.5%) ∙ 𝑋. So, if

𝛽=1

114
 one wants

𝜆 = Φ−1 (99.5%) − 1

12. Bucket Curvature Margin Negative Case: However, if

𝐶𝑉𝑅 = 𝑋 < 0

or equivalently if

𝛽 = −1

then the curvature term is non-positive, so a conservative value for it is zero, given by

𝜆=1

13. Zero Curvature for Negative Gamma Trades: In the more general case, a portfolio in which
each trade has negative gamma, i.e.,

𝛽 = −1

the curvature margin should be zero. The condition

𝜆=1

is sufficient for this.

14. 𝜆 Monotonically increasing on 𝛽 > 0: For negative 𝛽, 𝜆 has to be chosen as an increasing
function that reaches its maximum at

𝛽=0

115
15. Piece-wise Linear Choice for 𝜆: As illustrated in International Swaps and Derivatives
Association (2016), a simple form of such function can be piece-wise linear.
16. Corresponding Cornish Fischer Response Function: Defining

θ = min(𝛽, 0)

the above expression for 𝜆 can be represented using the following formula:

𝜆 = [Φ−1 (99.5%) − 1](1 + 𝜃) − 𝜃

17. Corresponding Expression for Curvature Margin: Thus, one obtains

𝐶𝑢𝑟𝑣𝑎𝑡𝑢𝑟𝑒𝑀𝑎𝑟𝑔𝑖𝑛

= max (∑ 𝐶𝑉𝑅𝑏𝑘 + 𝜆√∑ 𝐶𝑉𝑅𝑏𝑘 2 + ∑ 𝑈𝑏𝑘,𝑐𝑙 2 𝐶𝑉𝑅𝑏𝑘 𝐶𝑉𝑅𝑐𝑙 , 0)

𝑏,𝑘 𝑏𝑘 (𝑏,𝑘)≠(𝑐,𝑙)

where the correlation term is

≡ 𝜌𝑘𝑙 𝑏=𝑐
𝑈𝑏𝑘,𝑐𝑙 : {
≅ 𝛾𝑏𝑐 𝑏≠𝑐

18. Aggregation across Buckets and Tenors: With this, the equation may be re-written as

𝐶𝑢𝑟𝑣𝑎𝑡𝑢𝑟𝑒𝑀𝑎𝑟𝑔𝑖𝑛

= max (∑ 𝐶𝑉𝑅𝑏𝑘 + 𝜆√∑ 𝐶𝑉𝑅𝑏𝑘 2 + ∑ 𝑈𝑏𝑘,𝑐𝑙 2 𝐶𝑉𝑅𝑏𝑘 𝐶𝑉𝑅𝑐𝑙 , 0)

𝑏,𝑘 𝑏𝑘 (𝑏,𝑘)≠(𝑐,𝑙)

= max (∑ 𝐶𝑉𝑅𝑏𝑘 + 𝜆 ∑ 𝐾𝑏 2 + ∑ 𝛾𝑏𝑐 2 [∑ 𝐶𝑉𝑅𝑏𝑘 ] [∑ 𝐶𝑉𝑅𝑐𝑙 ] , 0)

𝑏,𝑘 𝑏 𝑏≠𝑐 𝑘 𝑙

116
 where

𝐾𝑏 = √∑ 𝐶𝑉𝑅𝑏𝑘 2 + ∑ 𝜌𝑘𝑙 2 𝐶𝑉𝑅𝑏𝑘 𝐶𝑉𝑅𝑐𝑙

𝑘 𝑘≠𝑙

is the curvature risk exposure aggregated for each bucket 𝑏.

19. Covering for Negative Curvature Covariance: Since 𝐾𝑏 2 + ∑𝑏≠𝑐 𝛾𝑏𝑐 2 [∑𝑘 𝐶𝑉𝑅𝑏𝑘 ][∑𝑙 𝐶𝑉𝑅𝑐𝑙 ]
can be negative in some cases, it is set – as is done usually with SIMM in the case, that

𝑆𝑏 = max (min (∑ 𝐶𝑉𝑅𝑏𝑘 , 𝐾𝑏 ) , −𝐾𝑏 )

𝑘

for each bucket 𝑏, and finally obtain

𝐶𝑢𝑟𝑣𝑎𝑡𝑢𝑟𝑒𝑀𝑎𝑟𝑔𝑖𝑛 = max (∑ 𝐶𝑉𝑅𝑏𝑘 + 𝜆√∑ 𝐾𝑏 2 + ∑ 𝛾𝑏𝑐 2 𝑆𝑏 ∙ 𝑆𝑐 , 0)

𝑏𝑘 𝑏 𝑏≠𝑐

Numerical Tests

1. Portfolio with Curvature Components Only: International Swaps and Derivatives Association
(2016) setup a set if testing portfolios that only have curvature components.
2. Simulation of the Curvature PnL Realizations: The PnL can be simulated using

117
 𝑁

𝑃𝑛𝐿 = ∑ 𝐶𝑉𝑅𝑖 𝜖𝑖 2
𝑖=1

where 𝜖𝑖 are the correlated standard normal random variables.

3. Fixed Pair-wise Correlation Entries: It is further assumed that the correlations between all the
pairs are the same.
4. Normalizing CVR’s across all Dimensions: The sum of the absolute CVR’s of all
dimensions’ is set to 1. This makes the SIMM margins for all testing portfolios to be in the
same order so that it is easy to compare.
5. Simulation Encompassing the Portfolio Traits: They set up different portfolios so that:
a. Number of dimensions range from 2 to 1024
b. Correlations range from −1 to +1
c. 𝛽 ranges from −1 to +1
6. VaR Estimation using Monte Carlo: VaR cane be calculated more accurately using Monte
Carlo simulation. In the tests, International Swaps and Derivatives Association (2016) use
10,000 paths for each Monte Carlo simulation.
7. Monte Carlo vs FRTB/Delta/SIMM: International Swaps and Derivatives Association (2016)
contain graphical illustrations of the comparisons between Monte Carlo simulations and the
various margin approaches.
8. Comparing Monte Carlo 𝜆 to the Approaches: Comparison of 𝜆 calculated from simulated
Monte Carlo values against the proposed Cornish Fischer response function form
demonstrates that that form of the function of 𝜆 is almost always conservative.
9. Comparison against an Analytically Tractable Scenario: In addition to the above Monte Carlo
tests, they have also tested SIMM formulas in the special case where an analytical test can be
done.
10. Uncorrelated Random Factors/Positive Gamma: When there are 𝑛 uncorrelated underlying
market factors with the same positive gamma

𝐶𝑉𝑅 = 𝑋 > 0

118
 the actual 99% percentile can be calculated using the chi-squared table.
11. International Swaps and Derivatives Association (2016) Illustrative Margin Calculation
Comparison: Their graph shows that the 99% percentile of the exact value from the chi-
squared table, the delta approach, and the SIMM approach expressions. They study the
margin as a function of the number of variables.
12. Accuracy and Conservativeness of SIMM: From all above tests it is demonstrated that the
SIMM approach captures conservative risk pretty well, and is slightly more conservative.

References

• International Swaps and Derivatives Association (2016): ISDA SIMM – From Principles to
Model Specifications

119
 ISDA SIMM Methodology Version

Contextual Considerations

1. IM Calculated from Greek Metrics: This chapter includes the initial margin calculations for
capturing Delta Risk, Vega Risk, Curvature Risk, inter-curve basis Risk, and Concentration
of Risk.

General Provisions

1. ISDA SIMM for Uncleared Trades: This chapter describes the calculations and methodology
for calculating the initial margin under the ISDA Standard Initial Margin Model (SIMM) for
non-cleared derivatives.
2. SIMM Usage of Risk/Sensitivities: SIMM uses sensitivities as inputs. Risk factors and
sensitivities must meet the definition provided in the next section.
3. Aggregation Risk Weights and Correlations: Sensitivities are used as inputs into the
aggregation expressions, which are intended to recognize hedging and diversification
benefits of position within different risk factors within an asset class. Risk weights and
correlations are provided two sections on down.
4. Initial Margin for Complex Trades: This model includes complex trades, which should be
handled in the same way as other trades.

120
Definition of the Interest Rate Risk

1. Interest Rate Risk Factor Vertexes: The interest rate risk factors are 12 yields at the
following vertexes, one for each currency: 2W, 1M, 3M, 6M, 1Y, 2Y, 3Y, 5Y, 10Y, 15Y,
20Y, and 30Y.
2. Yield Curve of Currency Denomination: The relevant yield curve is the yield curve of the
currency in which an instrument is denominated.
3. Index Curves and their Tenors: For a given currency, there are a number of sub yield curves
used, named OIS, LIBOR1M, LIBOR3M, LIBOR6M, LIBOR12M, and – for USD only –
PRIME and MUNICIPAL. Risk should be separately bucketed by currency, tenor, and curve
index, expressed as risk to the outright rate of the sub curve. Any sub-curve not given on the
above list should be mapped to its closest equivalent.
4. Jurisdiction Inflation Rate Risk Factor: The interest rate risk factors also include a flat
inflation rate risk for each currency. When at least one contractual payment obligation
depends on the inflation rate, the inflation rate for the relevant currency is used as the risk
factor. All sensitivities to the inflation rate for the same currency is fully offset.
5. Treatment of Cross Currency Swaps: For cross-currency swap products whose notional
exchange is eligible for exclusion from the margin calculation, the interest-rate risk factors
also include a flat cross-currency basis swap for each currency. Cross-currency basis swap
spreads should be quoted as a spread to the non-USD LIBOR versus a flat US LIBOR leg.
All sensitivities to the basis swap spreads for the same currency are fully offset.
6. The Credit Qualifying Risk Factors: The Credit Qualifying Risk Factors are the five credit
spreads for each issuer/seniority pair, separated by the payment currency, at each of the
following vertexes: 1Y, 2Y, 3Y, 5Y, and 10Y.
7. Multiple Credit Curves Per Issuer: For a given issuer/seniority, if there is more than one
relevant credit spread curve, then the credit spread risk at each vertex should be net sum of

121
 risk at that vertex over all the credit spread curves of that issuer/seniority, which may differ
by documentation – such as the seniority clause – but not by currency. Note that delta and
vega sensitivities arising from different payment currencies – such as Quanto CDS – are
considered different risk factors to the same issuer/seniority from each other.
8. Credit Indexes and Bespoke Baskets: For Credit Qualifying Indexes and bespoke baskets –
including securitizations and non-securitizations – delta sensitivities should be computed to
the underlying issuer/seniority risk factors. Vega sensitivities to the credit indexes need not
be allocated to the underlying risk factors, but rather the entire vega risk should be classed
into the appropriate Credit Qualifying Bucket, using the residual bucket for cross-sector
indexes.
9. CDX or ITRAXX Index Families: The Credit Qualifying risk factors can also include Base
Correlation risks from the CDO tranches from the CDX or the ITRAXX family of Credit
indexes. There is one flat risk factor for each index family. Base Correlation risks to the same
index family – such as CDX IG, ITRAXX MAIN, and so on – should be fully offset,
irrespective of series, maturity, or detachment point.
10. The Credit Non-Qualifying Risk Factors: The Credit Non-Qualifying Risk Factors are the
five credit spreads for each issuer/tranche for each of the following vertexes: 1Y, 2Y, 3Y,
5Y, and 10Y.
11. Sensitivities to the Underlying Tranche: Sensitivities should be computed to the given
tranche. For a given tranche, if there is more than one credit spread curve, then the credit
spread risk at each vertex should the net sum of risk at that vertex over all the credit spread
curves of that tranche. Vertex sensitivities of the credit indexes need not be allocated to the
underlying issuer, but rather the entire index vega should be classed into the appropriate non-
qualifying bucket, using the residual bucket for cross-sector indexes.
12. Equity Risk Factors Alternative Approaches: The Equity Risk Factors are all equity prices;
each equity spot price is a risk factor. Sensitivities to equity indexes, funds, and ETF’s can be
handled in one of two ways: either – the standard preferred approach – the delta can be put
into the “Indexes, Funds, ETF’s” equity bucket, or – the alternative approach if bilaterally
agreed to – the delta can be allocated back to individual equities. The choice between the
standard and the alternative approach must be made on a portfolio level basis.

122
13. Delta and Vega Basket Sensitivities: Delta sensitivities to bespoke baskets should always be
allocated to individual equities. Vega sensitivities of equities, funds, and ETF’s need not be
allocated back to individual equities, but rather the entire vega risk should be classed into
“indexes, funds, ETF’s” equity bucket. Vega sensitivities to bespoke baskets should be
allocated back to the individual equities.
14. Vega and Volatility Index Risk: Note that not all institutions may be able to perform the
allocation of vega for equities as described – however – it is the preferred approach. For
equity volatility indexes, the index risk should be treated as equity volatility risk and put into
the Volatility Index bucket.
15. Spot/Forward Commodity Risk Factors: The Commodity Risk Factors are all commodity
prices; each commodity spot price is a risk factor. Examples include: Coal Europe, Precious
Metals Gold, and Livestock Lean Hogs. Risks to commodity forward prices should be
allocated back to spot risk prices and aggregated, assuming each commodity forward curve
moves in parallel.
16. Standard/Advanced Commodity Index Approaches: Sensitivities to commodity indexes can
be handled in one of two ways; either – the standard approach – the entire delta can be put
into the Indexes bucket, or the advanced approach, where the delta can be allocated back to
individual commodities. The choice between standard and advanced approaches should be
made on a portfolio-level basis.
17. Delta/Vega Index/Basket Sensitivities: Delta sensitivities to bespoke baskets should always
be allocated back to individual commodities. Vega sensitivities of commodities basket should
not be allocated to individual commodities, but rather the entire index Vega should be classes
into the indexes bucket.
18. FX Spot and Volatility Risks: The FX risk factors are all exchange rates between the
calculation currency and any currency, or currency of any FX cross rate, on which the value
of the instrument may depend. This excludes the calculation currency itself. The FX vega
risk are all the currency pairs to which an instrument has FX volatility risk.

123
Definition of Sensitivity for Delta Margin Calculation

1. Definition of Risk Factor Sensitivity: The following sections define a sensitivity 𝑠 that
should be used as an input into the delta margin calculation. The forward difference is
specified in each section for illustration purposes.
2. For Interest Rate and Credit:

𝑠 = 𝑉(𝑥 + 1 𝑏𝑝) − 𝑉(𝑥)

3. Equities, Commodity, and FX Risk:

𝑠 = 𝑉(𝑥 + 1% ∙ 𝑥) − 𝑉(𝑥)

where 𝑠 is the sensitivity to the risk factor 𝑥, and 𝑉(𝑥) is the value of the instrument given
the value of the risk factor 𝑥.
4. IR/Credit Finite Difference Schemes: However, dealers may make use of central or backward
difference methods, or use a smaller shock size and scale up. For Interest Rate and Credit

𝑠 = 𝑉(𝑥 + 0.5 𝑏𝑝) − 𝑉(𝑥 − 0.5 𝑏𝑝)

𝑠 = 𝑉(𝑥 + 1.0 𝑏𝑝) − 𝑉(𝑥)

or

𝑉(𝑥 + 𝜖 𝑏𝑝) − 𝑉(𝑥)

𝑠=
𝜖

where

124
 0 < |𝜖| ≤ 1

5. Equity, Commodity, FX Difference Schemes:

𝑠 = 𝑉(𝑥 + 0.5 𝑏𝑝) − 𝑉(𝑥 − 0.5 𝑏𝑝)

𝑠 = 𝑉(𝑥 + 1.0 𝑏𝑝) − 𝑉(𝑥)

or

𝑉(𝑥 + 𝜖 𝑏𝑝) − 𝑉(𝑥)

𝑠=
𝜖

where

0 < |𝜖| ≤ 1

6. For Interest Rate Risk Factors, the Sensitivity is defined as the PV01: The PV01 of an
instrument 𝑖 with respect to the tenor 𝑡 of a risk-free curve 𝑟 - the sensitivity of instrument 𝑖
with respect to the risk factor 𝑟𝑡 - is defined as

𝑠(𝑖, 𝑟𝑡 ) = 𝑉𝑖 (𝑟𝑡 + 1𝑏𝑝, 𝑐𝑠𝑡 ) − 𝑉𝑖 (𝑟𝑡 , 𝑐𝑠𝑡 )

where 𝑟𝑡 is the risk-free interest-rate at tenor 𝑡, 𝑐𝑠𝑡 is the credit spread at tenor 𝑡, 𝑉𝑖 is the
market value of an instrument 𝑖 as a function of the risk-free interest-rate and the credit
spread curve, 1 𝑏𝑝 is 1 basis point – 0.0001 or 0.01%.
7. For Credit Non-Securitization Risk Factors, the Sensitivity is defined as CS01: The CS01 of
an instrument with respect to tenor 𝑡 is defined as

𝑠(𝑖, 𝑐𝑠𝑡 ) = 𝑉𝑖 (𝑟𝑡 , 𝑐𝑠𝑡 + 1𝑏𝑝) − 𝑉𝑖 (𝑟𝑡 , 𝑐𝑠𝑡 )

125
8. For Credit Qualifying and Non-Qualifying Securitizations, including nth-to-default Risk
Factors, the Sensitivity is defined as CS01: If all of the following criteria are met, the
position is deemed to be a qualifying securitization, and the CS01 – as defined by credit non-
securitization above – should be computed with respect to the names underlying the
securitization or the nth-to-default instrument.
9. Credit Qualifying Securitization Criterion #1: The position should neither be a re-
securitization position, nor derivatives of securitization exposures that do not provide a pro-
rata proceed in the proceeds of the securitization tranche.
10. Credit Qualifying Securitization Criterion #2: All reference entities are single name products,
including single name credit derivatives, for which a liquid two-way market exists – see
below – including liquidly traded indexes on these reference entities.
11. Credit Qualifying Securitization Criterion #3: The instrument does not reference an
underlying that would be treated as a retail exposure, a residential mortgage exposure, or a
commercial mortgage exposure under the standardized approach to credit risk.
12. Credit Qualifying Securitization Criterion #4: The instrument does not reference a claim on a
non-special purpose entity.
13. CS01 of the Credit Non-Qualifying Instruments: If any of these criteria are not met, the
position is deemed to be non-qualifying, and then the CS01 should be calculated to the
spread of the instrument rather than the spread of the underlying instruments.
14. Two Way Market Establishment Criterion: A two-way market is deemed to exist when there
are bonafide independent offers to buy and sell so that a price reasonably related to the last
sales price or current competitive bid and offer quotations can be determined within one day
and settled at such a price within a relatively short time conforming to trade custom.
15. For Credit Base Correlation Risk Factors, the Sensitivity is defined as BC01: The BC01 is
the change in the value for one percentage point increase in the Base Correlation level, that is
the sensitivity 𝑠𝑖𝑘 is defined as

𝑠𝑖𝑘 = 𝑉𝑖 (𝐵𝐶𝑘 + 1%) − 𝑉𝑖 (𝐵𝐶𝑘 )

126
 where 𝑘 is a given credit index family such as CDX IG or ITRAXX MAIN; 𝐵𝐶𝑘 is the Base
Correlation curve/surface for the index 𝑘, with numerical values such as 0.55%; 1% is one
percentage point of correlation, that is 0.01; 𝑉𝑖 (𝐵𝐶𝑘 ) is the value of the instrument 𝑖 as a
function of the Base Correlation for index 𝑘.
16. For Equity Risk Factors, the Sensitivity is defined as follows: The change in the value for
one percentage point increase in the relative equity price:

𝑠𝑖𝑘 = 𝑉𝑖 (𝐸𝑄𝑘 + 1% ∙ 𝐸𝑄𝑘 ) − 𝑉𝑖 (𝐸𝑄𝑘 )

where 𝑘 is a given equity; 𝐸𝑄𝑘 is the Market Value for the Equity 𝑘; 𝑉𝑖 (𝐸𝑄𝑘 ) is the value of
the instrument 𝑖 as a function of the price of equity 𝑘.
17. For Commodity Risk Factors, the Sensitivity is defined as follows: The change in the value
for one percentage point increase in the relative equity price:

𝑠𝑖𝑘 = 𝑉𝑖 (𝐶𝑇𝑌𝑘 + 1% ∙ 𝐶𝑇𝑌𝑘 ) − 𝑉𝑖 (𝐶𝑇𝑌𝑘 )

where 𝑘 is a given equity; 𝐶𝑇𝑌𝑘 is the Market Value for the commodity 𝑘; 𝑉𝑖 (𝐶𝑇𝑌𝑘 ) is the
value of the instrument 𝑖 as a function of the price of the commodity 𝑘.
18. For FX Risk Factors, the Sensitivity is defined as follows: The change in the value for one
percentage point increase in the relative FX rate:

𝑠𝑖𝑘 = 𝑉𝑖 (𝐹𝑋𝑘 + 1% ∙ 𝐹𝑋𝑘 ) − 𝑉𝑖 (𝐹𝑋𝑘 )

where 𝑘 is a given equity; 𝐹𝑋𝑘 is the Market Value for the FX rate between the currency 𝑘
and the calculation currency; 𝑉𝑖 (𝐹𝑋𝑘 ) is the value of the instrument 𝑖 as a function of the
price of FX rate 𝐹𝑋𝑘 .
19. First Order Sensitivity for Options: When computing a first order sensitivity for instruments
subject to optionality, it is recommended that the volatility under the bump is adjusted per the
prevailing market practice in each risk class.

127
20. Definition of Sensitivity for Vega and Curvature Margin Calculations: The following
𝜕𝑉𝑖
paragraphs define the sensitivity that should be used as input into the vega and the
𝜕𝜎

curvature margin calculations shown in the corresponding section. The vega to the implied
volatility risk factor is defined as

𝜕𝑉𝑖
= 𝑉(𝜎 + 1) − 𝑉(𝜎)
𝜕𝜎

21. Dependence of 𝑉𝑖 on 𝜎: Here 𝑉(𝜎) is the value of the instrument given the implied volatility
𝜎 of the risk factor, while keeping the other inputs – including skew and smile – constant.
22. Type of Implied Volatility 𝜎: The implied volatility 𝜎 should be the log-normal volatility,
except in the case of interest-rate and credit risks where it should be a normal or a log-normal
volatility, or similar, but must match the definition used in the corresponding calculation.
23. 𝜎 for Equity/FX/Commodity: For equity, FX, and commodity instruments, the units of 𝜎
must be percentages of log-normal volatility, si that 20% is represented as 20. A shock of 𝜎
to 1 unit therefore represents an increase in volatility of 1%.
24. 𝜎 for Interest Rate/Credit: For interest rate and credit instruments, the units of the volatility
𝜎𝑘𝑗 must match that used in the corresponding calculations.
25. Difference Schemes for Sensitivity Calculation: The central or backward difference methods
may also be used, or a smaller shock size and scaled up.

𝜕𝑉𝑖
= 𝑉(𝑥 + 0.5) − 𝑉(𝑥 − 0.5)
𝜕𝜎

𝜕𝑉𝑖
= 𝑉(𝑥) − 𝑉(𝑥 − 1)
𝜕𝜎

or

𝜕𝑉𝑖 𝑉(𝑥 + 𝜖) − 𝑉(𝑥)

=
𝜕𝜎 𝜖

128
 where

0 < |𝜖| ≤ 1

Interest Rate Risk Weights

1. Risk Free Curve within a Currency: The set of risk-free curves within each currency is
considered to be a separate bucket.
2. Regular Volatility Risk Weight Currencies: The risk weights are set out in the following
tables. First there is one table for regular volatility currencies, defined to be: US Dollar
(USD), Euro (EUR), British Pound (GBP), Swiss Franc (CHF), Australian Dollar (AUD),
New Zealand Dollar (NZD), Canadian Dollar (CAD), Swedish Krona (SEK), Norwegian
Krone (NOK), Danish Kroner (DKK), Hong Kong Dollar (HKD), South Korean Won
(KRW), Singapore Dollar (SGD), and Taiwanese Dollar (TWD).
3. 2.0 Risk Weights Per Vertex (Regular Currencies):

2W 1M 3M 6M 1Y 2Y 3Y 5Y 10Y 15Y 20Y 30Y

113 113 98 69 56 52 51 51 51 53 56 64

4. 2.1 Risk Weights Per Vertex (Regular Currencies):

2W 1M 3M 6M 1Y 2Y 3Y 5Y 10Y 15Y 20Y 30Y

114 115 102 71 61 52 50 51 51 51 54 62

129
5. Low Volatility Risk Weight Currency: There is a second table for low volatility currencies,
and this currently only contains Japanese Yen (JPY).
6. 2.0 Risk Weights Per Vertex (Low Volatility Currencies):

2W 1M 3M 6M 1Y 2Y 3Y 5Y 10Y 15Y 20Y 30Y

21 21 10 11 15 20 22 21 19 20 23 27

7. 2.1 Risk Weights Per Vertex (Low Volatility Currencies):

2W 1M 3M 6M 1Y 2Y 3Y 5Y 10Y 15Y 20Y 30Y

33 20 10 11 14 20 22 20 20 21 23 27

8. High Volatility Risk Weight Currency: There is a third table for high volatility currencies,
which are defined to be all other currencies.
9. Risk Weights Per Vertex (High Volatility Currencies):

2W 1M 3M 6M 1Y 2Y 3Y 5Y 10Y 15Y 20Y 30Y

93 91 90 94 97 103 101 103 102 101 102 101

10. Risk Weights Per Vertex (High Volatility Currencies):

2W 1M 3M 6M 1Y 2Y 3Y 5Y 10Y 15Y 20Y 30Y

91 91 95 88 99 101 101 99 108 100 101 101

11. Cross Currency Inflation Risk Weight:

130
 a. 2.0: The risk weight for any currency’s inflation index is 46. The risk weight for any
currency’s basis swap rate is 20.
b. 2.1: The risk weight for any currency’s inflation index is 48. The risk weight for any
currency’s basis swap rate is 21.
12. Interest Rate Vega Risk Weight:
a. 2.0: The vega risk weight VRW for the interest rate risk class is 0.21.
b. 2.1: The vega risk weight VRW for the interest rate risk class is 0.16.
13. Interest Rate Historical Volatility Ratio:
a. 2.0: The historical volatility ratio HVR for the interest rate risk class is 1.00.
b. 2.1: The historical volatility ratio HVR for the interest rate risk class is 0.62.
14. 2.0 Interest Rate Tenors Correlation Matrix: The matrix on aggregated weighted sensitivities
or risk exposures shown below should be used.

2W 1M 3M 6M 1Y 2Y 3Y 5Y 10Y 15Y 20Y 30Y

2W 1.00 1.00 0.79 0.67 0.53 0.42 0.37 0.30 0.22 0.18 0.16 0.12

1M 1.00 1.00 0.79 0.67 0.53 0.42 0.37 0.30 0.22 0.18 0.16 0.12

3M 0.79 0.79 1.00 0.85 0.69 0.57 0.60 0.42 0.32 0.25 0.23 0.20

6M 0.67 0.67 0.85 1.00 0.86 0.76 0.59 0.59 0.47 0.40 0.37 0.32

1Y 0.53 0.53 0.69 0.86 1.00 0.93 0.87 0.77 0.63 0.57 0.54 0.50

2Y 0.42 0.42 0.57 0.76 0.93 1.00 0.98 0.90 0.77 0.70 0.67 0.63

3Y 0.37 0.37 0.50 0.69 0.87 0.98 1.00 0.96 0.84 0.78 0.75 0.77

5Y 0.30 0.30 0.42 0.59 0.77 0.90 0.96 1.00 0.93 0.89 0.86 0.82

10Y 0.22 0.22 0.32 0.47 0.63 0.77 0.84 0.93 1.00 0.98 0.96 0.94

15Y 0.18 0.18 0.25 0.40 0.57 0.70 0.78 0.89 0.98 1.00 0.99 0.98

131
 20Y 0.16 0.16 0.23 0.37 0.54 0.67 0.75 0.86 0.96 0.99 1.00 0.99

30Y 0.12 0.12 0.20 0.32 0.50 0.63 0.71 0.82 0.94 0.98 0.99 1.00

15. 2.1 Interest Rate Tenors Correlation Matrix: The matrix on aggregated weighted sensitivities
or risk exposures shown below should be used.

2W 1M 3M 6M 1Y 2Y 3Y 5Y 10Y 15Y 20Y 30Y

2W 1.00 0.63 0.59 0.47 0.31 0.22 0.18 0.14 0.09 0.06 0.04 0.05

1M 0.63 1.00 0.79 0.67 0.52 0.42 0.37 0.30 0.23 0.18 0.15 0.13

3M 0.59 0.79 1.00 0.84 0.68 0.56 0.50 0.42 0.32 0.26 0.24 0.21

6M 0.47 0.67 0.84 1.00 0.86 0.76 0.69 0.60 0.48 0.42 0.38 0.33

1Y 0.31 0.52 0.68 0.86 1.00 0.94 0.89 0.80 0.67 0.60 0.57 0.53

2Y 0.22 0.42 0.56 0.76 0.94 1.00 0.98 0.91 0.79 0.73 0.70 0.66

3Y 0.18 0.37 0.50 0.69 0.89 0.98 1.00 0.96 0.87 0.81 0.78 0.74

5Y 0.14 0.30 0.42 0.60 0.80 0.91 0.96 1.00 0.95 0.91 0.88 0.84

10Y 0.09 0.23 0.32 0.48 0.67 0.79 0.87 0.95 1.00 0.98 0.97 0.94

15Y 0.06 0.18 0.26 0.42 0.60 0.73 0.81 0.91 0.98 1.00 0.99 0.97

20Y 0.04 0.15 0.24 0.38 0.57 0.70 0.78 0.88 0.97 0.99 1.00 0.99

30Y 0.05 0.13 0.21 0.33 0.53 0.66 0.74 0.84 0.94 0.97 0.99 1.00

16. Correlation between Sub-Curve Pairs:

132
 a. 2.0: For sub-curves, the correlation between any two pairs 𝜙𝑖𝑗 in the same currency is
0.98.
b. 2.1: For sub-curves, the correlation between any two pairs 𝜙𝑖𝑗 in the same currency is
0.98.
17. IR/Inflation Rate/Volatility Correlation:
a. 2.0: For aggregated weighted sensitivities or risk exposures, the correlation between
the inflation rate and any yield for the same currency (and the correlation between the
inflation volatility and any interest-rate volatility for the same currency) is 29%.
b. 2.1: For aggregated weighted sensitivities or risk exposures, the correlation between
the inflation rate and any yield for the same currency (and the correlation between the
inflation volatility and any interest-rate volatility for the same currency) is 33%.
18. IR/Cross Currency/Inflation Volatility Correlation:
a. 2.0: For aggregated weighted sensitivities or risk exposures, the correlation between
the cross-currency basis swap spread and any yield or inflation rate for the same
currency is 20%.
b. 2.1: For aggregated weighted sensitivities or risk exposures, the correlation between
the cross-currency basis swap spread and any yield or inflation rate for the same
currency is 19%.
19. Correlation used for Different Currencies:
a. 2.0: The parameter

𝛾𝑏𝑐 = 23%

should be used for aggregating across different currencies.

b. 2.1: The parameter

𝛾𝑏𝑐 = 21%

should be used for aggregating across different currencies.

133
Credit Qualifying: Risk Weights

1. Credit Quality/Sector Risk Exposure: Sensitivities or risk exposures to an issuer/seniority

should first be assigned to a bucket according to the following table:

Bucket Credit
Sector
Number Quality

1 Sovereigns including Central Banks

2 Financials including Government-backed Financials

3 Basic Materials, Energy, Industrials

Investment
4 Grade (IG) Consumer

5 Technology, Telecommunications

Health-care, Utilities, Local Governments, Government-backed

6
Corporates (non-financial)

7 Sovereigns including Central Banks

8 Financials including Government-backed Financials

9 High Yield Basic Materials, Energy, Industrials

(HY) and
10 Non-rated Consumer

11 (NR) Technology, Telecommunications

Health-care, Utilities, Local Governments, Government-backed

12
Corporates (non-financial)

134
2. Position/Sensitivities under different Currencies: Sensitivities must be distinguished
depending upon the payment currency of the trade – such as Quanto CDS and non-quanto
CDS. No initial netting or aggregation is applied between position sensitivities from different
currencies – except as described for the situation below.
3. 2.0 Vertex Risk Weight by Bucket: The risk weights should be used for all vertexes (1Y, 2Y,
3Y, 5Y, 10Y) according to bucket, as set out in the following table.

Bucket Risk Weight

1 85

2 85

3 73

4 49

5 48

6 43

7 161

8 238

9 151

10 210

11 149

12 102

Residual 238

135
4. 2.1 Vertex Risk Weight by Bucket: The risk weights should be used for all vertexes (1Y, 2Y,
3Y, 5Y, 10Y) according to bucket, as set out in the following table.

Bucket Risk Weight

1 69

2 107

3 72

4 55

5 48

6 41

7 166

8 187

9 177

10 187

11 129

12 136

Residual 187

5. Base Correlation/Vega Risk Weight:

a. 2.0: The vega risk weight VRW for the credit risk class is 0.27. The Base Correlation
risk weight is 20 for all index families.
b. 2.1: The vega risk weight VRW for the credit risk class is 0.27. The Base Correlation
risk weight is 19 for all index families.

136
Credit Qualifying: Correlations

1. 2.0 Same Bucket Risk Factor Correlation: The correlation parameter 𝜌𝑘𝑙 applicable to
sensitivity or risk exposure pairs within the same bucket are set out in the following table:

Same Issuer/Seniority, different

Different Issuer/Seniority
Vertex or Currency

Aggregate
97% 45%
Sensitivities

Residual Bucket 50% 50%

2. 2.1 Same Bucket Risk Factor Correlation: The correlation parameter 𝜌𝑘𝑙 applicable to
sensitivity or risk exposure pairs within the same bucket are set out in the following table:

Same Issuer/Seniority, different

Different Issuer/Seniority
Vertex or Currency

Aggregate
96% 39%
Sensitivities

Residual Bucket 50% 50%

3. Quanto-Currency/Base Correlation Values: Here currency refers to the payment currency of

sensitivity if there are sensitivities to multiple payment currencies – such as Quanto CDS and
non-Quanto CDS – which will not be fully offset.

137
 a. 2.0: The correlation parameter 𝜌𝑘𝑙 applying to the Base Correlation risks across
different indexes/families is 10%.
b. 2.1: The correlation parameter 𝜌𝑘𝑙 applying to the Base Correlation risks across
different indexes/families is 5%.
4. 2.0 Different Bucket Risk Factor Calculations: The correlation bucket parameters applying to
sensitivities of risk exposure pairs across different non-residual buckets is set out in the
following table:

Bucket 1 2 3 4 5 6 7 8 9 10 11 12

1 1.00 0.42 0.39 0.39 0.40 0.38 0.39 0.34 0.37 0.39 0.37 0.31

2 0.42 1.00 0.44 0.45 0.47 0.45 0.33 0.40 0.41 0.44 0.43 0.37

3 0.39 0.44 1.00 0.43 0.45 0.43 0.32 0.35 0.41 0.42 0.40 0.36

4 0.39 0.45 0.43 1.00 0.47 0.44 0.40 0.34 0.39 0.43 0.39 0.36

5 0.40 0.47 0.45 0.47 1.00 0.47 0.31 0.35 0.40 0.44 0.42 0.37

6 0.38 0.45 0.43 0.44 0.47 1.00 0.30 0.34 0.38 0.40 0.39 0.38

7 0.39 0.33 0.32 0.40 0.31 0.30 1.00 0.28 0.31 0.31 0.30 0.26

8 0.34 0.40 0.35 0.34 0.35 0.34 0.28 1.00 0.34 0.35 0.33 0.30

9 0.37 0.41 0.41 0.39 0.40 0.38 0.31 0.34 1.00 0.40 0.37 0.32

10 0.39 0.44 0.42 0.43 0.44 0.40 0.31 0.35 0.40 1.00 0.40 0.35

11 0.37 0.43 0.40 0.39 0.42 0.39 0.30 0.33 0.37 0.40 1.00 0.34

12 0.31 0.37 0.36 0.36 0.37 0.38 0.26 0.30 0.32 0.35 0.34 1.00

138
5. 2.1 Different Bucket Risk Factor Calculations: The correlation bucket parameters applying to
sensitivities of risk exposure pairs across different non-residual buckets is set out in the
following table:

Bucket 1 2 3 4 5 6 7 8 9 10 11 12

1 1.00 0.38 0.36 0.36 0.39 0.35 0.34 0.32 0.34 0.33 0.34 0.31

2 0.38 1.00 0.41 0.41 0.43 0.40 0.29 0.38 0.38 0.38 0.38 0.34

3 0.36 0.41 1.00 0.41 0.42 0.39 0.30 0.34 0.39 0.37 0.38 0.35

4 0.36 0.41 0.41 1.00 0.43 0.40 0.28 0.33 0.37 0.38 0.38 0.34

5 0.39 0.43 0.42 0.43 1.00 0.42 0.31 0.35 0.38 0.39 0.41 0.36

6 0.35 0.40 0.39 0.40 0.42 1.00 0.27 0.32 0.34 0.35 0.36 0.33

7 0.34 0.29 0.30 0.28 0.31 0.27 1.00 0.24 0.28 0.27 0.27 0.26

8 0.32 0.38 0.34 0.33 0.35 0.32 0.24 1.00 0.33 0.32 0.32 0.29

9 0.34 0.38 0.39 0.37 0.38 0.34 0.28 0.33 1.00 0.35 0.35 0.33

10 0.33 0.38 0.37 0.38 0.39 0.35 0.27 0.32 0.35 1.00 0.36 0.32

11 0.34 0.38 0.38 0.38 0.41 0.36 0.27 0.32 0.35 0.36 1.00 0.33

12 0.31 0.34 0.35 0.34 0.36 0.33 0.26 0.29 0.33 0.32 0.33 1.00

Credit Non-Qualifying Risk

139
1. Non-Qualifying Credit Risk Spread: Sensitivities to credit-spread risk arising from non-
qualifying securitization positions are treated according to the risk weights and the
correlations as specified in the following paragraphs.
2. Credit Non-Qualifying Bucket Classifications: Sensitivities or risk exposures should first be
assigned to a bucket according to the following table:

Bucket Credit Quality Sector

1 Investment Grade (IG) RMBS/CMBS

High Yield (HY) and Not-Rated

2 RMBS/CMBS
(NR)

Residual

3. 2.0 Credit Non-Qualifying Risk Weights: The risk weights are set out in the following table:

Bucket Number Risk Weight

1 140

2 2000

Residual 2000

4. 2.1 Credit Non-Qualifying Risk Weights: The risk weights are set out in the following table:

Bucket Number Risk Weight

1 150

2 1200

140
 Residual 1200

a) 2.0: The vega risk weight VRW for Credit Non-qualifying is 0.27.
b) 2.1: The vega risk weight VRW for Credit Non-qualifying is 0.27.

Credit Non-Qualifying – Correlations

1. 2.0 Non-Qualifying Correlation – Same Bucket: For other buckets, the correlation parameter
𝜌𝑘𝑙 applicable to sensitivity or risk exposure pairs within the same bucket is set out in the
following table:

Same Underlying Names (more Different Underlying

than 80% Overlap in Notional Names (less than 80%
Terms) Overlap in Notional Terms)

Aggregate
57% 27%
Sensitivities

Residual Bucket 50% 50%

2. 2.1 Non-Qualifying Correlation – Same Bucket: For other buckets, the correlation parameter
𝜌𝑘𝑙 applicable to sensitivity or risk exposure pairs within the same bucket is set out in the
following table:

141
 Same Underlying Names (more Different Underlying
than 80% Overlap in Notional Names (less than 80%
Terms) Overlap in Notional Terms)

Aggregate
57% 20%
Sensitivities

Residual Bucket 50% 50%

3. 2.0 Non-Qualifying Correlations: Different Buckets: The correlation parameter 𝛾𝑏𝑐

applicable to sensitivity or risk exposure pairs across different buckets is set out in the
following table:

Correlation

Non-residual Bucket to Non-residual Bucket 0.21

4. 2.1 Non-Qualifying Correlations: Different Buckets: The correlation parameter 𝛾𝑏𝑐

applicable to sensitivity or risk exposure pairs across different buckets is set out in the
following table:

Correlation

Non-residual Bucket to Non-residual Bucket 0.16

Equity Risk Weights

142
1. Sector-Based Equity Risk Classification: Sensitivities or risk exposures should first be
assigned to a bucket according to the definitions below in the following table:

Bucket
Size Region Sector
Number

Consumer Goods and Services, Transportation and Storage,

1
Administrative and Service Activities, Utilities

2 Telecommunications and Industrials

Emerging
Markets Basic Materials, Energy, Agriculture, Manufacturing,
3
Mining, and Quarrying

Financials including Government-backed Financials, Real

4
Estate Activities, Technology
Large
Consumer Goods and Services, Transportation and Storage,
5
Administrative and Service Activities, Utilities

6 Telecommunications and Industrials

Developed
Markets Basic Materials, Energy, Agriculture, Manufacturing,
7
Mining, and Quarrying

Financials including Government-backed Financials, Real

8
Estate Activities, Technology

Emerging
9 All Sectors
Markets
Small
Developed
10 All Sectors
Markets

143
 11 All All Indexes, Funds, and ETF’s

12 All All Volatility Indexes

2. Large vs. Small Market Capitalization: Large is defined as a market capitalization equal to or
greater than USD 2 billion and small is defined as a market capitalization of less than USD 2
billion.
3. Global Aggregate of Market Cap: Market Capitalization is defined as the sum of the market
capitalizations of the same legal entity or a group of legal entities across all stock markets
globally.
4. Jurisdictions of the Developed Markets: The developed markets are defined as: Canada, US,
Mexico, the Euro area, the non-Euro area Western European countries – the UK, Norway,
Denmark, Sweden, and Switzerland – Japan, Oceania – Australia and New Zealand –
Singapore, and Hong Kong.
5. Determination of the Allocation Bucket: The sectors definition is the one generally used in
the market. When allocating an equity position in a particular bucket, the bank must prove
that the equity issuer’s most material activity indeed corresponds to the bucket’s definition.
Acceptable proof may be external provider’s information, or internal analysis.
6. Multi-national Cross-Sector Issuers: For multinational multi-sector equity issuers, the
allocation to a particular bucket must be done according to the most material region and the
sector the issuer operates in.
7. 2.0 Sector Based Risk Weight Assignment: If it is not possible to allocate a position to one of
these buckets – for example because data on categorical variables is not available – the
position must then be allocated to a residual bucket. Risk weights should be assigned to each
notional position as in the following table:

Bucket Risk Weight

1 25

144
 2 32

3 29

4 27

5 18

6 21

7 25

8 22

9 27

10 29

11 16

12 16

Residual 32

8. 2.1 Sector Based Risk Weight Assignment: If it is not possible to allocate a position to one of
these buckets – for example because data on categorical variables is not available – the
position must then be allocated to a residual bucket. Risk weights should be assigned to each
notional position as in the following table:

Bucket Risk Weight

1 24

2 30

3 31

145
 4 25

5 21

6 22

7 27

8 24

9 33

10 34

11 17

12 17

Residual 34

9. 2.0 Equity Risk Historical Volatility Ratio: The historical volatility ratio HVR for equity risk
class is 0.65.
10. 2.1 Equity Risk Historical Volatility Ratio: The historical volatility ratio HVR for equity risk
class is 0.59.
11. 2.0 Equity Class Vega Risk Weight: The vega risk weight VRW for the equity risk class is
0.28 for all buckets except bucket 12 for which the vega risk weight is 0.64.
12. 2.1 Equity Class Vega Risk Weight: The vega risk weight VRW for the equity risk class is
0.28 for all buckets except bucket 12 for which the vega risk weight is 0.63.

Equity Correlations

146
1. 2.0 Correlation within a Single Equity Bucket: The correlation parameter 𝜌𝑘𝑙 applicable to
sensitivity or risk exposure pairs within the same bucket are set out in the following table:

Bucket Correlation

1 14%

2 20%

3 19%

4 21%

5 24%

6 35%

7 34%

8 34%

9 20%

10 24%

11 62%

12 62%

Residual 0%

2. 2.1 Correlation within a Single Equity Bucket: The correlation parameter 𝜌𝑘𝑙 applicable to
sensitivity or risk exposure pairs within the same bucket are set out in the following table:

Bucket Correlation

147
 1 14%

2 20%

3 25%

4 23%

5 23%

6 32%

7 35%

8 32%

9 17%

10 16%

11 51%

12 51%

Residual 0%

3. 2.0 Correlations across Equity Buckets: The correlation parameters 𝜌𝑘𝑙 applicable to
sensitivity or risk exposure pairs across different non-residual buckets are set out in the
following table:

Bucket 1 2 3 4 5 6 7 8 9 10 11 12

1 1.00 0.15 0.14 0.16 0.10 0.12 0.10 0.11 0.13 0.09 0.17 0.17

2 0.15 1.00 0.16 0.17 0.10 0.11 0.10 0.11 0.14 0.09 0.17 0.17

148
 3 0.14 0.16 1.00 0.19 0.14 0.17 0.18 0.17 0.16 0.14 0.25 0.25

4 0.16 0.17 0.19 1.00 0.15 0.18 0.18 0.18 0.18 0.15 0.28 0.28

5 0.10 0.10 0.14 0.15 1.00 0.28 0.23 0.27 0.13 0.21 0.35 0.35

6 0.12 0.11 0.17 0.18 0.28 1.00 0.30 0.34 0.16 0.26 0.45 0.45

7 0.10 0.10 0.18 0.18 0.23 0.30 1.00 0.29 0.16 0.24 0.41 0.41

8 0.11 0.11 0.17 0.18 0.27 0.34 0.29 1.00 0.16 0.26 0.44 0.44

9 0.13 0.14 0.16 0.18 0.13 0.16 0.16 0.16 1.00 0.13 0.24 0.24

10 0.09 0.09 0.14 0.15 0.21 0.26 0.24 0.26 0.13 1.00 0.33 0.33

11 0.17 0.17 0.25 0.28 0.35 0.45 0.41 0.44 0.24 0.33 1.00 0.62

12 0.17 0.17 0.25 0.28 0.35 0.45 0.41 0.44 0.24 0.33 0.62 1.00

4. 2.1 Correlations across Equity Buckets: The correlation parameters 𝜌𝑘𝑙 applicable to
sensitivity or risk exposure pairs across different non-residual buckets are set out in the
following table:

Bucket 1 2 3 4 5 6 7 8 9 10 11 12

1 1.00 0.16 0.16 0.17 0.13 0.15 0.15 0.15 0.13 0.11 0.19 0.19

2 0.16 1.00 0.20 0.20 0.14 0.16 0.16 0.16 0.15 0.13 0.20 0.20

3 0.16 0.20 1.00 0.22 0.15 0.19 0.22 0.19 0.16 0.15 0.25 0.25

4 0.17 0.20 0.22 1.00 0.17 0.21 0.21 0.21 0.17 0.15 0.27 0.27

5 0.13 0.14 0.15 0.17 1.00 0.25 0.26 0.23 0.14 0.17 0.32 0.32

149
 6 0.15 0.16 0.19 0.21 0.25 1.00 0.30 0.31 0.16 0.21 0.38 0.38

7 0.15 0.16 0.22 0.21 0.23 0.30 1.00 0.29 0.16 0.21 0.38 0.38

8 0.15 0.16 0.19 0.21 0.26 0.31 0.29 1.00 0.17 0.21 0.39 0.39

9 0.13 0.15 0.16 0.17 0.14 0.16 0.16 0.17 1.00 0.13 0.21 0.21

10 0.11 0.13 0.15 0.15 0.17 0.21 0.21 0.21 0.13 1.00 0.25 0.25

11 0.19 0.20 0.25 0.27 0.32 0.38 0.38 0.39 0.21 0.25 1.00 0.51

12 0.19 0.20 0.25 0.27 0.32 0.38 0.38 0.39 0.21 0.25 0.51 1.00

Commodity Risk Weights

1. 2.0 Risk Weights for Commodity Buckets: The risk weights depend on the commodity type;
they are set out in the following table:

Bucket Commodity Risk Weight

1 Coal 19

2 Crude 20

3 Light Ends 17

4 Middle Distillates 18

5 Heavy Distillates 24

6 North American Natural Gas 20

150
 7 European Natural Gas 25

8 North American Power 41

9 European Power 24

10 Freight 91

11 Base Metals 20

12 Precious Metals 19

13 Grains 16

14 Softs 15

15 Livestock 10

16 Other 91

17 Indexes 17

2. 2.1 Risk Weights for Commodity Buckets: The risk weights depend on the commodity type;
they are set out in the following table:

Bucket Commodity Risk Weight

1 Coal 19

2 Crude 20

3 Light Ends 17

4 Middle Distillates 19

5 Heavy Distillates 24

151
 6 North American Natural Gas 22

7 European Natural Gas 26

8 North American Power 50

9 European Power 27

10 Freight 54

11 Base Metals 20

12 Precious Metals 20

13 Grains 17

14 Softs 14

15 Livestock 10

16 Other 54

17 Indexes 16

3. 2.0 Commodity Class Historical Volatility Ratio: The historical volatility ratio HVR for the
commodity risk class is 0.80.
4. 2.1 Commodity Class Historical Volatility Ratio: The historical volatility ratio HVR for the
commodity risk class is 0.80.
5. 2.0 Commodity Class Vega Risk Weight: The vega risk weight VRW for the commodity risk
class is 0.38.
6. 2.1 Commodity Class Vega Risk Weight: The vega risk weight VRW for the commodity risk
class is 0.38.

152
Commodity Correlations

1. 2.0 Commodity Correlations within the same Bucket: The correlation parameters 𝜌𝑘𝑙
applicable to sensitivity or risk exposure pairs within the same bucket are set out in the
following table:

Bucket Correlation

1 0.30

2 0.97

3 0.93

4 0.98

5 0.97

6 0.92

7 1.00

8 0.58

9 1.00

10 0.10

11 0.55

12 0.64

13 0.71

14 0.22

153
 15 0.29

16 0.00

17 0.21

2. 2.1 Commodity Correlations within the same Bucket: The correlation parameters 𝜌𝑘𝑙
applicable to sensitivity or risk exposure pairs within the same bucket are set out in the
following table:

Bucket Correlation

1 0.27

2 0.97

3 0.92

4 0.97

5 0.99

6 1.00

7 1.00

8 0.40

9 0.73

10 0.13

11 0.53

12 0.64

154
 13 0.63

14 0.26

15 0.26

16 0.00

17 0.38

3. 2.0 Commodity Correlations among Different Buckets: The correlation parameters 𝛾𝑏𝑐
applicable to sensitivity or risk exposure pairs across different buckets are set out in the
following table:

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17

1 1.00 0.18 0.15 0.20 0.25 0.08 0.09 0.20 0.27 0.00 0.15 0.02 0.06 0.07 -0.04 0.00 0.06

2 0.18 1.00 0.89 0.94 0.93 0.32 0.22 0.27 0.24 0.09 0.45 0.21 0.32 0.28 0.17 0.00 0.37

-
3 0.15 0.89 1.00 0.87 0.88 0.25 0.16 0.19 0.12 0.10 0.26 0.19 0.17 0.10 0.00 0.27
0.01

4 0.20 0.94 0.87 1.00 0.92 0.29 0.22 0.26 0.19 0.00 0.32 0.05 0.20 0.22 0.13 0.00 0.28

5 0.25 0.93 0.88 0.92 1.00 0.30 0.26 0.22 0.28 0.12 0.42 0.23 0.28 0.19 0.17 0.00 0.34

-
6 0.08 0.32 0.25 0.29 0.30 1.00 0.13 0.57 0.05 0.14 0.57 0.13 0.17 0.01 0.00 0.26
0.02

7 0.09 0.22 0.16 0.22 0.26 0.13 1.00 0.07 0.80 0.19 0.16 0.05 0.17 0.18 0.00 0.00 0.18

8 0.20 0.27 0.19 0.26 0.22 0.57 0.07 1.00 0.13 0.06 0.16 0.03 0.10 0.12 0.06 0.00 0.23

9 0.27 0.24 0.12 0.19 0.28 0.05 0.80 0.13 1.00 0.15 0.17 0.05 0.15 0.13 -0.03 0.00 0.13

10 0.00 0.09 0.10 0.00 0.12 0.14 0.19 0.06 0.15 1.00 0.07 0.07 0.17 0.10 0,02 0.00 0.11

155
11 0.15 0.45 0.26 0.32 0.42 0.57 0.16 0.16 0.17 0.07 1.00 0.34 0.20 0.21 0.16 0.00 0.27

- -
12 0.02 0.21 0.05 0.23 0.05 0.03 0.05 0.07 0.34 1.00 0.17 0.26 0.11 0.00 0.14
0.01 0.02

13 0.06 0.32 0.19 0.20 0.28 0.13 0.17 0.10 0.15 0.17 0.20 0.17 1.00 0.35 0.09 0.00 0.22

14 0.07 0.28 0.17 0.22 0.19 0.17 0.18 0.12 0.13 0.10 0.27 0.26 0.35 1.00 0.06 0.00 0.20

-
15 -0.04 0.17 0.10 0.13 0.17 1.00 0.00 0.06 0.02 0.16 0.11 0.09 0.06 1.00 0.00 0.16
0.03

16 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00

17 0.06 0.37 0.27 0.28 0,34 0.26 0.18 0.23 0.13 0.11 0.21 0.14 0.22 0.20 0.16 0.00 1.00

4. 2.1 Commodity Correlations among Different Buckets: The correlation parameters 𝛾𝑏𝑐
applicable to sensitivity or risk exposure pairs across different buckets are set out in the
following table:

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17

1 1.00 0.16 0.11 0.19 0.22 0.12 0.22 0.02 0.27 0.08 0.11 0.05 0.04 0.06 0.01 0.00 0.10

2 0.16 1.00 0.89 0.94 0.93 0.32 0.24 0.19 0.21 0.06 0.39 0.23 0.39 0.29 0.13 0.00 0.66

3 0.11 0.89 1.00 0.87 0.88 0.17 0.17 0.13 0.12 0.03 0.24 0.04 0.27 0.19 0.08 0.00 0.61

4 0.19 0.94 0.87 1.00 0.92 0.37 0.27 0.21 0.21 0.03 0.36 0.16 0.27 0.28 0.09 0.00 0.64

5 0.22 0.93 0.88 0.92 1.00 0.29 0.26 0.19 0.23 0.10 0.40 0.27 0.38 0.30 0.15 0.00 0.64

6 0.12 0.32 0.17 0.37 0.29 1.00 0.19 0.60 0.18 0.09 0.22 0.09 0.14 0.16 0.10 0.00 0.37

7 0.22 0.24 0.17 0.27 0.26 0.19 1.00 0.06 0.68 0.16 0.21 0.10 0.24 0.25 -0.01 0.00 0.27

8 0.02 0.19 0.13 0.21 0.19 0.60 0.06 1.00 0.12 0.01 0.10 0.03 0.02 0.07 0.10 0.00 0.21

156
9 0.27 0.21 0.12 0.21 0.23 0.18 0.68 0.12 1.00 0.05 0.16 0.03 0.19 0.16 -0.01 0.00 0.19

10 0.08 0.06 0.03 0.03 0.10 0.09 0.16 0.01 0.05 1.00 0.08 0.04 0.05 0.11 0.02 0.00 0.00

11 0.11 0.39 0.24 0.36 0.40 0.22 0.21 0.10 0.16 0.08 1.00 0.34 0.19 0.22 0.15 0.00 0.34

12 0.05 0.23 0.04 0.16 0.27 0.09 0.10 0.03 0.03 0.04 0.34 1.00 0.14 0.26 0.09 0.00 0.20

13 0.04 0.39 0.27 0.27 0.38 0.14 0.24 0.02 0.19 0.05 0.19 0.14 1.00 0.30 0.16 0.00 0.40

14 0.06 0.29 0.19 0.28 0.30 0.16 0.25 0.07 0.16 0.11 0.22 0.26 0.30 1.00 0.09 0.00 0.30

- -
15 0.01 0.13 0.08 0.09 0.15 0.10 0.10 0.02 0.15 0.09 0.16 0.09 1.00 0.00 0.16
0.01 0.01

16 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00

17 0.10 0.66 0.61 0.64 0.64 0.37 0.27 0.21 0.19 0.00 0.34 0.20 0.40 0.30 0.16 0.00 1.00

Foreign Exchange Risk

1. 2.0 Foreign Exchange Risk Weights: A unique risk weight equal to 8.2 applies to all FX
sensitivities or risk exposures. The historical volatility ratio HVR for the FX risk class is
0.60. The vega risk weight VRW for the FX volatility 0.33.
2. 2.1 Foreign Exchange Risk Weights: A unique risk weight equal to 8.1 applies to all FX
sensitivities or risk exposures. The historical volatility ratio HVR for the FX risk class is
0.63. The vega risk weight VRW for the FX volatility 0.30.
3. 2.0 Foreign Exchange Correlations: A unique correlation 𝜌𝑘𝑙 equal to 0.5 applies to all pairs
of FX sensitivities or risk exposures.
4. 2.1 Foreign Exchange Correlations: A unique correlation 𝜌𝑘𝑙 equal to 0.5 applies to all pairs
of FX sensitivities or risk exposures.

157
5. Single Bucket Foreign Exchange Sensitivities: All foreign exchange sensitivities are
considered to be within a single bucket within the FX risk class, so no inter-bucket
aggregation is necessary. Note that the cross-bucket curvature calculations are still required
on a single bucket.

Concentration Thresholds

1. Asset Class/Bucket Concentration Thresholds: The concentration thresholds in this section

are defined for the asset-class-specific buckets specified earlier. For those cases where the
same concentration threshold applies to the related range of buckets, the tables in this section
specify a precise range of applicable buckets in the Bucket column and give a narrative
description of that group of buckets in the Risk Group column.
2. 2.0 Interest Rate Risk - Delta Concentration Thresholds: The delta concentration thresholds
for interest rate risk – inclusive of inflation risk – are given by the currency group:

Currency Risk Group Concentration Threshold (USD mm/bp)

High Volatility 8

Regular Volatility, Well Traded 230

Regular Volatility, less Well Traded 28

Low Volatility 82

3. 2.1 Interest Rate Risk - Delta Concentration Thresholds: The delta concentration thresholds
for interest rate risk – inclusive of inflation risk – are given by the currency group:

158
 Currency Risk Group Concentration Threshold (USD mm/bp)

High Volatility 12

Regular Volatility, Well Traded 210

Regular Volatility, less Well Traded 27

Low Volatility 170

4. Concentration Threshold Currency Risk Group: The Currency Risk Groups used in
establishing concentration thresholds for Interest Rate Risk are as follows:
a. High Volatility => All other currencies
b. Regular Volatility, Well Traded => USD, EUR, GBP
c. Regular Volatility, Less Well Traded => AUD, CAD, CHF, DKK, HKD, KRW,
NOK, NZD, SEK, SGD, TWD
d. Low Volatility => JPY
5. 2.0 Credit Spread Risk – Delta Concentration Thresholds: The delta concentration thresholds
for credit spread risk are given by credit spread risk are given by the Credit Risk Group and
Bucket.

Bucket (s) Credit Risk Group Concentration Threshold (USD mm/bp)

Qualifying

Sovereigns including Central

1, 7 0.95
Banks

2-6, 8-12 Corporate Entities 0.29

Residual Non-Classified 0.29

Non-Qualifying

159
 1 IG (RMBS and CMBS) 9.50

2 HY/non-rated (RMBS and CMBS) 0.50

Residual Not Classified 0.50

6. 2.1 Credit Spread Risk – Delta Concentration Thresholds: The delta concentration thresholds
for credit spread risk are given by credit spread risk are given by the Credit Risk Group and
Bucket.

Bucket (s) Credit Risk Group Concentration Threshold (USD mm/bp)

Qualifying

Sovereigns including Central

1, 7 1.00
Banks

2-6, 8-12 Corporate Entities 0.24

Residual Non-Classified 0.24

Non-Qualifying

1 IG (RMBS and CMBS) 9.50

2 HY/non-rated (RMBS and CMBS) 0.50

Residual Not Classified 0.50

7. 2.0 Equity Risk – Delta Concentration Thresholds: The delta concentration thresholds for
equity risk are given by bucket.

Bucket (s) Equity Risk Group Concentration Threshold (USD mm/bp)

160
 1-4 Emerging Markets – Large Cap 3.3

5-8 Developed Markets – Large Cap 30.0

9 Emerging Markets – Small Cap 0.6

10 Developed Markets – Small Cap 2.3

Indexes, Funds, ETF’s, and

11-12 900.0
Volatility Indexes

Residual Not Classified 0.6

8. 2.1 Equity Risk – Delta Concentration Thresholds: The delta concentration thresholds for
equity risk are given by bucket.

Bucket (s) Equity Risk Group Concentration Threshold (USD mm/bp)

1-4 Emerging Markets – Large Cap 8.4

5-8 Developed Markets – Large Cap 26.0

9 Emerging Markets – Small Cap 1.8

10 Developed Markets – Small Cap 1.9

Indexes, Funds, ETF’s, and

11-12 540.0
Volatility Indexes

Residual Not Classified 1.8

9. 2.0 Commodity Risk – Delta Concentration Thresholds: The delta concentration thresholds
for commodity risk are given by:

161
 Bucket (s) Commodity Risk Group Concentration Threshold (USD mm/bp)

1 Coal 1,400

2 Crude Oil 20,000

3-5 Oil Fractions 3,500

6-7 Natural Gas 6,400

8-9 Power 2,500

10 Freight – Dry or Wet 300

11 Base Metals 2,900

12 Precious Metals 7,600

13-15 Agricultural 3,900

16 Other 300

17 Indexes 12,000

10. 2.1 Commodity Risk – Delta Concentration Thresholds: The delta concentration thresholds
for commodity risk are given by:

Bucket (s) Commodity Risk Group Concentration Threshold (USD mm/bp)

1 Coal 700

2 Crude Oil 3,600

3-5 Oil Fractions 2,700

6-7 Natural Gas 2,600

162
 8-9 Power 1,900

10 Freight – Dry or Wet 52

11 Base Metals 2,000

12 Precious Metals 3,200

13-15 Agricultural 1,100

16 Other 52

17 Indexes 5,200

11. 2.0 FX Risk – Delta Concentration Thresholds: The delta concentration thresholds for FX
risk are given by the FX Risk Group:

FX Risk Group Concentration Threshold (USD mm/bp)

Category 1 8,400

Category 2 1,900

Category 3 560

12. 2.1 FX Risk – Delta Concentration Thresholds: The delta concentration thresholds for FX
risk are given by the FX Risk Group:

FX Risk Group Concentration Threshold (USD mm/bp)

Category 1 9,700

Category 2 2,900

163
 Category 3 450

13. FX Risk Concentration Threshold Classifications: Currencies were placed in three categories
as those for delta risk weights, constituted as follows:
a. Category 1 => Significantly Material – USD, EUR, JPY, GBP, CAD, AUD, CHF
b. Category 2 => Frequently Traded – BRL, CNY, HKD, INR, KRW, MXN, NOK,
NZD, RUB, SEK, SGD, TRY, ZAR
c. Category 3 => Others – All other currencies
14. 2.0 Interest Rate Risk - Vega Concentration Thresholds: The Vega Concentration thresholds
for Interest Rate Risk are:

Currency Risk Group Concentration Threshold (USD mm/bp)

High Volatility 110

Regular Volatility, Well Traded 2,700

Regular Volatility, less Well Traded 150

Low Volatility 960

15. 2.1 Interest Rate Risk - Vega Concentration Thresholds: The Vega Concentration thresholds
for Interest Rate Risk are:

Currency Risk Group Concentration Threshold (USD mm/bp)

High Volatility 120

Regular Volatility, Well Traded 2,200

Regular Volatility, less Well Traded 190

164
 Low Volatility 770

The currency risk groups used in establishing the concentration thresholds correspond to the
Concentration Threshold Currency Risk Group above.

16. 2.0 Credit Spread Risk - Vega Concentration Thresholds: The Vega Concentration thresholds
for Credit Spread Risk are:

Credit Risk Group Concentration Threshold (USD mm/bp)

Qualifying 290

Non-Qualifying 65

17. 2.1 Credit Spread Risk - Vega Concentration Thresholds: The Vega Concentration thresholds
for Credit Spread Risk are:

Credit Risk Group Concentration Threshold (USD mm/bp)

Qualifying 250

Non-Qualifying 54

18. 2.0 Equity Risk - Vega Concentration Thresholds: The Vega Concentration thresholds for
Equity Risk are:

Bucket (s) Equity Risk Group Concentration Threshold (USD mm/bp)

1-4 Emerging Markets – Large Cap 700

5-8 Developed Markets – Large Cap 7,300

165
 9 Emerging Markets – Small Cap 70

10 Developed Markets – Small Cap 300

Indexes, Funds, ETF’s, and

11-12 21,000
Volatility Indexes

Residual Not Classified 70

19. 2.1 Equity Risk - Vega Concentration Thresholds: The Vega Concentration thresholds for
Equity Risk are:

Bucket (s) Equity Risk Group Concentration Threshold (USD mm/bp)

1-4 Emerging Markets – Large Cap 20

5-8 Developed Markets – Large Cap 2,300

9 Emerging Markets – Small Cap 43

10 Developed Markets – Small Cap 250

Indexes, Funds, ETF’s, and

11-12 8,100
Volatility Indexes

Residual Not Classified 43

20. 2.0 Commodity Risk - Vega Concentration Thresholds: The Vega Concentration thresholds
for Commodity Risk are:

Bucket (s) Commodity Risk Group Concentration Threshold (USD mm/bp)

1 Coal 250

166
 2 Crude Oil 2,000

3-5 Oil Fractions 510

6-7 Natural Gas 1,900

8-9 Power 870

10 Freight – Dry or Wet 220

11 Base Metals 450

12 Precious Metals 740

13-15 Agricultural 370

16 Other 220

17 Indexes 430

21. 2.1 Commodity Risk - Vega Concentration Thresholds: The Vega Concentration thresholds
for Commodity Risk are:

Bucket (s) Commodity Risk Group Concentration Threshold (USD mm/bp)

1 Coal 250

2 Crude Oil 1,800

3-5 Oil Fractions 320

6-7 Natural Gas 2,200

8-9 Power 780

10 Freight – Dry or Wet 99

167
 11 Base Metals 420

12 Precious Metals 650

13-15 Agricultural 570

16 Other 99

17 Indexes 330

22. 2.0 FX Risk - Vega Concentration Thresholds: The Vega Concentration thresholds for FX
Risk are:

FX Risk Group Concentration Threshold (USD mm/bp)

Category 1 – Category 1 4,000

Category 1 – Category 2 1,900

Category 1 - Category 3 320

Category 2 – Category 2 120

Category 3 – Category 3 110

Category 3 - Category 3 110

23. 2.1 FX Risk - Vega Concentration Thresholds: The Vega Concentration thresholds for FX
Risk are:

FX Risk Group Concentration Threshold (USD mm/bp)

Category 1 – Category 1 2,000

168
 Category 1 – Category 2 1,000

Category 1 - Category 3 320

Category 2 – Category 2 410

Category 3 – Category 3 210

Category 3 - Category 3 150

The Currency categories used in establishing Concentration Thresholds for FX are identified
under FX Risk – Delta Concentration Thresholds.

24. 2.0 Correlation between Risk Classes within Products: The correlation parameters 𝜓𝑟𝑠
applying to initial margin risk classes within a single product class are set out in the
following table.

Credit
Interest Credit
Class Non- Equity FX Commodity
Rate Qualifying
Qualifying

Interest
1.00 0.28 0.18 0.18 0.30 0.22
Rate

Credit
0.28 1.00 0.30 0.66 0.46 0.27
Qualifying

Credit
Non- 0.18 0.30 1.00 0.23 0.25 0.18
Qualifying

Equity 0.18 0.66 0.23 1.00 0.39 0.24

FX 0.30 0.46 0.25 0.39 1.00 0.32

169
 Commodity 0.22 0.27 0.18 0.24 0.32 1.00

25. 2.1 Correlation between Risk Classes within Products: The correlation parameters 𝜓𝑟𝑠
applying to initial margin risk classes within a single product class are set out in the
following table.

Credit
Interest Credit
Class Non- Equity FX Commodity
Rate Qualifying
Qualifying

Interest
1.00 0.25 0.15 0.19 0.30 0.26
Rate

Credit
0.25 1.00 0.26 0.65 0.45 0.24
Qualifying

Credit
Non- 0.15 0.26 1.00 0.23 0.25 0.18
Qualifying

Equity 0.19 0.65 0.23 1.00 0.39 0.24

FX 0.30 0.45 0.25 0.39 1.00 0.32

Commodity 0.26 0.24 0.18 0.24 0.32 1.00

Additional Initial Margin Expressions

170
1. Additional Initial Margin – Standardized Expressions: Standardized formulas for calculating
additional initial margin are as follows:

𝐴𝑑𝑑𝑖𝑡𝑖𝑜𝑛𝑎𝑙𝐼𝑛𝑖𝑡𝑖𝑎𝑙𝑀𝑎𝑟𝑔𝑖𝑛
= 𝐴𝑑𝑑𝑂𝑛𝐼𝑀 + (𝑀𝑆𝑅𝐴𝑇𝐸𝑆𝐹𝑋 − 1)𝑆𝐼𝑀𝑀𝑅𝐴𝑇𝐸𝑆𝐹𝑋
+ (𝑀𝑆𝐶𝑅𝐸𝐷𝐼𝑇 − 1)𝑆𝐼𝑀𝑀𝐶𝑅𝐸𝐷𝐼𝑇 + (𝑀𝑆𝐸𝑄𝑈𝐼𝑇𝑌 − 1)𝑆𝐼𝑀𝑀𝐸𝑄𝑈𝐼𝑇𝑌
+ (𝑀𝑆𝐶𝑂𝑀𝑀𝑂𝐷𝐼𝑇𝑌 − 1)𝑆𝐼𝑀𝑀𝐶𝑂𝑀𝑀𝑂𝐷𝐼𝑇𝑌

where 𝐴𝑑𝑑𝑂𝑛𝐼𝑀 is defined as follows:

𝐴𝑑𝑑𝑂𝑛𝐼𝑀 = 𝐴𝑑𝑑𝑂𝑛𝐹𝑖𝑥𝑒𝑑 + 𝐴𝑑𝑑𝑂𝑛𝐹𝑎𝑐𝑡𝑜𝑟𝑃 𝑁𝑜𝑡𝑖𝑜𝑛𝑎𝑙𝑃

where 𝐴𝑑𝑑𝑂𝑛𝐹𝑖𝑥𝑒𝑑 is a fixed add-on amount, 𝐴𝑑𝑑𝑂𝑛𝐹𝑎𝑐𝑡𝑜𝑟𝑃 is the add-on factor for each
affected product 𝑃 expressed as a percentage of the notional (e.g., 5%), and 𝑁𝑜𝑡𝑖𝑜𝑛𝑎𝑙𝑃 is the
total notional of the product – sum of the absolute trade notionals. In such use – where a
variable notional is involved – current notional amount should be used.
2. Multiplicative Scales for Product Classes: The four variables - 𝑀𝑆𝑅𝐴𝑇𝐸𝑆𝐹𝑋 , 𝑀𝑆𝐶𝑅𝐸𝐷𝐼𝑇 ,
𝑀𝑆𝐸𝑄𝑈𝐼𝑇𝑌 , and 𝑀𝑆𝐶𝑂𝑀𝑀𝑂𝐷𝐼𝑇𝑌 are the four multiplicative scales for the four product classes
RatesFX, Credit, Equity, and Commodity. Their values can be individually specified to be
more than 1.0 – with 1.0 being the default and the minimum value.

Structure of the Methodology

1. Six Classes of Risk Factors: There are six risk classes – Interest Rate, Credit (Qualifying),
Credit (Non-Qualifying), Equity, Commodity, and FX – and the margin for each risk class is

171
 defined to be the sum of the Delta Margin, the Vega Margin, the Curvature Margin, and the
Base Correlation Margin – if applicable – for that risk class.
2. Four Classes of Sensitivity Margins: That is
𝐼𝑀𝑋 = 𝐷𝑒𝑙𝑡𝑎𝑀𝑎𝑟𝑔𝑖𝑛𝑋 + 𝑉𝑒𝑔𝑎𝑀𝑎𝑟𝑔𝑖𝑛𝑋 + 𝐶𝑢𝑟𝑣𝑎𝑡𝑢𝑟𝑒𝑀𝑎𝑟𝑔𝑖𝑛𝑋 +
𝐵𝑎𝑠𝑒𝐶𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑖𝑜𝑛𝑀𝑎𝑟𝑔𝑖𝑛𝑋

for each risk class 𝑋, where the 𝐵𝑎𝑠𝑒𝐶𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑖𝑜𝑛𝑀𝑎𝑟𝑔𝑖𝑛𝑋 is only present in the Credit
Qualifying risk class.
3. Four Classes of Products Needing Margin: There are four product classes:
a. Interest Rates and Foreign Exchange (RatesFX)
b. Credit
c. Equity
d. Commodity
4. Product Class per Marginable Trade: Every trade is assigned to an individual product class
and SIMM is considered separately for each product class.
5. Isolating Risk Factors across Products: Buckets are still defined in risk terms, but within each
product class the risk class takes its component risks only from trades of that product class.
For example, equity derivatives would have risk only in the interest rate risk class as well as
from the equity risk class; but all those risks are kept separate from the risks of the trades in
the RatesFX product class.
6. Product SIMM from Risk Factor IM: Within each product class, the initial margin (IM) for
each of the risk classes is calculated as above. The total margin for that product class is given
by

𝑆𝐼𝑀𝑀𝑃𝑅𝑂𝐷𝑈𝐶𝑇 = √∑ 𝐼𝑀𝑟 2 + ∑ ∑ 𝜓𝑟𝑠 𝐼𝑀𝑟 𝐼𝑀𝑠

𝑟 𝑟 𝑠≠𝑟

where 𝑃𝑅𝑂𝐷𝑈𝐶𝑇 is one of the four product classes above, and the sums on 𝑟 and 𝑠 are taken
over the six risk classes. The correlation matrix 𝜓𝑟𝑠 of the correlations between the risk
classes is given earlier.

172
7. Portfolio SIMM as Linear Sum: The total SIMM is the sum of these four product class
SIMM values:

𝑆𝐼𝑀𝑀 = 𝑆𝐼𝑀𝑀𝑅𝐴𝑇𝐸𝑆𝐹𝑋 + 𝑆𝐼𝑀𝑀𝐶𝑅𝐸𝐷𝐼𝑇 + 𝑆𝐼𝑀𝑀𝐸𝑄𝑈𝐼𝑇𝑌 + 𝑆𝐼𝑀𝑀𝐶𝑂𝑀𝑀𝑂𝐷𝐼𝑇𝑌

8. Product Specific SIMM Add-On: The SIMM equation can be extended to incorporate
notional based add-ons for specified products and/or multipliers to the individual product
class SIMM values. The section on SIMM add-ons contains the modified version of SIMM
in that case.

Interest Rate Risk Delta Margin

1. Approach for IR Delta Margin: The following step-by-step approach to capture delta risk
should be applied to capture delta risk for the interest-rate risk class only.
2. Sensitivity to Tenor/Risk Factor: Find a net sensitivity across instruments to each risk factor
(𝑘, 𝑖) where 𝑘 is the rates tenor and 𝑖 is the index name of the sub-yield curve as defined in
the sections that outline the interest rate risk class.
3. Risk Weight applied to the Sensitivity: Weight the net sensitivity 𝑠𝑘,𝑖 to each risk factor (𝑘, 𝑖)
by the corresponding risk weight 𝑅𝑊𝑘 according to the vertex structure laid out in the section
on Interest Rate Risk Weights.
4. Risk Weighted Vertex Sensitivity Expression:

𝑊𝑆𝑘,𝑖 = 𝑅𝑊𝑘 𝑠𝑘,𝑖 𝐶𝑅𝑏

is where 𝐶𝑅 is the concentration risk factor defined as

173
 ∑𝑘,𝑖 𝑠𝑘,𝑖
𝐶𝑅𝑏 = max (1, √ )
𝑇𝑏

for the concentration threshold 𝑇𝑏 defined for each currency 𝑏.

5. Treating Inflation/Cross Currency Sensitivities: Note that inflation sensitivities to currency 𝑏
are included in ∑𝑘,𝑖 𝑠𝑘,𝑖 but cross-currency basis swap sensitivities are not. Neither should the
cross-currency basis swap sensitivities be scaled by the concentration factor.
6. Single Curve Sensitivity Roll Up: The weighted sensitivities should then be aggregated
within each currency. The sub-curve correlations 𝜙𝑖𝑗 and the tenor correlation parameters 𝜌𝑘𝑙
are set out in the Section on Interest Rate Risk Correlations.
7. Single Curve Composite Sensitivity - Expression:

𝐾 = √∑ 𝑊𝑆𝑘,𝑖 2 + ∑ ∑ 𝜙𝑖𝑗 𝜌𝑘𝑙 𝑊𝑆𝑘,𝑙 𝑊𝑆𝑖,𝑗

𝑖,𝑘 𝑖,𝑘 (𝑗,𝑙)≠(𝑖,𝑘)

8. Multiple Currency Weighted Roll Up: Delta amounts should then be aggregated across
currencies within the risk class. The correlation parameters 𝛾𝑏𝑐 applicable are set out in the
Section on Interest Rate Risk factor correlations.
9. Multiple Currency Delta Margin Expression:

𝐷𝑒𝑙𝑡𝑎𝑀𝑎𝑟𝑔𝑖𝑛 = √∑ 𝐾𝑏 2 + ∑ ∑ 𝛾𝑏𝑐 𝑔𝑏𝑐 𝑆𝑏 𝑆𝑐

𝑏 𝑏 𝑐≠𝑏

where

𝑆𝑏 = max (min {∑ 𝑊𝑆𝑖,𝑘 , 𝐾𝑏 } , −𝐾𝑏 )

𝑖,𝑘

174
 and

min(𝐶𝑅𝑏 , 𝐶𝑅𝑐 )
𝑔𝑏𝑐 =
max(𝐶𝑅𝑏 , 𝐶𝑅𝑐 )

for all currencies 𝑏 and 𝑐.

Non Interest Rate Risk Classes

1. Non-IR Delta Margin Approaches: The following step-by-step approach to capture delta risk
should be applied separately to each risk class other than interest rates.
2. Sensitivity to Tenor/Risk Factor: Find the net sensitivity across instruments to each risk
factor 𝑘, which are defined in the sections for each risk class.
3. Risk Weight Applicability to Sensitivity: Weight the net sensitivity 𝑠𝑘 to each risk factor 𝑘
by the corresponding risk weight 𝑅𝑊𝑘 according to the bucketing structure for each risk class
set out in the Section Credit Qualifying Risk.
4. Risk Weighted Vertex Sensitivity Expression:

𝑊𝑆𝑘 = 𝑅𝑊𝑘 𝑠𝑘 𝐶𝑅𝑘

is where 𝐶𝑅𝑘 is the concentration risk factor defined as

|∑𝑗 𝑠𝑗 |
𝐶𝑅𝑘 = max (1, √ )
𝑇𝑏

175
 for credit spread risk with the sum 𝑗 taken over all the issuers and seniorities as the risk factor
𝑘 irrespective of the tenor of the payment currency, and

|𝑠𝑘 |
𝐶𝑅𝑘 = max (1, √ )
𝑇𝑏

for equity, commodity, and FX risk where 𝑇𝑏 is the concentration threshold for the bucket –
or FX category - 𝑏 as given in the appropriate section.
5. Incorporating the Base Correlation Risk: Note that the base correlation sensitivities are not
included in the concentration risk, and the concentration risk for these factors should be taken
as 1.
6. Roll Up within Risk Factor: Weighted sensitivities should then be aggregated within each
bucket. The buckets and the correlation parameters applicable to each risk class are set out in
the section on Credit Qualifying Risk.
7. Single Risk Factor Composition Expression:

𝐾 = √∑ 𝑊𝑆𝑘 2 + ∑ ∑ 𝜌𝑘𝑙 𝑓𝑘𝑙 𝑊𝑆𝑘 𝑊𝑆𝑙

𝑘 𝑘 𝑙≠𝑘

where

min(𝐶𝑅𝑘 , 𝐶𝑅𝑙 )
𝑓𝑘𝑙 =
max(𝐶𝑅𝑘 , 𝐶𝑅𝑙 )

8. Roll Up across Risk Factors: Delta margin amounts should then be aggregated across buckets
in each risk class. The correlation parameters 𝛾𝑏𝑐 applicable to each risk class are set out
earlier.
9. Cross Risk Factor Composition - Expression:

176
 𝐷𝑒𝑙𝑡𝑎𝑀𝑎𝑟𝑔𝑖𝑛 = √∑ 𝐾𝑏 2 + ∑ ∑ 𝛾𝑏𝑐 𝑆𝑏 𝑆𝑐 + 𝐾𝑅𝐸𝑆𝐼𝐷𝑈𝐴𝐿
𝑏 𝑏 𝑐≠𝑏

where

𝑆𝑏 = max (min {∑ 𝑊𝑆𝑖,𝑘 , 𝐾𝑏 } , −𝐾𝑏 )

𝑖,𝑘

for all risk factors in bucket 𝑏.

10. Margin Requirements for Volatility Instruments: Instruments that are option or include an
option – including pre-payment – or have volatility sensitivity – instruments subject to
optionality – are subject to additional margin risks such as vega and curvature risk – as
described down below. Instruments not subject to optionality with no volatility sensitivity are
not subject to vega risk and curvature risk.
11. Approaches for Vega Risk Exposure: The following step-by-step to capture vega risk
exposure should be applied separately to each risk class.
12. Risk Factor/Tenor ATM Volatility: For interest-rate and credit instruments, the volatility 𝜎𝑘𝑗
for risk-factor 𝑘 and maturity 𝑗 is defined to be the implied at-the-money (ATM) volatility of
the swaption with expiry time equal to the tenor 𝑘 at some swap maturity 𝑗. The volatility can
be quoted as normal, log-normal, or something similar.
13. Inflation Risk Factor ATM Volatility: In the case where 𝑘 is the inflation risk factor, the
inflation volatility 𝜎𝑘𝑗 of the inflation swaption of type 𝑗 is defined to the at-the-money
volatility of the swaption, where the type 𝑗 comprises an initial inflation observation date and
a final inflation observation date.
14. Buckets Mirroring Interest-Rate Factors: The option expiry date shall be defined to be the
final inflation observation date, and risk should be defined on a set of option expiries equal to
the same-tenor bucket as the interest rate delta. The volatility can be quoted as normal, log-
normal, or something similar.

177
15. Volatility for Equity/FX/Commodity: For Equity, FX, and Commodity instruments, the
volatility 𝜎𝑘𝑗 of the risk factor 𝑘 at each vol-tenor 𝑗 is given by the expression

𝑅𝑊𝑘 √365⁄14
𝜎𝑘𝑗 =
𝛼

where

𝛼 = Φ−1 (0.99)

𝑅𝑊𝑘 vol-tenor 𝑗 is the option expiry index, which should use the same tenor-buckets as the
interest-rate delta risk: 2W, 1M, 3M, 6M, 1Y, 2Y, 3Y, 5Y, 10Y, 15Y, 20Y, and 30Y. Here
Φ−1 (0.99) is the 99th percentile of the standard normal distribution.
16. Commodity Index Volatility Risk Weights: For commodity index volatilities, the risk weight
to use is that of the Indexes bucket.
17. FX Delta Sensitivity Risk Weight: For FX vega – which depends upon a pair of currencies –
the risk weight to use here is the common risk weight for the FX delta sensitivity given
explicitly in the corresponding section.
18. Instrument Level Vega Risk Expression: The vega risk for each instrument 𝑖 to the risk factor
𝑘 is estimated using the expression

𝜕𝑉𝑖𝑗
𝑉𝑅𝑖𝑘 = 𝐻𝑉𝑅𝑐 ∑ 𝜎𝑗𝑘
𝜕𝜎
𝑗

19. Product Specific Vega Value applicable: Here 𝜎𝑗𝑘 is the volatility defined over the last five
clause points.
𝜕𝑉𝑖𝑗
20. Instrument-Specific Price-Vega Sensitivity: is the sensitivity of the price of the
𝜕𝜎

instrument 𝑖 with respect to the implied at-the-money volatility – vega – as defined later, but
must match the definition above.

178
21. Historical Volatility Risk Class Correction: 𝐻𝑉𝑅𝑐 is the historical volatility ratio for the risk
class concerned 𝑐 as set out in the section on Equity Risk, which corrects for the inaccuracy
in the volatility estimate 𝜎𝑗𝑘 . The historical volatility ratio for the interest rate and the credit
risk classes is fixed at 1.0.
22. 5Y Interest Rate Swap Vega: For example, the 5Y interest rate vega is the sum of all the vol-
weighted interest rate caplet and swaption vegas which expire in 5 years’ time; the USDJPY
FX vega is the sum of all the vol-weighted USD/JPY FX vegas.
23. Gross Vega for Inflation Products: For inflation, the inflation vega is the sum of all vol-
weighted inflation swaption vegas in the particular currency.
24. Instrument Net Vega Risk Exposure: Find a net vega risk exposure 𝑉𝑅𝑘 across instruments 𝑖
to each risk factor 𝑘 – which are defined in the later sections – as well as the vega
concentration risk factor.
25. Portfolio/Factor IR Vega Risk: For interest-rate vega risk, these are given by the formulas

𝑉𝑅𝑘 = 𝑉𝑅𝑊 (∑ 𝑉𝑅𝑖𝑘 ) 𝑉𝐶𝑅𝑘

𝑖

where

|∑𝑖,𝑘 𝑉𝑅𝑖𝑘 |
𝑉𝐶𝑅𝑏 = max (1, √ )
𝑉𝑇𝑏

where 𝑏 is the bucket which contains the risk factor 𝑘.

26. Portfolio/Factor Credit Risk Vega: For credit spread vega risk, the corresponding formulas
are

𝑉𝑅𝑘 = 𝑉𝑅𝑊 (∑ 𝑉𝑅𝑖𝑘 ) 𝑉𝐶𝑅𝑘

𝑖

179
 where

|∑𝑖,𝑗 𝑉𝑅𝑖𝑗 |
𝑉𝐶𝑅𝑘 = max (1, √ )
𝑉𝑇𝑏

where the sum 𝑗 is taken over tenors of same issuer/seniority curve as the risk factor 𝑘,
irrespective of the tenor or the payment currency.
27. Equity/FX/Commodity Vega Risk: For equity, FX, and commodity vega risks, the
corresponding formulas are

𝑉𝑅𝑘 = 𝑉𝑅𝑊 (∑ 𝑉𝑅𝑖𝑘 ) 𝑉𝐶𝑅𝑘

𝑖

where

|∑𝑘 𝑉𝑅𝑖𝑘 |
𝑉𝐶𝑅𝑘 = max (1, √ )
𝑉𝑇𝑏

28. Vega Weights for the Risk Class: Here 𝑉𝑅𝑊 is the vega risk weight for the risk class
concerned as set out in the corresponding sections, and 𝑉𝑇𝑏 is the vega concentration
threshold for the bucket – or FX category 𝑏 - as given in the corresponding section.
29. Index Volatilities for Risk Classes: Note that there is a special treatment for index volatilities
in Credit Qualifying, Equity, and Commodity Risk Classes.
30. Vega Exposure across Risk Class: The vega risk exposure should then be aggregated within
each bucket. The buckets and the correlation parameters applicable to each risk class are set
out in the Sections on Risk Weights and Correlations.
31. Cross Factor Vega Margin Expression:

180
 𝑉𝑒𝑔𝑎𝑀𝑎𝑟𝑔𝑖𝑛 = √∑ 𝐾𝑏 2 + ∑ ∑ 𝛾𝑏𝑐 𝑔𝑏𝑐 𝑆𝑏 𝑆𝑐 + 𝐾𝑅𝐸𝑆𝐼𝐷𝑈𝐴𝐿
𝑏 𝑏 𝑐≠𝑏

where

𝑆𝑏 = max (min {∑ 𝑉𝑅𝑘 , 𝑉𝑏 } , −𝐾𝑏 )

𝑘=1

for all risk factors in bucket 𝑏.

32. Outer Correlation Adjustment Factors 𝑔𝑏𝑐 : The outer correlation adjustment factors 𝑔𝑏𝑐 are
identically 1.0 for all risk classes other than interest rates, and for interest rates they are
defined as

min(𝑉𝐶𝑅𝑏 , 𝑉𝐶𝑅𝑐 )
𝑔𝑏𝑐 =
max(𝑉𝐶𝑅𝑏 , 𝑉𝐶𝑅𝑐 )

33. Approach underlying Tension/Curvature Risk: The following step-by-step approach to

capture curvature risk exposure should be applied to each risk class.
34. Expression for Tension/Curvature Risk: The curvature risk exposure for each instrument 𝑖 to
risk factor 𝑘 is estimated using the expression

𝜕𝑉𝑖
𝐶𝑉𝑅𝑖𝑘 = ∑ 𝑆𝐹(𝑡𝑗𝑘 )𝜎𝑗𝑘
𝜕𝜎
𝑗

𝜕𝑉𝑖
35. Pairwise Volatility and Vega: Here 𝜎𝑗𝑘 and are the volatility and the vega defined in the
𝜕𝜎

items above.
36. Incorporating Standard Option Expiry Time: 𝑡𝑗𝑘 is the expiry time in calendar days from the
valuation date until the expiry of the standard option corresponding to this volatility and
vega.

181
37. Scaling Function Vega/Gamma Linkage: is the value of the scaling function obtained from
the linkage between vega and gamma for vanilla options:

14 𝑑𝑎𝑦𝑠
𝑆𝐹(𝑡) = 0.5 min (1, )
𝑡 𝑑𝑎𝑦𝑠

38. Scaling Function Dependence on Expiry: The scaling function is a function of expiry only, is
independent of both the vega and the vol, and is show in the table below.

2w 1m 3m 6m 12m 2y 3y 5y 10y

0.500 0.230 0.077 0.038 0.019 0.010 0.006 0.004 0.002

39. Conversion of Tenors to Days: Here the tenors are converted to calendar days using the
convention that 12m equals 365 calendar days, with pro-rata scaling for other tenors so that

365
1𝑚 = 𝑑𝑎𝑦𝑠
12

and

5𝑦 = 365 × 5 𝑑𝑎𝑦𝑠

40. Cross Factor Curvature Risk Aggregation: The curvature risk should then be aggregated
within each bucket using the following expression:

𝐾𝑏 = √∑ 𝐶𝑉𝑅𝑘 2 + ∑ ∑ 𝜌𝑘𝑙 2 𝐶𝑉𝑅𝑏𝑘 𝐶𝑉𝑅𝑏𝑙

𝑘 𝑘 𝑙≠𝑘

182
 where 𝜌𝑘𝑙 is the correlation applicable to each risk class as set out in the section on risk
weights and correlations. Note the use of 𝜌𝑘𝑙 2 rather than 𝜌𝑘𝑙 .
41. Instrument Risk Factor Curvature Exposure: The curvature risk exposure 𝐶𝑉𝑅𝑖𝑘 can then be
netted across instrument 𝑖 to each risk factor 𝑘, which are defined in the Sections on Risk
Factors and Sensitivities. Note that the same special treatment as for vega applies for indexes
in credit, equity, and commodity risk exposures.
42. The Non-Residual Curvature Standard Derivation Scaler: Margin should then be aggregated
across buckets within each risk class:

∑𝑏,𝑘 𝐶𝑉𝑅𝑏𝑘
𝜃 = min ( , 0)
∑𝑏,𝑘|𝐶𝑉𝑅𝑏𝑘 |

and

𝜆 = [Φ−1 (0.995)2 − 1](1 + 𝜃) − 𝜃

where the sums are taken over all the non-residual buckets in the risk class, and Φ−1 (0.995)
is the 99.5th percentile of the standard normal distribution.
43. Non-Residual Curvature Margin Expression: Then the non-residual curvature margin is

𝐶𝑢𝑟𝑣𝑎𝑡𝑢𝑟𝑒𝑀𝑎𝑟𝑔𝑖𝑛𝑁𝑂𝑁−𝑅𝐸𝑆𝐼𝐷𝑈𝐴𝐿 = max (∑ 𝐶𝑉𝑅𝑏𝑘 + 𝜆√∑ 𝐾𝑏 2 + ∑ ∑ 𝛾𝑏𝑐 2 𝑆𝑏 𝑆𝑐 , 0)

𝑏,𝑘 𝑏 𝑏 𝑐≠𝑏

where

𝑆𝑏 = max (min {∑ 𝑉𝑅𝑘 , 𝑉𝑏 } , −𝐾𝑏 )

𝑘=1

44. The Residual Curvature Margin Expression: Similarly the residual equivalents are defined as

183
 ∑𝑘 𝐶𝑉𝑅𝑅𝐸𝑆𝐼𝐷𝑈𝐴𝐿,𝑘
𝜃𝑅𝐸𝑆𝐼𝐷𝑈𝐴𝐿 = min ( , 0)
∑𝑘|𝐶𝑉𝑅𝑅𝐸𝑆𝐼𝐷𝑈𝐴𝐿,𝑘 |

and

𝜆𝑅𝐸𝑆𝐼𝐷𝑈𝐴𝐿 = [Φ−1 (0.995)2 − 1](1 + 𝜃𝑅𝐸𝑆𝐼𝐷𝑈𝐴𝐿 ) − 𝜃𝑅𝐸𝑆𝐼𝐷𝑈𝐴𝐿

𝐶𝑢𝑟𝑣𝑎𝑡𝑢𝑟𝑒𝑀𝑎𝑟𝑔𝑖𝑛𝑅𝐸𝑆𝐼𝐷𝑈𝐴𝐿 = max (∑ 𝐶𝑉𝑅𝑅𝐸𝑆𝐼𝐷𝑈𝐴𝐿,𝑘 + 𝜆𝑅𝐸𝑆𝐼𝐷𝑈𝐴𝐿 𝐾𝑅𝐸𝑆𝐼𝐷𝑈𝐴𝐿 , 0)

𝑘

45. Applying the Risk Factor Correlations: Here the correlation parameters 𝛾𝑏𝑐 applicable to
each risk class are set out in the Sections on Risk Weights and Correlations. Note the use of
𝛾𝑏𝑐 2 rather than 𝛾𝑏𝑐 .
46. Expression for Composite Curvature Margin: The total curvature margin is defined to be the
sum of the two curvature terms

𝐶𝑢𝑟𝑣𝑎𝑡𝑢𝑟𝑒𝑀𝑎𝑟𝑔𝑖𝑛 = 𝐶𝑢𝑟𝑣𝑎𝑡𝑢𝑟𝑒𝑀𝑎𝑟𝑔𝑖𝑛𝑅𝐸𝑆𝐼𝐷𝑈𝐴𝐿 + 𝐶𝑢𝑟𝑣𝑎𝑡𝑢𝑟𝑒𝑀𝑎𝑟𝑔𝑖𝑛𝑁𝑂𝑁−𝑅𝐸𝑆𝐼𝐷𝑈𝐴𝐿

47. Interest Rate Curvature Margin Scaler: For the interest rate risk class only, the Curvature
Margin must be multiplied by a scaler of 2.3. This provisional adjustment addresses a known
weakness in the expression that converts gamma into curvature, which will be properly
addressed in a later version of the model.
48. Base Correlation Model Credit Charge: Credit Qualifying Only – Instruments whose prices is
sensitive to the correlation between the defaults of different credits within an index or a
basket – such as CDO tranches – are subject to Base Correlation margin charges described
below. Instruments not sensitive to base correlation are not subject to Base Correlation
margin requirements.
49. Base Correlation Risk Exposure Approach: The following step-by-step approach to capture
the Base Correlation risk exposure should be applied to the Credit Qualifying Risk Class.

184
50. Base Correlation Risk Factor Sensitivity: The net sensitivity across instruments to each Base
Correlation risk factor 𝑘 is calculated, where 𝑘 is the index family such as CDX/IG.
51. Risk Weight applied to Sensitivity: Weight the net sensitivity 𝑠𝑘 to each risk factor 𝑘 by the
corresponding risk weight 𝑅𝑊𝑘 specified in the section on Credit Qualifying Risk:

𝑊𝑆𝑘 = 𝑅𝑊𝑘 𝑠𝑘

52. Aggregation of the Weighted Sensitivities: Weighted sensitivities should then be aggregated
to calculate the Base Correlation Margin as follows:

𝐵𝑎𝑠𝑒𝐶𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑖𝑜𝑛𝑀𝑎𝑟𝑔𝑖𝑛 = √∑ 𝑊𝑆𝑘 2 + ∑ ∑ 𝜌𝑘𝑙 𝑊𝑆𝑘 𝑊𝑆𝑙

𝑘 𝑘 𝑙≠𝑘

The correlation parameters are set out in the section on Credit Qualifying Risk Weights and
Correlations.

References

• International Swaps and Derivatives Association (2016): ISDA SIMM 2.0 Methodology
• International Swaps and Derivatives Association (2017): ISDA SIMM 2.1 Methodology

185
 Dynamic Initial Margin Impact on Exposure

Abstract

1. VM and IM Collateralized Positions: This chapter leverages the new framework for
collateralized exposure modeling introduced by Andersen, Pykhtin, and Sokol (2017b) to
analyze credit risk positions collateralized with both initial and variation margin. Special
attention is paid to the dynamics BCBS-IOSCO uncleared margin rules soon to be mandated
for bilateral inter-dealer trading in OTC derivatives markets.
2. Insufficiency of BCBS IOSCO Rules: While these rules set the initial margin at 99th 2-week
percentile level and aim to all but eliminate portfolio close-out risk, this chapter demonstrates
that the trade flow effects can result in exposures being reduced significantly less than
expected.
3. Efficient IM Simulation on an MC Path: The analysis is supplemented with several practical
schemes for estimating IM on a simulation path, and for improving the speed and the stability
of the exposure simulation.
4. Handling Trade Flow Exposure Spikes: This chapter also briefly discusses potential ways to
adjust the margin framework to more effectively deal with exposures arising from trade flow
events.

Introduction

186
1. Collateralization based on Variation Margin: Collateralization has long been a way of
mitigating counterparty risk in OTC bilateral trading. The most common collateral
mechanism is variation margin (VM) which aims to keep the threshold gap between
portfolio value and posted collateral below a certain, possibly stochastic, threshold. While it
is
2. Imperfect Collateralization under VM Schemes: Even when the thresholds for the VM are set
to zero, however, there remains residual exposure to the counterparty default resulting from a
sequence of contractual and operational time lags, starting from the last snapshot of the
market for which the counterparty would post in full the required VM to the termination date
after the counterparty’s default. The various collateral mechanisms, including the precise
definition of the variation margin thresholds, are typically captured in the ISDA Credit
Support Annex (CSA) – the portfolio level legal agreement that supplements the ISDA
Master Agreement.
3. MPoR - Margin Period of Risk: The aggregation of these lags results in a time period called
the Margin Period of Risk (MPoR) during which the gap between the portfolio value and
collateral can widen. The length of the MPoR is a critical input to any model of collateral
exposure.
4. IM Supplementing the VM Collateral: Posting of initial margin (IM) to supplement VM
provides the dealers with a mechanism to reduce the residual exposure resulting from market
risk over MPoR. While it is often believed that the IM is posted strictly in addition to the
VM, many CSAs intermingle the two types of collateral by letting IM affect the threshold
computation of VM.
5. Genesis and Structure of IM: Historically, IM in bilateral trading has been mostly reserved
for dealer counterparties deemed as high-risk – e.g., hedge funds – and typically done as a
trade level calculation, established in term sheets at the transaction time of each trade. This
type of IM posting is normally deterministic and either stays fixed over the life of a trade or
amortizes down according to a pre-specified schedule.
6. Ne Basel Rules for IM: In the inter-dealer bilateral OTC world, changes to the long-standing
VM and IM collateral practices are now imminent. BCBS and IOSCO proposed (Basel

187
 Committee on Banking Supervision (2013)) and later finalized (Basel Committee on Banking
Supervision (2015)) new uncleared margin rules for bilateral trading.
7. Key Features of the UMR: Under UMR, VM thresholds are forced to zero, and IM must be
posted bilaterally into segregated accounts at the netting set level, by either using an internal
model or by a lookup in a standardized schedule.
8. IM as a Horizon-Specific VaR: If an internal model is used, IM must be calculated as a
netting set Value-at-Risk (VaR) for a 99% confidence level. The horizon used in this
calculation equals 9 + 𝑎 business days, where 𝑎 is the re-margining period - 1 business day
under US rules.
9. No Cross-Asset Class Diversification: In these calculations, diversification across distinct
asset classes is not recognized, and calibration of the IM internal model for each asset class
must include a period of stress for that asset class. To reduce the potential for margin
disputes and to increase the overall market transparency, ISDA has proposed a standardized
sensitivity-based IM calculator known as SIMM (Standard Initial Margin Model)
(International Swaps and Derivatives Association (2016)). As a practical matter it is expected
that virtually all dealers will use SIMM for their day-to-day IM calculations.
10. Dynamic Nature of Initial Margin: Under UMR required levels of IM continuously change as
trade cash flows are paid, new trades are booked, or markets move, and dealers regularly
need to call for more IM or to return excess IM. This dynamic aspect of IM requirements
makes the modeling of the future exposures a challenge.
11. Modeling under Dynamics IM and VM: This chapter discusses modeling credit exposure in
the presence of dynamic IM and questions the conventional wisdom that IM essentially
eliminates counterparty risk. Leaning on the recent results from Andersen, Pykhtin, and
Sokol (2017a), it starts by formulating a general model of exposure in the presence of VM
and/or IM.
12. Simple Case - No Trade Flows: The resulting framework is first applied to the simple case
where no trade flows take place within the MPoR. For processes with Gaussian increments –
e.g., an Ito process – a limiting scale factor that converts the IM free expected exposure (EE)
to IM-protected EE is derived, for sufficiently small MPoR.

188
13. IM vs no IM EE Ratio: The universal value depends only on the IM confidence level and the
ratio of the IM horizon to the MPoR; it equals 0.85% at the BCBS-IOSCO confidence level
of 99%, provided the IM horizon equals the MPoR. While conceptually the IM and the
MPoR horizons are identical, a prudent MPoR for internal calculations may differ from the
regulatory minimum IM horizon.
14. No-Trade-Flow Exposure Reduction: While some deviations from this universal limit due to
a non-infinitesimal MPoR are to be expected, the reduction of EE by about 2 orders of
magnitude is, as will be demonstrated below, generally about right when no trade flows are
present within the MPoR.
15. Exposure Spikes from Trade Flows: For those periods for which trade flows do take place
within the MPoR, however, any trade payment flowing away from the dealer will result in a
spike in the EE profile. Without IM these spikes can make a fairly significant contribution to
the Credit Valuation Adjustment (CVA) – say, 20% of an interest rate swap’s total CVA
may originate with spikes – but the CVA would still mostly be determined by the EE level
between the spikes.
16. Exposure Spikes vs. Dynamic IM: This chapter shows that while IM is effective in
suppressing the EE between spikes, it will often fail to significantly suppress the spikes
themselves. As a result, the relative contribution of the spike to CVA is greatly increased in
the presence of IM – e.g., for a single interest rate swap, the spike’s contribution to the CVA
can be well about 90% for a position with IM.
17. Corresponding Impact on the CVA: Accounting for the spikes, the IM reduces the CVA by
much less than two orders of magnitude one might expect, with the reduction for the interest
rate swaps often being less than a factor of 10.
18. Estimating the Path-wise IM: The final part of this chapter discusses the practical approaches
to calculating the EE profiles in the presence of IM. The first step in this calculation is the
estimation of IM on simulation paths, which can be done by parametric regression or by
kernel regression.
19. IM Covering Few Netting Trades: When IM covers an insignificant number of trades in the
netting set, IM calculated on the path can be subtracted from the no-IM exposure realized on
that path to generate EE profiles.

189
20. IM covering most Netting Trades: However, when most trades of the netting set are covered
by the IM, this approach can be problematic because of excessive simulation noise and other
errors. An alternative approach that dampens the noise is proposed, and is generally more
accurate.
21. Suggested Alterations to the Exposure Rules: This chapter concludes by summarizing the
results and briefly discussing the possible modifications to trade and collateral
documentation that would make IM more effective in reducing residual counterparty risk.

Exposure in the Presence of IM and VM

1. VM/IM over single Netting Set: Consider a dealer D that has a portfolio of OTC derivatives
contracts traded with a counterparty C. Suppose for simplicity that the entire derivatives
portfolio is covered by a single netting agreement, which is supported by a margin agreement
that includes VM and may include IM on a subset of the portfolio.
2. Exposure of Client to Dealer: Quiet generally the exposure of D to the default of client C
measured at time 𝑡 – assumed to be the early termination time after C’s default – is given by

𝐸(𝑡) = [𝑉(𝑡) − 𝑉𝑀(𝑡) + 𝑈(𝑡) − 𝐼𝑀 (𝑡)]+

where 𝑉(𝑡) is the time 𝑡 portfolio value from D’s perspective; 𝑉𝑀(𝑡) is the VM available to
D at time 𝑡; 𝑈(𝑡) is the value of the trade flows scheduled to be paid by both D (negative)
and C (positive) up to time 𝑡, yet unpaid as of time 𝑡; 𝐼𝑀(𝑡) is the value of IM available to D
at time 𝑡.
3. Sign of VM and IM: Notice that VM can be positive – C posts VM – or negative – D posts
VM – from D’s perspective. On the other hand, IM is always positive as IM for both
counterparties is kept in segregated accounts, whereby IM posted by D does not contribute to
D’s exposure to the default of C.
190
4. Modeling Individual Terms in the Exposure: The above equation for 𝐸(𝑡) specifies the
exposure of D to C in a generic, model-free way. To add modeling detail, this chapter
assumes that D and C both post VM with zero-threshold and are required to post BCBS-
IOSCO compliant IM to a segregated account. The modeling of each of these terms VM, U,
and IM are dealt with turn by turn.

Modeling VM

1. Concurrent Dealer/Client VM Stoppage: The length of the 𝑀𝑃𝑜𝑅 denoted by 𝛿𝐶 defines the
last portfolio valuation date

𝑡0 = 𝑡 − 𝛿𝐶

prior to the termination date 𝑡 - after C’s default – for which C delivers VM to D. A common
assumption – denoted here as the Classical Model – assumes that D stops paying VM to C at
the exact same time C stops posting to D.
2. Expression for Classical Model VM: That is, the VM in the equation above is the VM
prescribed for the margin agreement for the portfolio valuation date

𝑡0 = 𝑡 − 𝛿𝐶

Ignoring minimum transfer amount and rounding, the prescribed VM in the Classical Model
is thus simply

𝑉𝑀𝐶𝐿𝐴𝑆𝑆𝐼𝐶𝐴𝐿 (𝑡) = 𝑉(𝑡𝑐 ) = 𝑉(𝑡 − 𝛿𝐶 )

191
3. Advanced Model Incorporating Operational Details: In the Advanced Model of Andersen,
Pykhtin, and Sokol (2017a), operational aspects and gamesmanship of margin disputes are
considered in more detail, leading to the more realistic assumption that D may continue to
post VM to C for some period of time, even after C stops posting.
4. Non-Concurrent Dealer/Client VM Stoppages: The model introduces another parameter

𝛿𝐷 ≥ 𝛿𝐶

that specifies that last portfolio valuation date

𝑡𝐷 = 𝑡 − 𝛿𝐷

for which D delivers VM to C. For the portfolio valuation dates

𝑇𝑖 ∈ [𝑡𝐶 , 𝑡𝐷 ]

D would post VM to C when the portfolio value decreases, but will receive no VM from C
when the portfolio value increases.
5. Expression for Advanced Model VM: This results in VM of

𝑚𝑖𝑛
𝑉𝑀𝐴𝐷𝑉𝐴𝑁𝐶𝐸𝐷 (𝑡) = 𝑉(𝑇𝑖 )
𝑇𝑖 ∈ [𝑡𝐶 , 𝑡𝐷 ]

This equation of course reduces to

𝑉𝑀𝐶𝐿𝐴𝑆𝑆𝐼𝐶𝐴𝐿 (𝑡) = 𝑉(𝑡𝑐 ) = 𝑉(𝑡 − 𝛿𝐶 )

when one sets

𝛿𝐷 = 𝛿𝐶

192
Modeling 𝑼

1. The Classical+ Version Trade Flow Currentness: In the most conventional version of the
Classical model – denoted here Classical+ - it is assumed that all trade flows are paid by
both C and D for the entire 𝑀𝑃𝑜𝑅 up to and including the termination date, i.e., in the time
interval [𝑡𝐶 , 𝑡] - time here is measured in discrete business days, so that [𝑢, 𝑠] is equivalent to
[𝑢 + 1 𝐵𝐷, 𝑠]. This assumption simply amounts to setting

𝑈𝐶𝐿𝐴𝑆𝑆𝐼𝐶𝐴𝐿+ (𝑡) = 0

2. Trade Flow Exposure Profile Spikes: One of the prominent features of the Classical+ model
is that the time 0 expectation of 𝐸(𝑡) – denoted 𝐸𝐸(𝑡) – will contain upward spikes
whenever there is a possibility of trade flows from D to C within the interval [𝑡𝐶 , 𝑡].
3. Absence of Client Margin Flows: These spikes appear because, by the classical model’s
assumption, C makes no margin payments during the 𝑀𝑃𝑜𝑅 and would consequently fail to
post an offsetting VM to D after D makes a trade payment to C. In the Classical model, D
will also not post VM to C in the event of trade payment from C to D, which results in a
negative jump in the exposure. However, these scenarios do not fully offset the scenarios
where D makes a trade payment because the zero floor in the exposure definitions effectively
limits the size of the downward exposure jump.
4. Sparse Fixed Time Exposure Grid: For dealers having a sparse fixed time exposure grid, the
alignment of grid nodes relative to trade flows will add numerical artifacts to genuine spikes,
causing EE exposure to appear and disappear as the calendar date moves. As a consequence,
an undesirable instability in the 𝐸𝐸 and the 𝐶𝑉𝐴 is introduced.

193
5. Classical- Version – No Trade Flows: An easy way to eliminate exposure spikes is to assume
that neither C nor D make any trade payment inside the 𝑀𝑃𝑜𝑅. The resulting model – here
denoted Classical- - consequently assumes that

𝑈𝐶𝐿𝐴𝑆𝑆𝐼𝐶𝐴𝐿− (𝑡) = 𝑇𝐹𝑁𝐸𝑇 (𝑡; (𝑡𝐶 , 𝑡])

where 𝑇𝐹𝑁𝐸𝑇 (𝑡; (𝑠, 𝑢]) denotes the time 𝑡 of all net trade flows payable in the interval (𝑠, 𝑢].
6. Implicit Simplifications in Classical- and Classical+: It should be evident that neither the
Classical- nor the Classical+ assumptions on trade flows are entirely realistic; in the
beginning of the 𝑀𝑃𝑜𝑅 both C and D are likely to make trade payments, while at the end of
the 𝑀𝑃𝑜𝑅 neither C no D are likely making any trade payments.
7. Last Dealer/Client Trade Flows: To capture this behavior, Andersen, Pykhtin, and Sokol
(2017a) add two more parameters in the model 𝛿𝐶 ′ and

𝛿𝐷 ′ ≤ 𝛿𝐶

that specify the last dates

𝑡𝐶 ′ = 𝑡 − 𝛿𝐶 ′
and

𝑡𝐷 ′ = 𝑡 − 𝛿𝐷 ′

for D at which trade payments are made prior to closeout at 𝑡.

8. Advanced Model Trade Flow Expansion: This results in unpaid trade flow terms of

𝑈𝐴𝐷𝑉𝐴𝑁𝐶𝐸𝐷 (𝑡) = 𝑇𝐹𝐶→𝐷 (𝑡; (𝑡𝐶 ′ , 𝑡𝐷 ′ ]) + 𝑇𝐹𝑁𝐸𝑇 (𝑡; (𝑡𝐷 ′ , 𝑡])

where an arrow indicates the direction of the trade flows and 𝐶 → 𝐷 (𝐷 → 𝐶) trade flows
have positive (negative) sign.

194
9. Advanced Model Exposure Profile Structure: The EE profiles obtained with the Advanced
Model contain spikes that are typically narrower and have a more complex structure than
spikes under the Classical+ model. Rather than being unwelcome noise, it is argued in
Andersen, Pykhtin, and Sokol (2017a) that spikes in EE profiles are important features that
represent actual risk.
10. Approximating the Trade Flow Spikes: To eliminate any numerical instability associated
with the spikes an approximation was proposed in Andersen, Pykhtin, and Sokol (2017a) for
calculation of the EE on a daily time grid without daily re-evaluation of the portfolio.

Modeling IM

1. Expression for Netting Set IM: Following the BCBS-IOSCO restrictions on diversification,
the IM is defined for the netting set as a sum of the IM’s over 𝐾 asset classes as

𝐼𝑀(𝑡) = ∑ 𝐼𝑀𝑘 (𝑡)

𝑘=1

Current netting rules have

𝐾=4

𝑘 can be one of Rates/FX, Credit, Equity, and Commodity.

2. Netting Set Granularity IM Value: Let 𝑉𝑘 (𝑡) denote the value at time 𝑡 of all trades in the
netting set to asset class 𝑘 and are subject to IM requirements. Note that only trades executed
after UMR go-live will be covered by the IM. The case where only a subset of the netting set
is covered by the IM will therefore be common in the near future.

195
3. Horizon/Confidence Based IM Definition: For an asset class 𝑘 we define IM as the quantile
at the confidence level 𝑞 of the clean portfolio value increment over BCBS-IOSCO IM
horizon 𝛿𝐼𝑀 - which may or may not coincide with 𝛿𝐶 - conditional on all the information
available at 𝑡𝐶 .
4. Mathematical Expression for the IM: That is

𝐼𝑀𝑘 (𝑡) = 𝑄𝑞 [𝑉𝑘 (𝑡𝐶 + 𝛿𝐼𝑀 ) + 𝑇𝐹𝑁𝐸𝑇,𝑘 (𝑡𝐶 + 𝛿𝐼𝑀 ; (𝑡𝐶 , 𝑡𝐶 + 𝛿𝐼𝑀 ]) − 𝑉𝑘 (𝑡𝐶 )| ℱ𝑡𝐶 ]

where 𝑄𝑞 (∙ | ℱ𝑠 ) denotes the quantile of confidence level 𝑞 conditional on information

available at time 𝑠. Note that the above expression assumes that C stops posting IM at the
same time it stops posting VM; hence IM is calculated as of

𝑡𝐶 = 𝑡 − 𝛿𝐶

Summary and Calibration

1. Classical+/Classical-/Advanced Model Exposures: To summarize, three different models

have been outlined in the generic exposure calculation

𝐸(𝑡) = [𝑉(𝑡) − 𝑉𝑀(𝑡) + 𝑈(𝑡) − 𝐼𝑀(𝑡)]+

for the Classical-, Classical+, and Advanced Models. Collecting results, one has

𝐸𝐶𝐿𝐴𝑆𝑆𝐼𝐶𝐴𝐿+ (𝑡) = [𝑉(𝑡) − 𝑉(𝑡 − 𝛿𝐶 ) − 𝐼𝑀(𝑡)]+

𝐸𝐶𝐿𝐴𝑆𝑆𝐼𝐶𝐴𝐿− (𝑡) = [𝑉(𝑡) − 𝑉(𝑡 − 𝛿𝐶 ) + 𝑇𝐹𝑁𝐸𝑇 (𝑡; (𝑡𝐶 , 𝑡]) − 𝐼𝑀(𝑡)]+

196
 𝑚𝑖𝑛
𝐸𝐴𝐷𝑉𝐴𝑁𝐶𝐸𝐷 (𝑡) = [𝑉(𝑡) − 𝑉(𝑇𝑖 ) + 𝑇𝐹𝐶→𝐷 (𝑡; (𝑡𝐶 ′ , 𝑡𝐷 ′ ]) + 𝑇𝐹𝑁𝐸𝑇 (𝑡; (𝑡𝐷 ′ , 𝑡])
𝑇𝑖 ∈ [𝑡𝐶 , 𝑡𝐷 ]
+
− 𝐼𝑀(𝑡)]

where for all the three models IM is computed as

𝐼𝑀𝑘 (𝑡) = 𝑄𝑞 [𝑉𝑘 (𝑡𝐶 + 𝛿𝐼𝑀 ) + 𝑇𝐹𝑁𝐸𝑇,𝑘 (𝑡𝐶 + 𝛿𝐼𝑀 ; (𝑡𝐶 , 𝑡𝐶 + 𝛿𝐼𝑀 ]) − 𝑉𝑘 (𝑡𝐶 )| ℱ𝑡𝐶 ]

2. IMA/CSA Dealer/Client Suspension Rights: In practice, the calibration of the time

parameters of the Advanced Model should be informed by the dealers’ legal rights and its
aggressiveness in pursuing them. For the former, Andersen, Pykhtin, and Sokol (2017a)
contain a full discussion of ISDA Master Agreements. But note that once a notice of
Potential Event of Default (PED) has been served, the suspension rights of ISDA Master
Agreement 2 (a) (iii) and its accompanying Credit Support Annex Paragraph 4 (a) allow a
bank to suspend all trade-related and collateral-related payments to its counterparty until the
PED has been cured.
3. Risks of Enforcing the Rights: The extent to which the suspension rights are actually
exercised in practice, however, depends on the dealer and its operating model, as well as the
specifics of each of the PED. One particular danger in the aggressive and immediate
enforcement of the suspension rights is that the original PED might be ruled unjustified.
Should this happen, the dealer can inadvertently commit a breach of contract, which,
especially in the presence of the cross-default provisions, can have serious consequences for
the dealer.
4. Calibrating the MPoR Lag Parameter: With the four time parameters, the Advanced Model
allows for a greater flexibility in modeling its risk tolerance and procedures for exercising its
suspension rights. A dealer may, in fact, calibrate these parameters differently for different
counter-parties to reflect its risk management practices towards counterparties of a given
type. One may make the time parameters stochastic, e.g., by making the legs a function of the

197
 exposure magnitude. This way one can, say, model the fact that a dealer might tighten its
operational controls when the exposures are high.
5. Prototypical Aggressive vs. Conservative Timelines: Additional discussions can be found in
Andersen, Pykhtin, and Sokol (2017a), which also provides some prototypical parameter
settings; the Aggressive Calibration (D can always sniff out financial distress in its clients
and is swift and aggressive in enforcing its legal rights) and the Conservative Calibration (D
is deliberate and cautious in enforcing its rights, and acknowledges potential for operational
errors and for rapid, unpredictable deterioration in client credit). For the numerical results in
this chapter, the values of the time parameters are mostly set in between the Aggressive and
the Conservative.

The Impact of IM: No Trade Flows within the MPoR

1. No Trade Flow IM EE Impact: This section examines the impact of IM on EE when there are
no trade flows within the MPoR. For simplicity the Classical Model is considered, and the
entire netting set is assumed to be covered by the IM and comprised of all trades belonging to
the same asset class. Results for the Advanced Models are similar, as shown in the later
sections.
2. Estimating the IM Efficiency Ratio: In the absence of trade flows on (𝑡𝐶 , 𝑡]

𝐸𝐶𝐿𝐴𝑆𝑆𝐼𝐶𝐴𝐿+ (𝑡) = [𝑉(𝑡) − 𝑉(𝑡 − 𝛿𝐶 ) − 𝐼𝑀(𝑡)]+

𝐸𝐶𝐿𝐴𝑆𝑆𝐼𝐶𝐴𝐿− (𝑡) = [𝑉(𝑡) − 𝑉(𝑡 − 𝛿𝐶 ) + 𝑇𝐹𝑁𝐸𝑇 (𝑡; (𝑡𝐶 , 𝑡]) − 𝐼𝑀(𝑡)]+

show that the Classical Model computes the expected exposure as

198
 +
𝐸𝐸(𝑡) = 𝔼 [{𝑉(𝑡) − 𝑉(𝑡𝐶 ) − 𝑄𝑞 [𝑉𝑅 (𝑡𝐶 + 𝛿𝐼𝑀 )| ℱ𝑡𝐶 ]} ]

𝑡𝐶 = 𝑡 − 𝛿𝐶

where 𝔼[∙] is the expectation operator. In the absence of IM this expression would be

𝐸𝐸0 (𝑡) = 𝔼[{𝑉(𝑡) − 𝑉(𝑡𝐶 )}+ ]

One of the objectives is to establish meaningful estimates of the IM Efficiency Ratio

𝐸𝐸(𝑡)
𝜆(𝑡) ≜
𝐸𝐸0 (𝑡)

Local Gaussian Approximation

1. Portfolio Value following Ito Process: Suppose that the portfolio value 𝑉(𝑡) follows an Ito
process:

Δ𝑉(𝑡) = 𝜇(𝑡)Δ𝑡 + 𝑠 𝑇 (𝑡)Δ𝑊(𝑡)

where 𝑊(𝑡) is a vector of independent Brownian motions, and 𝜇(𝑡) and 𝑠 𝑇 (𝑡) are well-
behaved processes – with 𝑠 𝑇 (𝑡) being vector-valued – adapted to 𝑊(𝑡). Notice that both 𝜇
and 𝑠 𝑇 may depend on the evolution of multiple risk factors prior to time 𝑡. For convenience
denote

𝜎(𝑡) = |𝑠 𝑇 (𝑡)|

199
2. Portfolio Increment as a Gaussian: Then, for a sufficiently small horizon 𝛿, the increment of
the portfolio value over [𝑡𝐶 , 𝑡𝐶 + 𝛿] conditional on ℱ𝑡𝐶 is well-approximated by a Gaussian
𝜎(𝑡𝐶 )
distribution with a mean 𝜇(𝑡𝐶 )𝛿 and a standard deviation .
√𝛿

3. Approximation for the Expected Exposure: Assuming 𝜎(𝑡𝐶 ) > 0 the drift term may be
ignored for small 𝛿. Under the Gaussian approximation above it is then straightforward to
approximate the expectation in

+
𝐸𝐸(𝑡) = 𝔼 [{𝑉(𝑡) − 𝑉(𝑡𝐶 ) − 𝑄𝑞 [𝑉𝑅 (𝑡𝐶 + 𝛿𝐼𝑀 )| ℱ𝑡𝐶 ]} ]

𝑡𝐶 = 𝑡 − 𝛿𝐶

in the closed form

𝐸𝐸(𝑡) ≈ 𝔼[𝜎(𝑡𝐶 )]√𝛿𝐶 [𝜙(𝑧(𝑞)) − 𝑧(𝑞)Φ(−𝑧(𝑞))]

𝛿𝐼𝑀 −1
𝑧(𝑞) ≜ √ Φ (𝑞)
𝛿𝐶

where 𝜙 and Φ are the standard Gaussian PDF and CDF, respectively.
4. Expression for IM Efficiency Ratio: Similarly

𝐸𝐸0 (𝑡) ≈ 𝔼[𝜎(𝑡𝐶 )]√𝛿𝐶 𝜙(0)

so that 𝜆 in

𝐸𝐸(𝑡)
𝜆(𝑡) ≜
𝐸𝐸0 (𝑡)

200
 is approximated by

𝜙(𝑧(𝑞)) − 𝑧(𝑞)Φ(−𝑧(𝑞))
𝜆(𝑡) ≈
𝜙(0)

5. 𝜆 Dependence: 𝑞/𝑀𝑃𝑜𝑅 Ratio: Interestingly, the multiplier 𝜆(𝑡) above is independent of 𝑡

and depends only on two quantities – the confidence level 𝑞 used for specifying the IM, and
the ratio of the IM horizon to the 𝑀𝑃𝑜𝑅. It is emphasized that the above expression for 𝜆(𝑡)
was derived under weak conditions; relying on the local normality assumptions for portfolio
value increments. Otherwise, the ratio is model-free; no further assumptions are made on the
distribution of the portfolio value or on the dependence of the local volatility on risk factors.
6. 𝜆 Exact under Brownian Motion: The special case for which the above expression for 𝜆
becomes exact is when the portfolio process follows a Brownian motion, a case discussed in
Gregory (2015). The point here is that the 𝜆(𝑡) above constitutes a small- 𝛿 limit, and a
useful approximation, for a much broader class of processes. In fact, it can be extended to
jump-diffusion processes, provided that the portfolio jumps are approximately Gaussian.
7. 𝜆 for the BCBS-IOSCO Parameters: Andersen, Pykhtin, and Sokol (2017b) illustrate the
value of 𝜆(𝑡) for the case

𝛿𝐼𝑀 = 𝛿𝐶

graphed as a function of 𝑞. For the value of the confidence level

𝑞 = 99%

specified by the BCBS-IOSCO framework

𝐸𝐸(𝑡)
𝜆(𝑡) ≜
𝐸𝐸0 (𝑡)

201
 results in a value of

𝜆 = 0.85%

i.e., IM is anticipated to reduce the expected exposure by a factor of 117.

Numerical Tests

1. Portfolio following Geometric Brownian Motion: Recall that the approximation

𝐸𝐸(𝑡)
𝜆(𝑡) ≜
𝐸𝐸0 (𝑡)

hinges on the 𝑀𝑃𝑜𝑅 being small, but it is ex-ante unclear if, say, the commonly used value
of 10 𝐵𝐷 is small enough. To investigate the potential magnitude of errors introduced by the
non-infinitesimal 𝑀𝑃𝑜𝑅, assume now – in a slight abuse of the notation – that 𝑉(𝑡) follows a
geometric Brownian motion with a constant volatility 𝜎 such that

∆𝑉(𝑡)
≅ 𝒪(∆𝑡) + 𝜎∆𝑊(𝑡)
𝑉(𝑡)

where 𝑊(𝑡) is one-dimensional Brownian motion.

2. Left/Right Portfolio Value Behavior: Compared to a Gaussian distribution, the distribution of
𝑉(𝑡) over a finite time interval is skewed left for

𝑉(0) < 0

202
 and right for

𝑉(0) > 0

with 𝑉 never crossing the origin. While this specification may seem restrictive, its only
purpose is to test the accuracy of the local Gaussian approximation.
3. Lognormal Portfolio Dynamics 𝜆 Estimate: Assuming for simplicity that

𝛿𝐼𝑀 = 𝛿𝐶

applying the calculations of the previous section to the lognormal setup, and neglecting terms
of order 𝛿 or higher, one gets

−1 (𝑞)
1 − Φ(Φ−1 (𝑞) − 𝜓𝜎√𝛿𝐶 ) − (1 − 𝑞)𝑒 𝜓𝜎√𝛿𝐶 Φ
𝜆(𝑡) = 𝜓
𝜎√𝛿
2Φ ( 2 𝐶 ) − 1

𝜓 ≜ 𝑠𝑖𝑔𝑛(𝑉(0))

4. 𝜆 Dependence for Lognormal Portfolios: While the multiplier in

𝜙(𝑧(𝑞)) − 𝑧(𝑞)Φ(−𝑧(𝑞))
𝜆(𝑡) ≈
𝜙(0)

depends only on the confidence level 𝑞 when

𝛿𝐼𝑀 = 𝛿𝐶

the multiplier in

203
 −1 (𝑞)
1 − Φ(Φ−1 (𝑞) − 𝜓𝜎√𝛿𝐶 ) − (1 − 𝑞)𝑒 𝜓𝜎√𝛿𝐶 Φ
𝜆(𝑡) = 𝜓
𝜎√𝛿
2Φ ( 2 𝐶 ) − 1

𝜓 ≜ 𝑠𝑖𝑔𝑛(𝑉(0))

additionally depends on the product 𝜎√𝛿𝐶 and on the sign of the portfolio exposure; the limit
for 𝛿𝐶 - or 𝜎 - approaching zero can be verified to be equal to

𝜙(𝑧(𝑞)) − 𝑧(𝑞)Φ(−𝑧(𝑞))
𝜆(𝑡) ≈
𝜙(0)

Andersen, Pykhtin, and Sokol (2017b) compare

𝜙(𝑧(𝑞)) − 𝑧(𝑞)Φ(−𝑧(𝑞))
𝜆(𝑡) ≈
𝜙(0)

to

−1 (𝑞)
1 − Φ(Φ−1 (𝑞) − 𝜓𝜎√𝛿𝐶 ) − (1 − 𝑞)𝑒 𝜓𝜎√𝛿𝐶 Φ
𝜆(𝑡) = 𝜓
𝜎√𝛿
2Φ ( 2 𝐶 ) − 1

𝜓 ≜ 𝑠𝑖𝑔𝑛(𝑉(0))

for various levels of 𝜎.

5. Long/Short Portfolio Volatility Impact: As expected, Andersen, Pykhtin, and Sokol (2017b)
show that the deviation of the local lognormal multiplier from the Gaussian one increases
with the volatility 𝜎 - more precisely, with the product 𝜎√𝛿𝐶 . Furthermore, the multiplier for
the case

204
 𝜓=1

is always greater – and the multiplier for the case

𝜓 = −1

is always lesser – than the limit case of 0.85% - an easily seen consequence of the fact that
the relevant distribution tail for the cases

𝜓=1

is thicker – and

𝜓 = −1

is thinner – than the tail of the Gaussian distribution, as discussed earlier.

6. No Trade Flow 𝜆 Estimate: Yet, overall the results illustrate that, for the case of the 10 𝐵𝐷
horizon, the local Gaussian approximation produces reasonable values of the scaling
multiplier at most levels of relative volatility. Certainly, the results support the idea that
introducing IM at the level of

𝑞 = 99%

should result in about two orders of magnitude reduction in the EE, when no trade flows
occur within the 𝑀𝑃𝑜𝑅.

The Impact of IM: Trade Flows within the MPoR

205
1. Impact of Trade Flows – Introduction: This section considers the efficacy of the IM in the
more complicated case where trade flows are scheduled to take place inside the 𝑀𝑃𝑜𝑅. In
practice, many portfolios produce trade flows every business day, so this case is of
considerable relevance. The introduction of cash flows will make the computation of EE both
more complex and more model-dependent, so additional care is needed here to distinguish
between the Classical+, the Classical- and the Advanced Models.
2. Trade Flows in the Classical- Model: Unless one operates under the assumption of the
Classical- model assumptions of no trade flows paid by either C or D within the 𝑀𝑃𝑜𝑅, any
possibility of D making a trade payment to C in the future results in a spike in the EE profile.
3. Origin of Trade Flow Spikes: These spikes will appear because the portfolio value will jump
following D’s payment, but C would fail to post or return the VM associated with the jump.
By way of example, suppose there is a trade payment due at future time 𝑢 where D is a net
payer.
4. Range of Trade Flow Spikes: A spike in the EE profile originating with this cash flow will
appear for a range of termination flows 𝑡 such that
a. 𝑢 lies within the 𝑀𝑃𝑜𝑅 – mathematically

𝑡 ∈ [𝑢, 𝑢 + 𝛿𝐶 )

b. D would actually make the payment.

5. Trade Flows in the Classical+ Model: The assumption in the Classical+ model is that D – as
well as C – would make contractual trade payments for the entire 𝑀𝑃𝑜𝑅 so a spike of width
𝛿𝐶 will appear in 𝐸𝐸(𝑡) for the range

𝑡 ∈ [𝑢, 𝑢 + 𝛿𝐶 )

6. Trade Flows in the Advanced Model: Here it is assumed that D would make the contractual
trade payments from the beginning 𝑡 − 𝛿𝐶 of the 𝑀𝑃𝑜𝑅 to the time 𝑡 − 𝑡𝐷 ′ so a spike of
width 𝛿𝐶 − 𝛿𝐷 ′ would appear in 𝐸𝐸(𝑡) for the range
206
 𝑡 ∈ [𝑢 + 𝛿𝐷 ′ , 𝑢 + 𝛿𝐶 )

7. Advanced vs. Classical Spike Width: While spikes produced by the Advanced model would
always be narrower than those produced by the Classical+ model, the former is often taller.
In particular, the Advanced Model, unlike the Classical+ Model, contains a range of trade
payment times

𝑡 ∈ [𝑡 − 𝛿𝐷 ′ , 𝑡 − 𝛿𝐶 )

where D would make a contractual trade payment but C would not.

8. Advanced vs Classical Spike Peak: This creates a narrow peak for the range

𝑡 ∈ [𝑢 + 𝛿𝐷 ′ , 𝑢 + 𝛿𝐶 )

at the left edge of the spike. This peak can be very high when C’s and D’s payments do not
net – e.g., because of payments in different currencies – and makes an extra contribution to
the EE that is not present in the Classical+ model.
9. IM Impact on the Margin Exposure: IM is now introduced to the exposure computation.
More specifically assume that the netting set is composed of trades belonging to a single
asset class, and in addition to the VM, is fully covered by dynamic IM as defined by UMR in

𝐼𝑀𝑘 (𝑡) = 𝑄𝑞 [𝑉𝑘 (𝑡𝐶 + 𝛿𝐼𝑀 ) + 𝑇𝐹𝑁𝐸𝑇,𝑘 (𝑡𝐶 + 𝛿𝐼𝑀 ; (𝑡𝐶 , 𝑡𝐶 + 𝛿𝐼𝑀 ]) − 𝑉𝑘 (𝑡𝐶 )| ℱ𝑡𝐶 ]

10. IM Impact on Exposure Spikes: As just discovered, in the areas between the EE spikes, IM
will reduce EE by a factor given approximately by

𝜙(𝑧(𝑞)) − 𝑧(𝑞)Φ(−𝑧(𝑞))
𝜆(𝑡) ≈
𝜙(0)

207
 The question remains on how the exposure spikes will be affected by IM. The answer to that
question is: It depends.
11. Payment Count Decrease with Time: As the simulation time passes, the portfolio will closer
to maturity, and fewer and fewer trade payments will remain. Because of this amortization
effect, the width of the distribution of the portfolio value increments on each path becomes
smaller, and being the VaR of increments, the IM is reduced.
12. VM vs Trade Spike Ratio: On the other hand, trade payments do not generally become
smaller as the trade approaches its maturity, and in fact can often become larger due to risk
factor diffusion. The effectiveness of the IM in reducing the EE spikes depends on the size of
the trade payments relative to the portfolio value increment distribution and can therefore in
many cases decline towards the maturity of the portfolio.

Expected Exposure – Numerical Example #1

1. Fix-Float Two-Way CSA: As an example, consider a 2Y interest rate swap under a two-way
CSA with zero threshold daily VM, but without IM. Assume that D pays a fixed rate of 2%
semi-annually, and receives a floating rate quarterly. Setting the initial interest rate curve at
2% - quarterly compounding – with 50% lognormal volatility, Andersen, Pykhtin, and Sokol
(2017b) illustrate the exposure profiles for the Classical+, the Classical-, and the Advanced
models.
2. Spikes in both Classical+/Advanced: As expected both the Classical+ and the Advanced
models produce spikes around the quarterly dates when the payment takes place. The nature
of the spikes depends upon whether a fixed payment is made or not.
3. Fixed Payment Time Points - Twice a Year: As C pays quarterly on its semi-annual fixed
payments, D will pay on average twice as much as C, resulting in high upward spikes in the
exposure profile. As discussed above, the width of the Classical+ model spikes is

208
 𝛿𝐶 = 10 𝐵𝐷

while the width of the Advanced model spikes is

𝛿𝐶 − 𝛿𝐷 ′ = 6 𝐵𝐷

4. Floating Only Payment Points - Twice a Year: On quarterly payment dates when no fixed
payment is due by D, C still pays a floating rate for the quarter. This results in downward
spikes in the EE profile – C makes a trade payment and defaults before D must return the
VM. The Classical+ model assumes that the trade payments are made by C (D) and the
margin payments are not made by D (C) for the entire 𝑀𝑃𝑜𝑅. Under these unrealistic
assumptions the downward spike width is

𝛿𝐶 = 10 𝐵𝐷

The Advanced model, on the other hand, assumes that C would make the trade payments
over the time interval 𝛿𝐶 − 𝛿𝐷 from which the interval 𝛿𝐶 − 𝛿𝐷 ′ over which D would pay
VM to C would be subtracted. In the aggregate, the width of the downward spikes in the
Advanced model is therefore

𝛿𝐶 − 𝛿𝐷 ′ = 2 𝐵𝐷

5. Intra-Spike Change in VM: Between the spikes, the EE profile produced by the Advanced
model is about 22% higher than the one produced by the Classical model. The difference
originates with the VM specifications – the Classical models assume that C and D stop
paying at the beginning of the 𝑀𝑃𝑜𝑅 – see

𝑉𝑀𝐶𝐿𝐴𝑆𝑆𝐼𝐶𝐴𝐿 (𝑡) = 𝑉(𝑡𝑐 ) = 𝑉(𝑡 − 𝛿𝐶 )

209
 while the Advanced model assumes that D would post VM for some time after D has stopped
doing so – see

𝑚𝑖𝑛
𝑉𝑀𝐴𝐷𝑉𝐴𝑁𝐶𝐸𝐷 (𝑡) = 𝑉(𝑇𝑖 )
𝑇𝑖 ∈ [𝑡𝐶 , 𝑡𝐷 ]

6. Impact of IM Numerical Setting: Let us consider the impact of IM on the EE profile for this
interest rate swap. All the assumptions about the 𝑀𝑃𝑜𝑅 and the CSA are retained, and the IM
horizon is set at

𝛿𝐼𝑀 = 𝛿𝐶 = 10 𝐵𝐷

Since the dynamics of the swap value are modeled with a single risk factor, the IM can be
conveniently calculated on each path exactly by stressing the risk factor to the 99% level
over the IM horizon each path. Andersen, Pykhtin, and Sokol (2017b) illustrate the EE
profile on the same scale as the no-IM result – in fact, the only part of the EE profile visible
in this scale is the upward spikes of D’s semi-annual payments of the fixed rate.
7. Numerical Comparison - Exposure between Spikes: To analyze the IM results, the exposure
profile is broken down into two categories. First, between the spikes the EE produced by the
Classical model with IM is 1.06% of the EE produced by the classical model without IM.
Since the interest rate is modeled by a lognormal process, this number is closer to the value
of 1.10% predicted by

−1 (𝑞)
1 − Φ(Φ−1 (𝑞) − 𝜓𝜎√𝛿𝐶 ) − (1 − 𝑞)𝑒 𝜓𝜎√𝛿𝐶 Φ
𝜆(𝑡) = 𝜓
𝜎√𝛿𝐶
2Φ ( )−1
2

𝜓 ≜ 𝑠𝑖𝑔𝑛(𝑉(0))

which is the lognormal case, and the value of 0.85% given by

210
 𝜙(𝑧(𝑞)) − 𝑧(𝑞)Φ(−𝑧(𝑞))
𝜆(𝑡) ≈
𝜙(0)

which is the local Gaussian approximation. Under the Advanced model, the IM scales down
EE to 1.00% of its value under VM alone, which is very similar to the Classical model case.
8. Numerical Comparison - Impact of Spikes: As discussed, the degree of suppression for the
spikes decreases with the simulation time, as the IM on all simulation paths shrinks with
time. Because the payments at the end of a period are determined at the beginning of the
period, the final payment is known exactly at time 1.75 years. So, there is in fact no IM
requirement in the period from 1.75 years to 2 years. As a result, the final spike is not
reduced at all.
9. Impact of Mismatch on Exposure: To examine the extent to which the unequal payment is
responsible for the spike dominance in the presence of IM, Andersen, Pykhtin, and Sokol
(2017b) move on to a 2Y IRS with quarterly payment on both legs, with all other model
assumptions being the same as in the previous example.
10. Numerical Analysis of Spike Reduction: For this swap, D is a net payer on the payment dates
only on approximately half of the scenarios, so the spike height will necessarily be reduced.
This is confirmed by Andersen, Pykhtin, and Sokol (2017b) where they compute the EE
profile for the quarterly-quarterly swap under the Classical ± and the Advanced models. As
is seen there, while the upward EE spikes are present around the payment dates in the
absence of IM, the height of the spikes is, as expected, significantly lower than before.
Nevertheless, the IM remains incapable of completely suppressing the spikes.

Expected Exposure: Numerical Exposure 2

211
1. Mismatched Fix-Float Pay Frequencies: To some extent, the spikes in the presence of IM are
particularly pronounced because of unequal payment frequency on the fixed and the floating
legs on the swap (e.g., semi-annual fixed against quarterly floating is the prevailing market
standard in the US).
2. Matched Fix-Float Pay Frequencies: Indeed, on the semi-annual payment dates, the fixed leg
pays, on average, twice as much as the floating leg, so D is a net payer on the vast majority
of the scenarios, which in turn results in sizeable semi-annual upward spikes when no IM is
present.

The Impact of IM on CVA

1. CVA as Period EE Proxy: CVA constitutes a convenient condensation of EE profiles into a

single number, of considerable practical relevance.
2. Numerical Estimation of the CVA: To demonstrate the impact of IM on CVA for the swap
examples above, Andersen, Pykhtin, and Sokol (2017b) show the CVA calculated from the
expected exposure profiles. To illustrate the impact of the VM on the uncollateralized
exposure, all CVA numbers are shown relative to the CVA of an otherwise uncollateralized
swap.
3. CVA in the Semi-Annual Case: Since the Classical- model does not have spikes, CVA in this
model is reduced by about 2 orders of magnitude by the IM, as one would expect from

𝜙(𝑧(𝑞)) − 𝑧(𝑞)Φ(−𝑧(𝑞))
𝜆(𝑡) ≈
𝜙(0)

The presence of spikes, however, reduces the effectiveness of the IM significantly; for the
case of semi-annual fixed payments, the CVA with IM is about 24% of the CVA without the
IM for the Classical+ model and 15% of the Advanced model.

212
4. CVA in Quarterly Float Case: For the case of quarterly fixed payments, the reduced height of
the spikes renders the IM noticeably more effective in reducing the CVA; the CVA with IM
here is about 9% of the CVA with IM for the Classical+ model and about 5% for the
Advanced model.
5. EE Spikes still dominate CVA: Nonetheless, when M is present, EE spikes still dominate the
CVA, as can be verified for the Classical+ or Advanced model to the CVA for the Classical-
model.
6. Importance of the Advanced Model: Overall the swap flows demonstrate that when the trade
flows within the MPoR are properly modeled, the IM at 99% VaR may not be sufficient to
achieve even a one order of magnitude reduction in the CVA. Also, since spikes dominate
CVA in the presence of IM, accurately modeling the trade flows within the MPoR is
especially important when the IM is present. Hence the Advanced model is clearly preferable
since neither the Classical+ nor the Classical- models produce reasonable CVA numbers for
portfolios with IM.
7. IM CVA Advanced Model Calibration: Finally, it should be noted that even if one uses only
the Advanced model, the impact of IM on CVA may vary significantly, depending on the
trade/portfolio details and the model calibration. In particular the following general
observation can be made.
8. Payment Frequency: Higher frequency of trade payments results in more EE spikes, thus
reducing the effectiveness of the IM.
9. Payment Size: Higher payment size relative to the trade/portfolio volatility results in higher
EE spikes thus reducing the effectiveness of the IM.
10. Payment Asymmetry: EE spikes are especially high for payment dates where only B pays or
where B is almost always the net payer. The presence of such payment dates reduces the
effectiveness of the IM.
11. Model Calibration: In the Advanced Model, the width of these spikes is determined by the
time interval within the 𝑀𝑃𝑜𝑅 where D makes trade payments, i.e., 𝛿𝐶 − 𝛿𝐷 ′ . Thus, larger
the 𝛿𝐶 – i.e., the 𝑀𝑃𝑜𝑅 – and/or smaller the 𝛿𝐷 ′ - i.e., the time interval within the 𝑀𝑃𝑜𝑅
where D does not make trade payments – would result in wider spikes, and thus reduced
effectiveness of IM.

213
12. Illustrating the Advanced Model Effects: Andersen, Pykhtin, and Sokol (2017b) illustrate
briefly the last point using additional calculation. In the swap examples above they use the
baseline timeline calibration of the Advanced model –

𝛿𝐶 = 10 𝐵𝐷

𝛿𝐷 = 8 𝐵𝐷

𝛿𝐶 ′ = 6 𝐵𝐷

𝛿𝐷 ′ = 4 𝐵𝐷

with 𝑀𝑃𝑜𝑅 being equal to the BCBS-IOSCO mandated IM horizon of

𝛿𝐼𝑀 = 10 𝐵𝐷

Their table shows CVA numbers and their ratios for two alternative calibrations of the
Advanced Model mentioned earlier.

Numerical Techniques – Daily Time Grid

1. EE Estimation under Daily Resolution: The discussion so far has made obvious the
importance of accurately capturing exposure spikes from trade flows. To achieve this one
would need to calculate the EE with a daily resolution, something that, if done by brute force
methods, likely will not be feasible for large portfolios.

214
2. Coarse Grid Portfolio - First Pass: In Andersen, Pykhtin, and Sokol (2017a), the authors
discuss a fast approximation that produces a reasonably accurate EE profile without a
significant increase in the computation time relative to the standard, coarse-grid calculations.
The method requires simulation of risk factors and trade flows with a daily resolution, but the
portfolio valuations – which are normally the slowest part of the simulation process – can be
computed on a much coarser time grid.
3. Second Pass using Brownian Bridge: Portfolio values on the daily grid are then obtained by
Brownian Bridge interpolation between the values at the coarse grid points, with careful
accounting for trade flows.

Calculation of the Path-wise IM

1. Need for Numerical IM Calculations: As per

𝐼𝑀(𝑡) = ∑ 𝐼𝑀𝑘 (𝑡)

𝑘=1

and

𝐼𝑀𝑘 (𝑡) = 𝑄𝑞 [𝑉𝑘 (𝑡𝐶 + 𝛿𝐼𝑀 ) + 𝑇𝐹𝑁𝐸𝑇,𝑘 (𝑡𝐶 + 𝛿𝐼𝑀 ; (𝑡𝐶 , 𝑡𝐶 + 𝛿𝐼𝑀 ]) − 𝑉𝑘 (𝑡𝐶 )| ℱ𝑡𝐶 ]

the calculation of the 𝐼𝑀 requires dynamic knowledge of the portfolio value increments
(P&L) across 𝐾 distinct asset classes. Since conditional distributions of the P&L are
generally not known, one must rely on numerical methods to calculate IM.
2. Path-wise IM using Regression Approach: Regression approaches are useful for this,
although the selection of regression variables for large diverse portfolios where IM would
depend on an impracticably large number of risk factors can be difficult.

215
3. Regression Variable - Value of𝑉𝑘 (𝑡𝐶 ): To simplify the regression approach, here the
portfolio value 𝑉𝑘 (𝑡𝐶 ) is chosen as the single regression variable for the asset class 𝑘.
Mathematically, the conditioning on ℱ𝑡𝐶 in

𝐼𝑀𝑘 (𝑡) = 𝑄𝑞 [𝑉𝑘 (𝑡𝐶 + 𝛿𝐼𝑀 ) + 𝑇𝐹𝑁𝐸𝑇,𝑘 (𝑡𝐶 + 𝛿𝐼𝑀 ; (𝑡𝐶 , 𝑡𝐶 + 𝛿𝐼𝑀 ]) − 𝑉𝑘 (𝑡𝐶 )| ℱ𝑡𝐶 ]

is replaced with the conditioning on 𝑉𝑘 (𝑡𝐶 ) with the hope that the projection model would
not have a material impact on the result.
4. Initial Margin Exposure Quantile Re-formulation: For Monte Carlo purposes these
assumptions approximate

𝐼𝑀𝑘 (𝑡) = 𝑄𝑞 [𝑉𝑘 (𝑡𝐶 + 𝛿𝐼𝑀 ) + 𝑇𝐹𝑁𝐸𝑇,𝑘 (𝑡𝐶 + 𝛿𝐼𝑀 ; (𝑡𝐶 , 𝑡𝐶 + 𝛿𝐼𝑀 ]) − 𝑉𝑘 (𝑡𝐶 )| ℱ𝑡𝐶 ]

with

𝐼𝑀𝑘,𝑚 (𝑡𝐶 ) ≈ 𝑄𝑞 [𝑉𝑘 (𝑡𝐶 + 𝛿𝐼𝑀 ) + 𝑇𝐹𝑁𝐸𝑇,𝑘 (𝑡𝐶 + 𝛿𝐼𝑀 ; (𝑡𝐶 , 𝑡𝐶 + 𝛿𝐼𝑀 ]) − 𝑉𝑘 (𝑡𝐶 )| 𝑉𝑘 (𝑡𝐶 )
= 𝑉𝑘,𝑚 (𝑡𝐶 )]

where 𝑚 designates the 𝑚𝑡ℎ simulation path and 𝑇𝐹𝑁𝐸𝑇,𝑘 (∙) is the time 𝑡𝐶 + 𝛿𝐼𝑀 ] value of
all net trade flows scheduled to be paid on the interval (𝑡𝐶 , 𝑡𝐶 + 𝛿𝐼𝑀 ] realized along the
simulation path 𝑚.
5. Portfolio Value Process for IM: Following Andersen, Pykhtin, and Sokol (2017a) the
discounting effects are ignored, and the P&L under the quantile in the right-hand side of

𝐼𝑀𝑘,𝑚 (𝑡𝐶 ) = 𝑄𝑞 [𝑉𝑘 (𝑡𝐶 + 𝛿𝐼𝑀 ) + 𝑇𝐹𝑁𝐸𝑇,𝑘 (𝑡𝐶 + 𝛿𝐼𝑀 ; (𝑡𝐶 , 𝑡𝐶 + 𝛿𝐼𝑀 ]) − 𝑉𝑘 (𝑡𝐶 )| 𝑉𝑘 (𝑡𝐶 )
= 𝑉𝑘,𝑚 (𝑡𝐶 )]

conditional on

216
 𝑉𝑘 (𝑡𝐶 ) = 𝑉𝑘,𝑚 (𝑡𝐶 )

is assumed to be Gaussian with zero drift such that simply

𝐼𝑀𝑘,𝑚 (𝑡𝐶 ) ≈ 𝜎𝑘,𝑚 (𝑡𝐶 )√𝛿𝐼𝑀 Φ−1 (𝑞)

1 2
𝜎𝑘,𝑚 2 (𝑡𝐶 ) ≜ 𝔼 [{𝑉𝑘 (𝑡𝐶 + 𝛿𝐼𝑀 ) + 𝑇𝐹𝑁𝐸𝑇,𝑘 (𝑡𝐶 + 𝛿𝐼𝑀 ; (𝑡𝐶 , 𝑡𝐶 + 𝛿𝐼𝑀 ]) − 𝑉𝑘 (𝑡𝐶 )} | 𝑉𝑘 (𝑡𝐶 )
𝛿𝐼𝑀

= 𝑉𝑘,𝑚 (𝑡𝐶 )]

6. Non-Gaussian Portfolio Value Process: Andersen and Pykhtin (2015) explored non-Gaussian
assumptions, using kernel regressions to estimate the first four conditional moments of the
P&L. As kernel regression for the third and the fourth moments is prone to instability, this
approach was deemed insufficiently robust.
7. Conditional Variance Estimation Parameters Regression: Estimation of the conditional
expectation in

1 2
𝜎𝑘,𝑚 2 (𝑡𝐶 ) ≜ 𝔼 [{𝑉𝑘 (𝑡𝐶 + 𝛿𝐼𝑀 ) + 𝑇𝐹𝑁𝐸𝑇,𝑘 (𝑡𝐶 + 𝛿𝐼𝑀 ; (𝑡𝐶 , 𝑡𝐶 + 𝛿𝐼𝑀 ]) − 𝑉𝑘 (𝑡𝐶 )} | 𝑉𝑘 (𝑡𝐶 )
𝛿𝐼𝑀

= 𝑉𝑘,𝑚 (𝑡𝐶 )]

may be computed by, say, parametric regression or kernel regression. In parametric

regression, the conditional variance is approximated as a parametric function – e.g., a
polynomial – of 𝑉𝑘 (𝑡𝐶 ) with function parameters estimated by least squares regression; see,
for instance, Anfuso, Aziz, Giltinan, and Loukopoulos (2017).
8. Local Standard Deviation Calculation Details: In defining the local standard deviation on
path 𝑚, Andersen, Pykhtin, and Sokol (2017b) follow Pykhtin (2009) and scale the
unconditional standard deviation of 𝑉𝑘 (𝑡𝐶 ) by a ratio of two probability densities; in the
denominator is the probability density of 𝑉𝑘 (𝑡𝐶 ) on path 𝑚; in the numerator is the
probability density that a normally distributed random variable with the same mean and

217
 standard deviation as 𝑉𝑘 (𝑡𝐶 ) would have on the path 𝑚. The actual calculation of the local
standard deviation is performed as presented in Pykhtin (2009).
9. Usage in Conjunction with SIMM: It is t be noted in passing that if a dealer uses an out-of-
model margin calculator – e.g., the SIMM method in International Swaps and Derivatives
Association (2016) – an adjustment will be needed to capture the difference between an in-
model IM computed as above by regression methods and the out-of-model margin calculator
actually used. Many possible methods could be contemplated here - e.g. a multiplicative
adjustment factor that aligns the two margin calculations at time 0.

Calculation of Path-wise Exposure

1. IM Adjusted Path-wise Exposure Estimation: Once the IM on a path is calculated, the

exposure on that path, in principle, can be obtained by directly subtracting the calculated IM
value from the no-IM exposure. While not of central importance to this chapter, it is to be
noted that the path-wise IM results allow for a straightforward computation of the Margin
Value Adjustment (MVA) – see, for e.g. Andersen, Duffie, and Song (2017).
2. Dominance of IM Covered Trades: This approach works well when all the trades covered by
an IM represent a reasonably small fraction of the netting set. However, when the netting set
is dominated by the IM covered trades, this approach suffers from two issues that have an
impact on the accuracy of the EE – and therefore CVA – calculations.
3. Simulation Noise: Suppose that all trades of the netting set are covered by IM and that

𝛿𝐼𝑀 = 𝛿𝐶

If all trades belong to the same asset class, then the non-zero exposure will be realized on
average only 1% of the time between the exposure spikes, as IM by design covers 99% of
the P&L increase. If multiple asset classes are present in the netting set, the percentages of

218
 non-zero realizations will be even less because of the disallowed diversification across asset
classes. Thus, EE calculated by direct exposure simulation will be extremely noisy between
spikes.
4. Non-Normality: IM is calculated under the assumption that, conditional on a path, the P&L
over the 𝑀𝑃𝑜𝑅 is Gaussian. When all – or most – of the netting is covered by IM, the EE
between the spikes calculated by a direct of the exposure is very sensitive to deviations from
local normality. If the conditional P&L distribution has a heavier (lighter) upper tail than the
Gaussian distribution, IM between the spikes will be understated (overstated) and EE will
therefore be overstated (understated).
5. Time 𝑡𝐶 Path-wise Expected Exposure: Bothe of these issues can be remedied by calculating
a time 𝑡𝐶 path-wise expected exposure for a time 𝑡 default, rather than the exposure itself. If
our target exposure measure is the unconditional time-0 EE, this substitution is valid, of
course, by the Law of Iterated Expectations.
6. Simplification of the Advanced Model: To proceed with this idea, one simplifies the
Advanced Model slightly - and assumes – as the Classical Model does – that D and C stop
posting margin simultaneously – i.e. that

𝛿𝐶 = 𝛿𝐷

Then

𝐾 +

𝐸𝐴𝐷𝑉𝐴𝑁𝐶𝐸𝐷,𝑚 (𝑡) = [𝑉𝑚 (𝑡) − 𝑉𝑚 (𝑡𝐶 ) + 𝑈𝐴𝐷𝑉𝐴𝑁𝐶𝐸𝐷,𝑚 (𝑡) − ∑ 𝐼𝑀𝑘 (𝑡)]

𝑘=1

where 𝑈𝐴𝐷𝑉𝐴𝑁𝐶𝐸𝐷,𝑚 (𝑡) is the realization of the right hand side of

𝑈𝐴𝐷𝑉𝐴𝑁𝐶𝐸𝐷,𝑚 (𝑡) = 𝑇𝐹𝐶→𝐷 (𝑡; (𝑡𝐶 ′ , 𝑡𝐷 ′ ]) + 𝑇𝐹𝑁𝐸𝑇 (𝑡; (𝑡𝐷 ′ , 𝑡])

on path 𝑚.

219
7. Expected Exposure Conditional on 𝑉𝑘 (𝑡𝐶 ) = 𝑉𝑘,𝑚 (𝑡𝐶 ): The expectation of this exposure
conditional on

𝑉𝑘 (𝑡𝐶 ) = 𝑉𝑘,𝑚 (𝑡𝐶 )

is, in the Advanced model

𝐸𝐸𝑚 (𝑡; 𝑡𝐶 ) ≜ 𝔼[𝐸𝐴𝐷𝑉𝐴𝑁𝐶𝐸𝐷 (𝑡) | 𝑉(𝑡𝐶 ) = 𝑉𝑚 (𝑡𝐶 )]

Averaging

𝐾 +

𝐸𝐴𝐷𝑉𝐴𝑁𝐶𝐸𝐷,𝑚 (𝑡) = [𝑉𝑚 (𝑡) − 𝑉𝑚 (𝑡𝐶 ) + 𝑈𝐴𝐷𝑉𝐴𝑁𝐶𝐸𝐷,𝑚 (𝑡) − ∑ 𝐼𝑀𝑘 (𝑡)]

𝑘=1

and

𝐸𝐸𝑚 (𝑡; 𝑡𝐶 ) ≜ 𝔼[𝐸𝐴𝐷𝑉𝐴𝑁𝐶𝐸𝐷 (𝑡) | 𝑉(𝑡𝐶 ) = 𝑉𝑚 (𝑡𝐶 )]

over all Monte Carlo paths will lead to the same result for 𝐸𝐸(𝑡; 𝑡𝐶 ) up to Monte Carlo
sample noise.
8. Evolution of the Underlying Portfolio: To calculate the right-hand side of

𝐸𝐸𝑚 (𝑡; 𝑡𝐶 ) ≜ 𝔼[𝐸𝐴𝐷𝑉𝐴𝑁𝐶𝐸𝐷 (𝑡) | 𝑉(𝑡𝐶 ) = 𝑉𝑚 (𝑡𝐶 )]

analytically, the mismatch of the portfolio value over the 𝑀𝑃𝑜𝑅 is assumed to be Gaussian.
Specifically, given

𝑉(𝑡𝐶 ) = 𝑉𝑚 (𝑡𝐶 )

220
 it is assumed that 𝑇𝐹𝑚 (𝑡; (𝑡𝐶 , 𝑡]) is known at time 𝑡𝐶 and equal to its realization on the path,
and that

𝑉(𝑡𝐶 ) = 𝑉𝑚 (𝑡𝐶 ) ~ 𝒩(𝑇𝐹𝑚 (𝑡; (𝑡𝐶 , 𝑡]), 𝜎𝑚 (𝑡𝐶 )√𝛿𝐶 )

where 𝑇𝐹𝑚 (𝑡; (𝑡𝐶 , 𝑡]) represents the sum of all scheduled portfolio payments (𝑡𝐶 , 𝑡] on that
path 𝑚, and 𝜎𝑚 2 (𝑡𝐶 ) is defined as

1
𝜎𝑚 2 (𝑡𝐶 ) ≜ 𝔼[{𝑉(𝑡) + 𝑇𝐹𝑚 (𝑡; (𝑡𝐶 , 𝑡]) − 𝑉𝑚 (𝑡𝐶 )}2 | 𝑉(𝑡𝐶 ) = 𝑉𝑚 (𝑡𝐶 )]
𝛿𝐶

9. Netting Set Level Portfolio Volatility: It is emphasized that the expression for 𝜎𝑚 2 (𝑡𝐶 ) above
differs from

1 2
𝜎𝑘,𝑚 2 (𝑡𝐶 ) ≜ 𝔼 [{𝑉𝑘 (𝑡𝐶 + 𝛿𝐼𝑀 ) + 𝑇𝐹𝑁𝐸𝑇,𝑘 (𝑡𝐶 + 𝛿𝐼𝑀 ; (𝑡𝐶 , 𝑡𝐶 + 𝛿𝐼𝑀 ]) − 𝑉𝑘 (𝑡𝐶 )} | 𝑉𝑘 (𝑡𝐶 )
𝛿𝐼𝑀

= 𝑉𝑘,𝑚 (𝑡𝐶 )]

in several important ways. First the portfolio used in

1
𝜎𝑚 2 (𝑡𝐶 ) ≜ 𝔼[{𝑉(𝑡) + 𝑇𝐹𝑚 (𝑡; (𝑡𝐶 , 𝑡]) − 𝑉𝑚 (𝑡𝐶 )}2 | 𝑉(𝑡𝐶 ) = 𝑉𝑚 (𝑡𝐶 )]
𝛿𝐶

spans the entire netting set rather than just the sub-portfolio covered by the IM and associated
with the asset class 𝑘. Second the length of the time horizon is 𝛿𝐶 which may be different
from 𝛿𝐼𝑀 . As before, 𝜎𝑚 2 (𝑡𝐶 ) can be calculated by parametric or kernel regression.
10. Unconditional Brownian Bridge Expected Exposure: Using this, one can calculate

𝐸𝐸𝑚 (𝑡; 𝑡𝐶 ) ≜ 𝔼[𝐸𝐴𝐷𝑉𝐴𝑁𝐶𝐸𝐷 (𝑡) | 𝑉(𝑡𝐶 ) = 𝑉𝑚 (𝑡𝐶 )]

221
 to obtain, in the Advanced model

𝐸𝐸𝑚 (𝑡) = 𝜎𝑚 (𝑡𝐶 )√𝛿𝐶 [𝑑𝑚 (𝑡)Φ(𝑑𝑚 (𝑡))] + 𝜙(𝑑𝑚 (𝑡))

−𝑃𝑇𝐹𝐴𝐷𝑉𝐴𝑁𝐶𝐸𝐷,𝑚 (𝑡; (𝑡𝐶 , 𝑡]) − ∑𝐾

𝑘=1 𝐼𝑀𝑘,𝑚 (𝑡)
𝑑𝑚 (𝑡) ≜
𝜎𝑚 (𝑡𝐶 )√𝛿𝐶

where – omitting arguments –

𝑃𝑇𝐹𝐴𝐷𝑉𝐴𝑁𝐶𝐸𝐷,𝑚 = 𝑇𝐹𝑚 − 𝑈𝐴𝐷𝑉𝐴𝑁𝐶𝐸𝐷,𝑚

are the net trade flows on the 𝑀𝑃𝑜𝑅 actually paid, according to the Advanced model.
11. Relaxing the 𝛿𝐷 = 𝛿𝐶 Simplification: Recall that

𝐸𝐸𝑚 (𝑡) = 𝜎𝑚 (𝑡𝐶 )√𝛿𝐶 [𝑑𝑚 (𝑡)Φ(𝑑𝑚 (𝑡))] + 𝜙(𝑑𝑚 (𝑡))

−𝑃𝑇𝐹𝐴𝐷𝑉𝐴𝑁𝐶𝐸𝐷,𝑚 (𝑡; (𝑡𝐶 , 𝑡]) − ∑𝐾

𝑘=1 𝐼𝑀𝑘,𝑚 (𝑡)
𝑑𝑚 (𝑡) ≜
𝜎𝑚 (𝑡𝐶 )√𝛿𝐶

was derived using the Advanced Model for the simplifying case

𝛿𝐷 = 𝛿𝐶

𝛿𝐷 < 𝛿𝐶

would, however, result in a significant understatement of the EE between spikes – e.g. for the
earlier swap sample, the understatement would be around 22%.
12. Handling the 𝛿𝐷 < 𝛿𝐶 Case using Scaling: Directly extending

222
 𝐸𝐸𝑚 (𝑡) = 𝜎𝑚 (𝑡𝐶 )√𝛿𝐶 [𝑑𝑚 (𝑡)Φ(𝑑𝑚 (𝑡))] + 𝜙(𝑑𝑚 (𝑡))

−𝑃𝑇𝐹𝐴𝐷𝑉𝐴𝑁𝐶𝐸𝐷,𝑚 (𝑡; (𝑡𝐶 , 𝑡]) − ∑𝐾

𝑘=1 𝐼𝑀𝑘,𝑚 (𝑡)
𝑑𝑚 (𝑡) ≜
𝜎𝑚 (𝑡𝐶 )√𝛿𝐶

to cover

𝛿𝐷 < 𝛿𝐶

is, however, not straightforward. Therefore Andersen, Pykhtin, and Sokol (2017b) propose a
simple scaling solution. Here the paths of the exposure without the IM are simulated first for
the Advanced model, i.e. for

𝑚𝑖𝑛
𝐸𝐴𝐷𝑉𝐴𝑁𝐶𝐸𝐷 (𝑡) = [𝑉(𝑡) − 𝑉(𝑇𝑖 ) + 𝑇𝐹𝐶→𝐷 (𝑡; (𝑡𝐶 ′ , 𝑡𝐷 ′ ]) + 𝑇𝐹𝑁𝐸𝑇 (𝑡; (𝑡𝐷 ′ , 𝑡])
𝑇𝑖 ∈ [𝑡𝐶 , 𝑡𝐷 ]

𝐾 +

− ∑ 𝐼𝑀𝑘 (𝑡)]
𝑘=1

with

𝐼𝑀(𝑡) = 0

Then for each exposure simulation time 𝑡 and each measurement path 𝑚 the exposure value
𝐸𝐸 (𝑡)
is multiplied by the ratio of 𝐸𝐸 𝑚 (𝑡) where 𝐸𝐸𝑚 (𝑡) is computed from
0,𝑚

𝐸𝐸𝑚 (𝑡) = 𝜎𝑚 (𝑡𝐶 )√𝛿𝐶 [𝑑𝑚 (𝑡)Φ(𝑑𝑚 (𝑡))] + 𝜙(𝑑𝑚 (𝑡))

−𝑃𝑇𝐹𝐴𝐷𝑉𝐴𝑁𝐶𝐸𝐷,𝑚 (𝑡; (𝑡𝐶 , 𝑡]) − ∑𝐾

𝑘=1 𝐼𝑀𝑘,𝑚 (𝑡)
𝑑𝑚 (𝑡) ≜
𝜎𝑚 (𝑡𝐶 )√𝛿𝐶

223
 and 𝐸𝐸0,𝑚 (𝑡) is the special case of 𝐸𝐸𝑚 (𝑡) where

𝐼𝑀(𝑡) = 0

Numerical Example

1. EE Computations using Different Approaches: To illustrate the benefits of the conditional EE

simulation method described above, Andersen, Pykhtin, and Sokol (2017b) turn to the 2Y
IRS with unequal payment frequencies considered earlier. They illustrate EE profile
comparisons using several computational approaches.
2. Expected Exposures at Spike - Comparison: First they focus their attention on the upward EE
spikes produced by trade payments at the 1-year point. Both the unconditional and the
conditional EE estimators here produce an almost identical spike, exceeding the benchmark
spike height – the consequence of using kernel regression and a Gaussian distribution to
estimate the IM.
3. Expected Exposure between Spikes - Comparison: They then show the EE exposure at a fine
exposure scale, allowing for a clear observation of the EE between spikes. There the
advantages of the conditional EE approach can be clearly seen.
4. Reduction in Simulation Noise: While both methods use Monte Carlo simulation with 5,000
paths, the simulation noise in the conditional EE approach is substantially less than that in the
unconditional EE approach. In fact, the conditional EE noise is even less than in the
benchmark EE results that were calculated using 50,000 paths.
5. Reduced Error from Non-Gaussian Dynamics: Andersen, Pykhtin, and Sokol (2017b) have
used a high-volatility lognormal interest rate model in their example, to produce significant
deviations in a 10 𝐷 horizon. This non-normality is the main reason for the deviation of the
conditional and the unconditional EE curves from the benchmark. In estimating
224
 +
𝐸𝐸(𝑡) = 𝔼 [{𝑉(𝑡) − 𝑉𝑘 (𝑡𝐶 ) − 𝑄𝑞 (𝑉𝑘 (𝑡𝐶 + 𝛿𝐼𝑀 )| ℱ𝑡𝐶 )} ]

𝑡𝐶 = 𝑡 − 𝛿𝐶

the conditional estimator uses a Gaussian distribution to approximate both the IM and the
portfolio increment, resulting in partial error cancellation. Such error cancellation does not
take place for the unconditional estimator which uses the empirical – here lognormal –
distribution for the portfolio increment, yet estimates IM from a Gaussian distribution.
Between the spikes the EE errors for the unconditional estimator are therefore significantly
larger here than for the conditional estimator.

Conclusion

1. Consequences from the BCBS-IOSCO IM Rules: There is universal agreement that the new
BCBS-IOSCO IM rules will lead to very substantial postings of margin amounts into
segregated accounts accompanied by inevitable increase in the funding costs (MVA) that the
dealers will face when raising funds for the IM. According to conventional wisdom, these
postings, while expensive, should effectively eliminate counter party risk.
2. Handling the Trade Flow Spikes: This chapter examines the degree to which the bilateral IM
required by the BCBS-IOSCO margin rules suppresses counter party exposure. As shown by
Andersen, Pykhtin, and Sokol (2017a) any trade flow to the defaulting party for which it does
not return margin during the MPoR causes a spike in the exposure profile.
3. Ignoring the Trade Flow Spikes: These spikes are often ignored by dealers a being spurious
or as being part of settlement risk. In reality these spikes are integral part of the exposure
profile and represent real risk that has previously materialized in many well-documented

225
 incidents, notably the non-payment of Lehman of reciprocal margin to trade payments that
arrived at around the time of the bankruptcy filing.
4. IM reduces Expected Exposure Considerably: The chapter shows that, under very general
assumptions the BCBS IOSCO IM specified as the 99% 10 𝐷 𝑉𝑎𝑅 reduces the exposure
between the spikes by a factor of over 100 but fails to suppress the spikes by a comparable
degree. This happens because IM is calculated without reference to trade payments, and is
based only on the changes to the portfolio value resulting from the risk factor variability. As
an example, Andersen, Pykhtin, and Sokol (2017b) show that IM reduces the CVA of a 2Y
IRS with VM by only a factor of 7.
5. Impracticality of Increasing the IM: While 𝑉𝑎𝑅 based 𝐼𝑀 fails to fully suppress the
contribution of exposure spikes to 𝐶𝑉𝐴 and 𝐸𝐴𝐷, increasing the 𝐼𝑀 to always exceed the
peak exposure would be impractical, and would require moving large amounts of collateral
back and forth in a matter of days.
6. IM CVA dominated by Spikes: Another important property of 𝐶𝑉𝐴 under full 𝐼𝑀 coverage is
that it is dominated by exposure spikes; in their 2Y IRS example, Andersen, Pykhtin, and
Sokol (2017b) find that spike contribution to the 𝐶𝑉𝐴 is about 95% in the presence of 𝐼𝑀
(compared to about 20% without 𝐼𝑀).
7. Modeling 𝐶𝑆𝐴 Time-line in 𝐼𝑀: Thus, in the presence of 𝐼𝑀, the focus of exposure modeling
should be on capturing the impact of trade payments, which involves making realistic
assumptions on what the dealer and the client are expected to make contingent on the client’s
default.
8. Computationally Feasible Daily Portfolio Value: Furthermore, to accurately calculate the
𝐶𝑉𝐴 mostly produced by narrow exposure spikes, one needs to produce exposure on a daily
time grid. A method for producing daily exposures without daily portfolio re-evaluations was
discussed above, along with other useful numerical techniques.
9. Approach taken by the CCP: A natural question to ask is why similar payment effects have
not been recognized in trading through central counterparties (CCPs), which also require 𝐼𝑀
posting that is typically based on 99% 𝑉𝑎𝑅 over the 𝑀𝑃𝑜𝑅. As it turns out, CCPs already
use a mechanism that amounts to netting of trade and margin payments.

226
10. Infeasibility of the CCP Approach: Unfortunately, the same approach cannot be adopted in
bilateral trading as it would require changing all of the existing trade documentation, which
is a practical impossibility.
11. Nettability of Trade/Margin Payments: While a trade payment and its reciprocal margin
payment cannot be netted in bilateral trading, this lag can be eliminated and the two
payments made to fall on the same day by making a simple change in the CSA.
12. CSA Amendment for Trade Payments: Specifically, if the CSA is amended to state that
known trade payments due to arrive prior to the scheduled margin payment date must be
subtracted from the portfolio valuation for the purposes of margin – technically this
amendment effectively sets the VM based on a 2-day portfolio forward value – then the call
for the reciprocal margin will happen ahead of time, and it will on the same day as the trade
payment – a no lag margin settlement.
13. Reduction in the 𝑀𝑃𝑜𝑅 Duration: From an IT and a back-office perspective, this change in
the CSA is relatively easy to align with existing mark-to-market and cash-flow processes,
and is beneficial in several ways. First it shortens the duration of the exposure spikes and the
𝑀𝑃𝑜𝑅 overall, reducing counter-party risk.
14. Margin vs 𝑀𝑇𝑀 2𝐷 Lag: Second it makes margin follow 𝑀𝑇𝑀 without a 2𝐷 lag, thereby
eliminating the need to use outside funding to fund hedging during this 2𝐷 period.
15. Trade/Margin Payments Reciprocal Concurrence: Finally, with the reciprocal trade and the
margin payments falling on the same day, payment-versus-payment services (𝑃𝑣𝑃) such as
CLS Bank (Galati (2002), Lindley (2008), Brazier (2015)) may be able to settle trade and
margin payments together, reducing residual counterparty risk even further.

References

• Andersen, L., D. Duffie, and Y. Song (2017): Funding Value Adjustment eSSRN

227
• Andersen, L. and M. Pykhtin (2015): Accounting for Dynamic Initial Margin in Credit
Exposure Models Risk Minds International
• Andersen, L., M. Pykhtin, and A. Sokol (2017a): Re-thinking Margin Period of Risk eSSRN
• Andersen, L., M. Pykhtin, and A. Sokol (2017b): Credit Exposure in the Presence of Initial
Margin eSSRN
• Anfuso, C., D. Aziz, K. Loukopoulos, and P. Giltinan (2017): A Sound Modeling and
Backtesting Framework for Forecasting Initial Margin eSSRN
• Basel Committee on Banking Supervision (2013): Basel III: The Liquidity Coverage Ratio
and Liquidity Risk Monitoring Tools
• Basel Committee on Banking Supervision (2015): Margin Requirements for Non-centrally
Cleared Derivatives
• Brazier, J. (2015): CLS, Markit Launch Cross-Currency FX Settlement Service Waters
Technology
• Galati, G. (2002): Settlement Risk in Foreign Exchange Markets and CLS Bank BIS
Quarterly Review December 55-66
• Gregory, J. (2015): The XVA Challenge: Counterparty Credit Risk, Funding, Collateral, and
Capital, 3rd Edition Wiley
• International Swaps and Derivatives Association (2016): ISDA SIMM Methodology
• Lindley, R. (2008): Reducing Foreign Exchange Settlement Risk BIS Quarterly Review
September 53-65
• Pykhtin, M. (2010): Collateralized Credit Exposure, in: Counterparty Credit Risk (editor: E.
Canabarro) Risk Books

228
 Basel III Framework for Backtesting Exposure Models

Abstract

1. Standard Practices for IMM Back-testing: A central component of the Basel III (B3)
standards is the Sound Practices for Backtesting (Basel Committee for Banking Supervision
(2010)), i.e., a summary of the strict regulatory guidances on how to validate and back-test
Internal Methods Models (IMM) for credit exposure.
2. Statistical Credit Exposure Backtesting Framework: In their work, Anfuso, Karyampas, and
Nawroth (2017) define a comprehensive framework to backtest credit exposure models,
highlighting the proposed features against the regulatory requirements.
3. Risk Factor Dynamical Evolution Backtesting: Their framework contains four main pillars.
First is the risk factor backtesting, i.e., assessment of the forecasting ability of the stochastic
differential equations (SDEs) used to describe the dynamics of the single factor.
4. Risk Factor Correlation Estimator Backtesting: Next is the correlations backtesting, i.e., the
assessment of the statistical estimators used to describe the cross-asset evolution.
5. Representative Firm Portfolio Backtesting: Third is the portfolio backtesting, i.e., the
assessment of the complete exposure model – SDEs + correlations + pricing – for portfolios
that are representative of the firm’s exposure.
6. Computation of the Capital Buffer: Last is the computation of the capital buffer, i.e., the
extra amount of capital that the firm should hold if the model framework is not adequate –
using the outcomes of the pillars above.
7. Distributional Tests for Collateralized/Uncollateralized: Anfuso, Karyampas, and Nawroth
(2017) show with concrete examples in the cases of collateralized and uncollateralized
models how to perform distributional tests with respect to different risk metrics.

229
8. Discriminatory Power Analysis across Forecasting Horizons: They produce discriminatory
power analysis for all the tests introduced, providing exact methods to aggregate backtesting
results across forecasting horizons.
9. Capital Remedies for Model Deficiencies: Most importantly, the third and the fourth pillars
define a sound quantitative approach for computing capital remedies for potential model
deficiencies.

Introduction

1. Validating and Backtesting IMM: The central pillar of the Basel III (B3) document is the
Sound Practices for Backtesting (Basel Committee for Banking Supervision (2010)), i.e., a
summary of strict regulatory guidelines on how to validate and backtest Internal Methods
Models (IMM) for credit exposure.
2. European Basel III - The CRD4 Requirements: Similarly, a series of requirements for CRD4
– the European equivalent of B3 – indicate and define backtesting and validation as core
component for good governance of IMM firms.
3. Importance of IMM for CVA: From a dealer perspective, the new regulatory changes
introduced with B3 – e.g. CVA capital charge – have stressed even more the importance of
IMM in making the capital costs of the businesses sustainable.
4. Importance of Backtesting for IMM: At the same time, the increasing complexity of the
capital framework requires a thorough approach to the validation and the monitoring of the
model performance.
5. Support from Regulators and Stake-holders: A sound backtesting methodology is therefore
the key tool to both prove to the regulators the soundness of the models and to assure the
stakeholders that the capital position of the firm is in sound modeling grounds.
6. Holistic Qualitative/Quantitative Model Assessment: The assessment of a model is a holistic
process that has both qualitative and quantitative elements. While the former may have a

230
 decisive weight for the choice of a given model many possible, the latter are the ones to be
considered for backtesting.
7. Validation of the NULL Hypothesis: In particular, the performance of a model should be
judged in terms of its forecasting ability. In statistical jargon, backtesting should address the
question: Can we reject the NULL hypothesis – i.e., the model – based on the available
historical data?
8. Statistical Testing of the Forecasting Ability: In the definition employed in this chapter,
backtesting is therefore a set of statistical tests that measures the forecasting ability of the
model using the data history available as comparison.
9. Assessment Metric - Aggregation of p-values: The final metric is based on the aggregation of
the given p-values of the single tests, rejecting the model if an a priori determined threshold
is breached.
10. Framework Features vs. Regulatory Guidelines: This chapter presents a complete framework
to backtest credit exposure models. The next section gives a brief overview of the relevant
metrics for the counterparty credit risk and summarizes the features of the framework against
the new regulatory guidelines.
11. Complete Backtesting Cycle Details: In subsequent sections the methodology is presented in
detail for the full backtesting cycle, i.e., the risk factor evolution models, the correlation
models, and the portfolio exposure metrics. Later sections show how to compute the capital
buffers based on the back-testing results. Finally, conclusions are drawn.

Basic Concepts and the Need for Backtesting

1. Credit Counterparty Exposure - Definition: Credit counterparty exposure is defined as the

amount a dealer A could potentially loos in the case that its counterparty B defaults.

231
2. Estimating the CP Exposure Distribution: The exposure – from A’s perspective – is
computed from the forecasted distribution of prices of the financial contracts that constitute
the portfolio of counterparty B at any future date.
3. Components required for the Exposure Estimation: The main building blocks required for the
computation are the following: first, the scenario simulations for the underlying risk factors –
generated with what is referred to here as the Risk Factor Evolution (RFE) models – and
second, the pricing at each scenario to generate the 𝑀𝑇𝑀(𝑡) distribution at any future date 𝑡.
4. Expression for the Expected Positive Exposure: The relevant exposure metric from the
regulatory perspective is the Expected (Positive) Exposure at time 𝑡 which is defined as:

𝐸𝐸(𝑡) = 𝔼[𝑀𝑇𝑀+ (𝑡)]

where

𝑀𝑇𝑀+ (𝑡) = max(𝑀𝑇𝑀(𝑡), 0)

5. RWA for CVA and Capital: The same exposure profile 𝐸𝐸(𝑡) enters in the computation of
the Risk Weighted Assets (RWA) of a given counterparty both for the CVA and for the
default capital charges.
6. Modeling the Risk Factor Evolution: To compute the 𝑀𝑇𝑀(𝑡) distribution – and most
importantly the 𝐸𝐸(𝑡) profile – at any future time 𝑡 one needs to forecast the evolution of the
risk factor values. Those risk factors can also be dependent on each other – correlation
assumption between the risk factors.
7. Loss Impact of RFE Mis-specification: The more accurately the RFE model is specified, the
more realistic is the exposure calculation. If the RFE is mis-specified the exposure figure can
be wrongly stated and the losses may occur with higher probability than expected in the case
of counterparty defaults.
8. Risk Neutral Measure for the CVA: As observed in Kenyon and Stamm (2012), there is a
potential for divergence in the choice for calibration for the RFE models. On the one hand,

232
 the 𝐸𝐸(𝑡) profile used for CVA calculation – a price – should be based on market
calibration.
9. Historical Measure for Capital Charges: On the other hand, default charges require a forecast
in the real-world measure and therefore a historical calibration would be most suitable.
10. Complication from the Dual Measures: The backtesting methodology described in the
following is agnostic to the choice of the model calibration. Nevertheless, by construction,
the regulatory requirements for backtesting models are addressed generally by historical
models. In view of Anfuso, Karyampas, and Nawroth (2017), this apparent dichotomy is one
of the key quantitative challenges for the industry after Basel III.
11. Regulatory Valuation Validation Framework: Following the guidances from regulators,
Anfuso, Karyampas, and Nawroth (2017) define a framework that has four main pillars.
12. Risk Factor Dynamics Backtesting: The first is Risk Factor backtesting, i.e., the assessment
of the forecasting ability of the Stochastic Differential Equations (SDEs) used to describe the
dynamics of the single risk factors. It can be seen that the calibration of the SDE – market
implied or historical – has a crucial influence on this assessment.
13. Risk Factor Correlation Backtesting: Next is the correlations backtesting, i.e., the assessment
of the estimators used to model the cross-asset evolution.
14. The Representative Portfolio Backtesting: Third is the portfolio backtesting, i.e., the
assessment of the complete exposure model – RFEs + Correlations + Pricing – for portfolios
that are representative of the firm’s exposure.
15. Model Reserve Capital Buffer Calculation: The computation of the capital buffer, i.e., the
extra amount of capital that the form should hold if the model framework is not adequate –
see the outcomes of the three pillars above – is the final pillar.
16. Diagnostic vs. Deficiency Remedy Pillars: The first three pillars are diagnostic whereas the
fourth comes as a remedy for potential deficiencies in the exposure models.
17. Regulatory Guidance vs. Framework Response: The section below summarizes the relevant
regulatory guidances for backtesting and how the framework in this chapter addresses them.
The specifics of the proposed solutions are described in more detail in the sections below.

233
Regulatory Guidances

1. Guidance #1: The performance of the market risk factor should be validated using
backtesting. The validation must be able to identify poor performances in individual risk
factors.
a. Methodology => The forecasting capability of the RFE models and their calibrations
is back tested at multiple backtesting horizons, making use of different distributional
tests.
b. Compliance => Full.
2. Guidance #2: Validation of the EPE models and all the relevant models that input into the
calculation of the EPE must be made using forecasts initialized on a number of historical
dates.
a. Methodology => The sampling of the backtesting is on a bi-weekly frequency,
spanning all the available data history. Since the sampling is very close, no statistical
bias due to the selection of the sampling frequency is introduced.
b. Compliance => Full.
3. Guidance #3: Historical backtesting on representative counterparty portfolios and market risk
factors must be part of the validation process. At regular intervals, as dictated by its
supervisor, the dealer must conduct backtesting on a number of representative counterparty
portfolios and the market risk factor models. The representative portfolios must be chosen
based on the sensitivity to the material risk factors and correlations to which a dealer is
exposed.
a. Methodology => The portfolio backtesting is performed with suitable metrics that do
not penalize the conservative estimates of the EPE. Representative counterparties can
be chosen by a given dealer based on their RWA contributions.
b. Compliance => Fully compliant from a methodology perspective. The dealer should
additionally ensure that the selected counterparties are representative.

234
4. Guidance #4: Backtesting of the EPE and all the relevant models that input into the
calculation of the EPE must be based on the recent performance.
a. Methodology => The framework is agnostic to the data history that is selected, since
that is an input. It can therefore be applied to assess recent and longer-term
performances – though the statistical significance of the results will not be equivalent
given the different amount of historical realizations that are being backtested.
b. Compliance => Full.
5. Guidance #5: The frequency with which the parameters of an EPE model are updated needs
to be assumed as part of the ongoing validation process.
a. Methodology => The calibration is fully accounted for the RF, the correlations, and
the portfolio backtesting.
b. Compliance => Full.
6. Guidance #6: Dealers need to unambiguously define what constitutes acceptable and
unacceptable performance for their EPE models and the models that input into the calculation
of the EPE and have a written policy in place that describes how unacceptable performance
will be remediated.
a. Methodology => The framework gives a quantitative probabilistic interpretation of
the performance that allows unambiguous acceptance or rejection of a model.
b. Compliance => Fully compliant from a model perspective. The dealer should
additionally ensure that the acceptance threshold is sufficiently conservative.
7. Guidance #7: IMM firms need to conduct hypothetical portfolio backtesting that is designed
to test risk factor model assumptions, e.g., the relationships between the tenors of the same
risk factors, and the modeled relationships between the risk factors.
a. Methodology => The chapter does backtest the model correlation assumptions with a
coherent extension of the methodology applied to the single risk factors.
b. Compliance => Full.
8. Guidance #8: Firms must backtest their EPE models and all relevant models that input into
the calculations of the EPE out to long horizons of at least one year.
a. Methodology => Multiple horizons - shorter and longer than one year - are backtested
at every level of granularity, i.e., RFs, correlations, and portfolio exposures.

235
 b. Compliance => Full.
9. Guidance #9: Firms must validate their EPE models and all relevant models that input into
the calculation of the EPE out to the time horizons commensurate with the maturity of trades
covered by the IMM waiver.
a. Methodology => Multiple horizons commensurate with the maturity if the trades are
backtested at every level of granularity, i.e., RF’s, correlations, and portfolio
exposures.
b. Compliance => Full.
10. Guidance #10: Prior to the implementation of a new EPE model or a new model that inputs
into the calculation of the EPE, a dealer must carry out a beck testing of the EPE model and
all relevant models that input into the calculation of the EPE at a number of distinct time
horizons using historical data for movements in the market risk factors for a range of
historical periods covering a wide range of market conditions.
a. Methodology => The framework described, because of its granularity and modularity,
can be applied for periodic regulatory back testing as well as initial validation of a
given model.
b. Compliance => Full.

RF Backtesting: The Backtesting Construction for Collateralized and

Uncollateralized Models

1. Basic Components of the Framework: The RFE models are the most atomic components of
the exposure framework.
2. Cross Horizon RF Backtesting: As for regulatory guidances 1, 8, and 9, their performance
should be assessed for different forecasting horizons and the predicted distributions should be
consistent with the realized history of the corresponding risk factors.

236
3. Probability Integral Transform (PIT) Definition: The statistical tool that is the basis for the
backtesting is the Probability Integral Transform (PIT – Gunther, Diebold, and Tay (1998),
Kenyon and Stamm (2012)) defined as

𝑟𝑛

𝐹(𝑟𝑛 ) = ∫ 𝜙(𝑥)𝑑𝑥
−∞

where 𝑟𝑛 is the realization of the given random variable and 𝜙(∙) is its predicted distribution.
4. Application to the RF Distribution: It is clear that when one applies PIT to a set of i.i.d.
variables 𝑟𝑛 using the correct distribution of 𝑟𝑛 the transformed set

𝑦𝑛 = 𝐹(𝑟𝑛 )

is uniformly distributed.
5. Statistical Metric for Mismatch/Departure: The distance between the transformed set 𝑦𝑛 and
𝑈[0, 1] distribution – in a statistical sense – can therefore be used as a goodness of the model
𝜙(∙) to describe the random variable 𝑟𝑛 .
6. RF Market Generation and Evolution: In practice, the input for RFE backtesting analysis is
the time series of the given risk factor and the model – the SDE and its calibration – used to
describe its evolution.
7. PIT Map onto [0, 1] Space: The PIT for the predicted distribution can be used to map the
realized values of a risk factor – or their variations, see below – to a set of values in the
interval [0, 1].
8. The Transformed Set Model Performance Assessment: The transformed set is then is used to
assess the model performance by applying the standard distribution tests.
9. Collateralized vs. Uncollateralized Time Grids: For collateralized and uncollateralized
models the construction differs because of the presence of the multiple time scales. In both
cases a grid of sampling points 𝑡𝑘 is used.

237
10. Criteria for Time Grid Selection: The sampling points define the origin of the backtesting
experiment and they should be on a sufficiently fine grid and for a sufficiently long time
series so as to ensure the following:
a. An acceptable discriminatory power for the test – in relation to the size of the data
history – and:
b. An absence of significant statistical bias caused by the sampling sequence – if the
sampling points are too distant, the backtesting results from an equivalent sequence
with different arbitrary starting point may differ.
11. Horizon used for Backtesting: The backtesting is carried out for an arbitrary set of horizons
{ℎ1 , ⋯ , ℎ𝑛 } and for every ℎ𝑖 a single result is produced.
12. Horizon Matching Firms’ Exposure Structure: The choice of the set {ℎ1 , ⋯ , ℎ𝑛 } should be so
as to reflect the portfolio structure of the firm.
13. Uncollateralized RF Model - Time Scale: In case of the uncollateralized RF models the
horizon ℎ𝑖 is the only timescale. At

𝑡 = 𝑡𝑘

the forecast forward distribution for

𝑡 = 𝑡𝑘 + ℎ𝑖

is constructed based on he given RF model, and on the filtration ℱ(𝑡𝑘 ) at 𝑡𝑘 .

14. PIT Based Transform to the [0, 1] Scale: Using the PIT transform – where 𝜙(∙) is given by
the forecast conditional distribution – the realized value of RF at

𝑡 = 𝑡𝑘 + ℎ𝑖

is mapped to a value in the [0, 1] scale.

15. Collateralized RF Model - MPoR Impact: For collateralized models the presence of the
margin period of risk (𝑀𝑃𝑜𝑅) should be additionally accounted for. The MPoR is the time

238
 required for the dealer to liquidate the collateral a given counterparty posted to finance its
exposure.
16. Collateralized RF Model - Primary Focus: Therefore, the primary focus of a collateralized
model is to describe the variation of the RF over the MPoR at any future horizon.
17. Distribution of the RF inside the MPoR: In the backtesting exercise 𝜙(∙) is the forecast RF
variation distribution in the interval [𝑡 = 𝑡𝑘 + ℎ𝑖 , 𝑡 = 𝑡𝑘 + ℎ𝑖 + 𝑀𝑃𝑜𝑅] conditional on ℱ(𝑡𝑘 )
and the realized value is the historical variation of RF in the same interval.
18. PIT Transformations on RF Realizations: The result of the PIT transformation on the
sampling sequence is a set of values ℱ(𝑟𝑡𝑘 ) in the interval [0, 1].
19. Statistical Properties of the PIT Transformation: At this stage the standard statistical tests are
applied to check for the different properties of the RF distribution.
20. Enhanced CDF Generalized Distribution Metric: In particular one can introduce a
generalized distance metric as

𝑑𝑤 = ∫[𝐹𝑛 (𝑥) − 𝐹(𝑥)]2 𝑤(𝑥)𝑑𝑥

Γ

where

Γ = [0, 1]

is the domain of 𝐹(𝑥), 𝐹𝑛 (𝑥) is the empirical cumulative distribution function CDF of the
𝐹(𝑟𝑡𝑘 ) values,

𝐹(𝑥) = 𝑥

is the CDF of the 𝑈[0, 1] distribution, and 𝑤(𝑥) is a weight function that can be chosen so as
to emphasize a given quantile domain. In Kenyon and Stamm (2012) it is suggested to link
𝑤(𝑥) to the portfolio structure.

239
21. Individual RF Realizations as Pillars: Conversely, in this chapter the risk factors are
backtested independently and the choice of 𝑤(𝑥) is unrelated to the portfolio composition.
The portfolio backtesting is a separate pillar as described above.
22. Testing across different Weight Distribution: The literature provides many different
statistical tests to check the match between the two distributions.
23. Cramer von Mises vs. Anderson Darling: This chapter considers the well-known Cramer-von
Mises – CVM with

𝑤(𝑥) = 1

and Anderson-Darling (AD)

1
𝑤(𝑥) =
𝐹(𝑥)[1 − 𝐹(𝑥)]

tests where the first focuses more on the center of the distribution whereas the latter focuses
more on the tails.
24. Distance Metric as a Single Value: It can be seen that 𝑑𝑤 is a single distance value obtained
from the realized cost at ℱ(𝑟𝑡𝑘 ).
25. Monte Carlo based p-Value Estimate: The final outcome of the backtesting should be a p-
Value. Therefore, the single realization 𝑑𝑤 is assigned a p-Value based on the construction of
the corresponding test statistic – distribution of the outcomes of 𝑑𝑤 – using Monte Carlo.
26. Generation of the Test Statistic Distribution: The test statistic for a given history, sampling
frequency, and horizon is produced by repeating on a large number of simulated paths the
back-testing calculation/mapping described above.
27. Calibrated RF Model Path Generation: The paths are generated with the same
model/calibration which is going to be used for calculating their corresponding 𝑑𝑤 ’s.
28. Horizon Realized RF Value Distribution: Therefore, the test statistic is the distribution of the
expected the correct model.

240
29. p-value from the 𝑑𝑤 Quantiles: With the computed quantile of the realized 𝑑𝑤 the p-value for
the back-testing can be derived.
30. Handling of the Overlapping Forecasting Horizons: Notice that this derivation of the statistic
allows for overlapping forecasting horizons since the auto-correlation among the 𝐹(𝑟𝑡𝑘 ) is
correctly reflected in the construction.
31. Backtesting Longer Dated Horizons: This feature is particularly useful in backtesting longer
horizons – see guidance #9, e.g., long-dated inflation/IR trades – for which the data history is
comparatively short in most of the cases.
32. Sample Test - CHFUSD Exchange Rate: Anfuso, Karyampas, and Nawroth (2017) show
uncollateralized and collateralized backtesting results for the CHFUSD exchange rate where
they apply the CVM backtest to the last 15 years of history with bi-weekly sampling
frequency for different forecasting horizons

ℎ𝑖 = {1𝑚, 3𝑚 1𝑦}

33. RFE Process Underlying the FX: The RFE is a Geometric Random Brownian Motion (GBM)
with drift

𝜇=0

and volatility calibrated using a rolling 1Y window.

34. Backtesting Collateralized/Uncollateralized Exposures: For the considered case, both the
uncollateralized and the collateralized cases

𝑀𝑃𝑜𝑅 = 2𝑤

backtesting gives acceptable results at all horizons.

35. Discriminatory Power PLUS Horizon Aggregation: The next sections analyze two further
aspects that were mentioned above – the discriminatory power of the statistical tests and the
aggregation of the backtesting results across RF’s and horizons.

241
Discriminatory Power RF Backtesting

1. Data Size Impact on Assessment: The assessment of the RF models depends crucially on the
amount of data history.
2. Model Mis-specification Sensitivity to Size: In case of availability of large data sets, the
backtesting can resolve very tiny model mis-specifications.
3. Less Data - Easier Backtesting: Conversely if the data is too few the model uncertainty will
be larger – especially for longer horizons – and it will be comparatively easier to pass
backtesting.
4. Quantitative Analysis of Data Sizes: To have a quantitative understanding of the above
described effect, Anfuso, Karyampas, and Nawroth (2017) ran backtesting analysis on the
CVM and the AD tests on synthetic data made from 1000 paths for 15 years generated by the
same stochastic model – Geometric Brownian Motion with annualize drift and volatility of

𝜇=0

and

𝜎 = 10%

5. Impact of Drift/Volatility Mis-specifications: For every one of these paths, the p-value is
determined for

ℎ𝑖 = {1𝑚, 3𝑚, 1𝑦}

and for different mis-specifications of 𝜇 and 𝜎.

242
6. Distance Metric as Sensitivity Metric: For a given ℎ𝑖 the average of the p-values across the
paths is an intuitive measure of the sensitivity of the test for the given data history w.r.t. a
mis-specified model calibration – for the correct model

〈𝑝〉 = 0.5

7. Illustration of the above Analysis: Anfuso, Karyampas, and Nawroth (2017) illustrate the
above analysis by highlighting the correctly specified model.
8. Backtesting Sensitivity to Shorter Horizons: As is evident from their tables, the backtesting
results are more sensitive at the shorter horizons because of the larger number of independent
observations taken into account:

ℎ𝑖 = 1𝑚 → 180

vs.

ℎ𝑖 = 15𝑦 → 15

independent observations.
9. AD vs. CVM Comparative Performance: It can also be noticed that AD slightly out-performs
CVM in detecting model mis-specifications for the constructed example.
10. Discriminatory Power Analysis Regulatory Reporting: In the documentation that a dealer
should provide to the regulators for the backtesting of the IMM, the discriminatory power
analysis is a useful complementary information to assess the tolerance of the backtesting
methodology.

The Aggregation of Backtesting Results

243
1. Lowest Granularity of Backtesting: The general indication from the regulators is to show the
RF backtesting results at the most granular level, i.e., single RF’s and horizons.
2. Asset Class and Horizon Aggregation: Nevertheless, the analysis should be complemented
with aggregated results that assess more holistically the performance of the exposure models
– e.g., by asset class and/or including several horizons.
3. Boot-strappable RF Probability Framework: The RF backtesting scheme presented above
allows for further aggregations within the same probabilistic framework.
4. Aggregation over multiple RF Horizons: The scheme considered here is for a single RF over
multiple horizons, i.e., the aim is to produce a single aggregated result where the importance
of the different horizons is given an arbitrary weighting function 𝜃(𝑖) with

∑ 𝜃(𝑖) = 1
𝑖

and

𝜃(𝑖) > 0 ∀ 𝑖

5. Incorporating the Dealer Portfolio Exposure: This case is of relevance in the common
situation where the same RF model is used for products/portfolios of very different maturities
and the assessment of mode performance should encompass many different time scales.
6. Standard Deviation of the Test Statistic: It can be seen that for a given sampling frequency
that the standard deviation of the test statistic distribution scales linearly with the horizon.
This is a direct consequence of the auto-correlation among back-testing experiments at
different sampling points that grows linearly with the length of the horizon.
7. Standard Deviation for GBM AD/CVM: Anfuso, Karyampas, and Nawroth (2017) illustrate
this behavior for both AD and CVM tests where the standard deviation has been determined
numerically for the case of GBM for different forecasting horizons.

244
8. Horizon Normalization of the Standard Deviation: As a consequence of the linearity, the test
normalized test distances

𝑑𝑤,𝑖
𝑑̅𝑤,𝑖 =
ℎ𝑖

with the normalization given by the forecasting horizon ℎ𝑖 - are measured in equivalent units.
9. Realized Historical Standard Deviation Metric: One can therefore define a single distance

𝑑𝑤,𝐴𝐺𝐺 = ∑ 𝜃(𝑖)𝑑̅𝑤,𝑖
𝑖

with the desired weightings across the horizons.

10. Realized Standard Deviation Metric: Therefore, the realized historical value for 𝑑𝑤,𝐴𝐺𝐺 is
derived straightforwardly from 𝑑̅𝑤,𝑖 and 𝜃(𝑖).
11. Path-wise Test Statistic Distribution: The corresponding distribution of the test statistic can
be computed by applying path by path the given test at different horizons and aggregating
𝑑𝑤,𝐴𝐺𝐺 using the equation above.
12. Equi-Weighted Horizon Discriminatory Analysis: Anfuso, Karyampas, and Nawroth (2017)
show the discriminatory power analysis – shown in the previous section – for the case of
equally weighted aggregation across horizons

1
𝜃(𝑖) =
𝑁ℎ

where 𝑁ℎ is the number of horizons considered.

Correlations Backtesting

245
1. Basel III CRD Correlation Backtesting Estimator: One of the novelties introduced by Basel
III and CRD4 for backtesting is the explicit requirement to check the correlation estimator –
see Guidance #7.
2. Capturing Bear Market Correlated Moves: This is an important ingredient of the exposure
framework, especially in periods of extreme bear markets when correlations rise significantly
(Nawroth, Anfuso, and Akesson (2014)).
3. Enhancing PIT for Correlations Testing: While the previous sections were about the
performance of single RF models, this section defines a method to measure how well a
correlated set of SDE’s for multiple RF’s describe the RF’s co-movements. As will be seen
the PIT methodology can be suitably generalized for this purpose.
𝑁(𝑁−1)
4. Backtesting Pair-wise Correlations: For a set of 𝑁 RF’s the aim is to backtest 2

correlations among all the – upper or lower – off-diagonal pairs.

5. Correlation Matrix for 𝑁 GBM’s: To explain the methodology consider the simplest model
framework, i.e., 𝑁 GBM’s with a given – e.g. historical – calibration for 𝜇⃗, 𝜎⃗, and the
correlation matrix [𝜌].
6. Synthetic Consolidated Random Factor Consolidation: Defining the process for 𝑅𝐹𝑖 as

1 2
𝑅𝐹𝑖 (𝑡) = 𝑅𝐹𝑖 (0)𝑒 𝜇𝑖 𝑡+𝜎𝑖 𝑊(𝑡)−2𝜎𝑖 𝑡

for every non-equivalent pair {𝑖, 𝑗} one can introduce the synthetic RF 𝑍𝑖𝑗 as follows:

1 1 1 1 𝜇 𝜇𝑗
(𝜎𝑖 +𝜎𝑗 )𝑡− (2+2𝜌𝑖𝑗 )−( 𝑖 + )𝑡
𝜎 𝜎 2 2 𝜎𝑖 𝜎𝑗
𝑍𝑖𝑗 (𝑡) = 𝑅𝐹𝑖 (𝑡) 𝑖 𝑅𝐹𝑗 (𝑡) 𝑗 𝑒

7. Volatility of the Synthetic Random Factor: By construction 𝑍𝑖𝑗 is a drift-less GBM with
volatility given by

246
 𝜎𝑖𝑗 = √2 + 2𝜌𝑖𝑗

8. Computing the Historical 𝑍𝑖𝑗 Realizations: The historical realizations of 𝑍𝑖𝑗 can be obtained
at every sampling point using the estimators for the marginal distributions of 𝑅𝐹𝑖 and 𝑅𝐹𝑗 -
i.e., the volatilities and the drifts 𝜎𝑖 (𝑡𝑘 ), 𝜎𝑗 (𝑡𝑘 ), 𝜇𝑖 (𝑡𝑘 ), and 𝜇𝑗 (𝑡𝑘 ), and the correlations
among 𝑅𝐹𝑖 and 𝑅𝐹𝑗 .
9. 𝑍𝑖𝑗 Volatility Measures 𝜌𝑖𝑗 : 𝑍𝑖𝑗 can be therefore backtested as was done for single RF’s, but
its volatility is a direct measure of the correlation to be verified.
10. Inadequate Marginal Distribution Estimators: It can be seen that if the estimators of the
marginal distributions are inadequate 𝑍𝑖𝑗 is likely to fail backtesting independently of 𝜌𝑖𝑗 .
11. Consequence of Inadequate Marginal Estimations: This feature is not a drawback but rather a
desirable property given the regulatory purpose of the backtesting analysis.
12. Need for Valid Marginal RF’s: The RFE models are perceived as the atomic components of
the exposure framework while the correlations are the second layer.
13. Valid 𝑍𝑖𝑗 Estimators and Invalid Marginals: Whenever the underlying RF models fail the
information on the correlation performance is of little value from a regulatory perspective.
14. SnP500 vs CHFUSD Correlation Tests: In their illustrations Anfuso, Karyampas, and
Nawroth (2017) present uncollateralized backtesting results for SnP500 index and CHFUSD
exchange rate.
15. GBM Martingale vs 1Y Rolling Volatility: They consider a correlated GBM with

𝜇=0

and the correlation 𝜌 estimated using a 1Y rolling window.

16. Negative Correlation between SnP00 and USDCHF: The correlation is mostly negative,
especially in correspondence of the lows of the SnP500.
17. Multi Horizon Correlations Backtesting: At the three-time horizons considered

ℎ𝑖 = {1𝑚, 3𝑚, 1𝑦}

247
 and for a data history of 15 years the correlation model passes backtesting at

𝐶𝐿 = 99%

- the two tickers pass RF backtesting independently for the same set of horizons. Results for
the single factors are discussed earlier.
18. Reusing Single Factor Discriminatory Power Analysis: The mapping to the single RF
problem allows the inheritance of all the results derived in that context – e.g., the
discriminatory power analysis can be obtained as for single RF as is illustrated by Anfuso,
Karyampas, and Nawroth (2017).
19. Collateralized/Uncollateralized Backtesting and Metric Aggregation: In particular the
collateralized and the uncollateralized models can both be backtested based on the single RF
factor scheme, and the correlation backtesting results can be aggregated at different horizons
using the method discussed in the previous section.
20. Aggregation across a Single Correlator: Given the large number of entries in the correlation
matrix for a given scenario, it is very convenient also to aggregate results across correlation
elements – to a given forecasting horizon.
21. Block Level Correlation Metric Aggregation: A powerful visualization of the correlation
testing is, e.g., aggregation by asset classes and the assignment of a given p-value for every
block of the correlation matrix.
22. Block Level CVM/AD PIT: To obtain such a result, the aggregated distance for a given
subset of the correlation matrix

Ω = {𝑖 ∈ 𝛼, 𝑗 ∈ 𝛽}

can be obtained by applying the chosen distribution tests (CVM or AD) to the union of all the
PIT’s of the synthetic RF’s

𝑍𝑖𝑗 ∈ Ω

248
23. Model Comparison using the Departure Metric: The corresponding test statistic distribution
can be derived with the model vs. model approach described in the previous sections,
considering the correlated paths of the RF’s in every scenario in Ω.
24. Cross RF Aggregation Technique: Anfuso, Karyampas, and Nawroth (2017) provide a
detailed illustration of such an example of the aggregation methodology.

Portfolio Backtesting

1. Backtesting the Underlying RF’s: The discussion so far has focused on the backtesting of the
underlyings.
2. Primary Focus of the Regulators: However, the primary regulatory focus is on the
performance on the overall regulatory framework, i.e., the ability of the dealer’s IMM
models to assess the RWA – i.e., the 𝐸𝐸(𝑡) profile and hence the capital – accurately or
conservatively – see Guidance #3.
3. RF vs. Regulatory Backtesting: It is obvious that the portfolio backtesting is conceptually
different from the RF’s and the correlations backtesting precisely owing to the above
statement.
4. Tolerance for Overstating the Exposure: While in the case of individual RF’s and
correlations the model that is systematically understating or overstating a certain quantile
domain is expected to fail, for portfolios and model feature that leads to systematically
overstating the 𝐸𝐸(𝑡) should not be penalized – at least from a regulatory perspective.
5. Asymmetry Adjusted PIT Capital Testing: The proposal of the third pillar of the framework –
the portfolio backtesting – is again based on the PIT, with suitable modifications that account
for the asymmetry discussed above.
6. Steps Involved in the Methodology: The methodology applies to both collateralized and
uncollateralized books, and comprises of the following steps.

249
7. Scheme Based Counterparty Identification: The identification of a set of counterparties is
based on, e.g., the B3 Default Capital RWA.
8. PIT for MM and 𝑟𝑡 :The construction of the empirical uniform with the PIT using

𝑟𝑛

𝐹(𝑟𝑛 ) = ∫ 𝜙(𝑥)𝑑𝑥
−∞

where 𝜙(∙) is given by the forecasted 𝑀𝑇𝑀 distribution for uncollateralized counterparties –
or by the ∆𝑀𝑇𝑀 distribution for the collateralized ones – and 𝑟𝑡 by the realized values for
uncollateralized counterparties – or by realized 𝑀𝑇𝑀 variation for collateralized ones.
9. Backtesting the Horizon-Appropriate Portfolio: At every sampling point

𝑡 = 𝑡𝑘

the composition of the portfolio should be the correct historical one. If no information on the
historical trade composition is available one can simply backtest the current portfolio in a
manner similar to what is discussed in the next section. The realized 𝑀𝑇𝑀 value is obtained
by re-pricing the same portfolio at

𝑡 = 𝑡𝑘 + ℎ

10. Statistically Conservative Test Portfolio #1: The statistical analysis of the 𝐹(𝑟𝑡𝑘 ) set based on
a test with a notion of conservatism embedded is referred to as the conservative portfolio
CPT.
11. Statistically Conservative Test Portfolio #2: The CPT is defined as

𝐶𝑃𝑇 = ∫[max(𝐹(𝑥) − 𝐹(𝑥𝑛 ), 0)]2 𝑤(𝑥)𝑑𝐹(𝑥)

Γ

250
 where 𝛤 and 𝐹(𝑥) are defined as for

𝑑𝑤 = ∫[𝐹𝑛 (𝑥) − 𝐹(𝑥)]2 𝑤(𝑥)𝑑𝐹(𝑥)

Γ

where 𝐹𝑛 (𝑥) is the empirical CDF from the forecasted MTM distribution and 𝑤(𝑥) is a
weight function.
12. Affirmation of the Conservative Behavior: For a given quantile the max(∙, 0) function
ensures that no test distance is accrued when the empirical uniform is more conservative than
the theoretical one.
13. Considering Strictly Positive Exposures: In the exposure language if

𝐹𝑛 (𝑥) < 𝐹(𝑥)

the contribution of the quantile 𝑥 to 𝐸𝐸 is greater or equal to the value from the exact model
– equality holds in the case of 𝑥 being negative when the contribution is zero.
14. Sample vs. Distribution Departure Characteristic: As shown in Anfuso, Karyampas, and
Nawroth (2017), in such cases as the mis-specification of the volatility of the MTM
distribution, 𝐹(𝑥) and 𝐹𝑛 (𝑥) can have multiple crossing points.
15. Low/High MTM Volatility Impact: This implies that the lower quantiles of the estimation
may be conservatively estimated while the higher values are below the correct values for low
MTM volatilities. The inverse is true for high MTM volatilities.
16. Mean of the MTM Distribution: Additionally, the EE sensitivity of the MTM volatility
depends on the mean of the MTM distribution.
17. IMT/ATM/OTM MTM Characteristic: Again, as shown in Anfuso, Karyampas, and Nawroth
(2017), in the two limiting cases of the MTM distribution – deep in and deep out of the
money, the EE is almost independent of the MTM volatility – as opposed to the more
intuitive linear dependence for an MTM distribution centered at 0.
18. Differential Ranking of the EE Quantiles: For the determination of 𝑤(𝑥) the different
quantiles should be ranked – for e.g., - their importance to the EE must be quantified.

251
19. Metric for the Quantile Distribution Importance: Their relative contribution𝛼(𝑞) can be
defined as

Φ−1 (𝑞+∆𝑞)
𝐿𝑖𝑚𝑖𝑡 1 1 1
𝛼(𝑞) = max ( ∫ 𝑥𝜙(𝑥)𝑑𝑥 , 0) ∙ = max(Φ−1 (𝑞), 0)
∆𝑞 → 0 ∆𝑞 𝐸𝐸 𝐸𝐸
Φ−1 (𝑞)

where 𝜙(𝑥) and Φ−1 (𝑥) are the probability density function and the inverse of the
cumulative density function of a reference MTM distribution.
20. Application of the Mean Value Theorem: The identity in the above equation can be derived
by applying the mean value theorem in the limit

∆𝑞 → 0

21. Normal MTM Distribution - Relative Contribution: It can be seen that in the case of a normal
MTM distribution, 𝛼(𝑞) has the following closed-form solution:

max(Φ𝒩(𝜇,𝜎) −1 (𝑞), 0) max(𝜇 + √2𝜎 2 𝐸𝐹 −1 (2𝑞 − 2), 0)

𝛼(𝑞, 𝜇, 𝜎) = = 𝜇 𝜇
𝐸𝐸𝒩(𝜇,𝜎) 𝜇Φ𝒩(0,1) (𝜎) + 𝜎ϕ𝒩(0,1) (𝜎)

where 𝐸𝐹 −1 (𝑥) is the inverse of the error function, and 𝜇 and 𝜎 are the mean and the
volatility of the Normal Distribution.
22. Same 𝜇 but varying 𝜎: Anfuso, Karyampas, and Nawroth (2017) show 𝛼(𝑞, 𝜇, 𝜎) for a set of
normal MTM distributions with equal volatility but different reasons.
23. Shape/Slope Dependence on 𝜇: The shape and the slope of 𝛼 vary dramatically with the level
of MTM.
24. Par Contract 𝛼 Volatility Dependence: It can be seen though that for

𝜇=0

252
 the standard case of collateralized portfolio 𝛼(𝑞, 𝜇, 𝜎) is independent of the volatility, i.e.

𝛼(𝑞, 𝜇 = 0, 𝜎) = 𝛼̅(𝑞)

25. Elliptical Distribution 𝜇 Scale Invariance: This additional scale invariance property is quite
robust and holds beyond the normality assumption, i.e., for all members of the elliptical
family.
26. Inadequacy of a Single 𝑤(𝑥): For real portfolios the level of MTM varies across a wide
spectrum of values and it is unlikely that a single choice of 𝑤(𝑥) can be optimal across all
cases.
27. Special Case - Fully Collateralized Portfolios: In their analysis their focus has been on an
MTM distribution centered at 0 – given also its special relevance for collateralized portfolios
– and the case

𝑤(𝑥) = 𝛼̅(𝑥)

is considered.
28. Application of the Appropriate 𝑤(𝑥): The dealer can fine-tune its own representation 𝑤(𝑥)
by looking, for e.g., to the historically realized MTM’s for the set of portfolios to be
backtested.
29. Test Statistic appropriate for Portfolio Backtesting: A final remark is in order for the
derivation of the test statistic for portfolio backtesting.
30. Multiple Times/MTM Sequence Simulation: The exact methodology – outlined for the RF’s
and correlations – would imply that every portfolio would have to compute its test statistic
using two nested Monte Carlo simulations – one for the path of the underlyings, and one to
generate the conditional MTM distribution at every sampling point and along every path.
31. Feasibility of the Dual Sequence Simulations: This approach can become computationally
very expensive and requires a sophisticated parallel implementation on a cluster for large
portfolios.

253
32. Options for avoiding Dual Simulations: If that is not feasible the following two alternative
options are available.
33. Portfolios for Proxying Key Trades: For every representative portfolio select the most
relevant trade and backtest only for the selected set – deriving the statistic exactly but for a
much smaller portfolio.
34. Portfolio for Proxying Key RF’s: Derive a representative statistic – used for all representative
portfolios – based on an archetypal portfolio with trade composition and level of auto-
correlation across sampling dates that are representative across of the set of portfolios to be
backtested.

Capital Buffer Calculation

1. Basel III Compliant Backtesting Assessment: The previous sections have shown how to run a
Basel III compliant backtesting and produce granular assessments for different IMM
components.
2. Operational/Capital Remediation to Model Under-performance: Once that diagnosis is
completed, in the case of unsatisfactory performance, the regulators expect the following
steps from the dealer:
a. A feedback loop so as to improve the models based on backtesting results
b. An intermediate remediation action to account for potential shortages of capital to
account for model deficiencies.
3. Capital Buffer Calculation – Requirement: This section discusses the fourth pillar of the
framework, i.e., the calculation of the capital buffer (CB).
4. Capital Buffer Calculation - Edge Cases: The CB should be able to efficiently interpolate the
two limits of a perfect forecasting model and its opposite, i.e., the case of a completely
inadequate estimation of the regulatory capital.

254
5. Capital Buffer Edge Case Estimates: In the former situation the CB is equal to 0, while in the
latter the sum of the estimated RWA and the buffer is bounded by regulatory capital
established with standard rules – the conservative regulatory guidances that non-IMM dealers
should follow for the calculation of the RWA.
6. Penalty for Model Mis-specification: Additionally, regulators expect the capital buffer to be
punitive, i.e.,

𝑅𝑊𝐴𝑊𝑀 × (1 + 𝐶𝐵𝑊𝑀 ) ≥ 𝑅𝑊𝐴𝑅𝑀

where WM and RM stand for wrong and right models and the CB has been defined as the
multiplicative factor be applied to the IMM 𝑅𝑊𝐴.
7. Mandatory Penalty for Model Mis-specification: The role of the above inequality is to ensure
that the dealer does not have any advantage in using the wrong models, i.e., that there is a
capital incentive in adopting an adequate exposure framework.
8. Uniqueness of the Capital Buffer Estimation: The features of the CB stated above do not
characterize it in a unique way.
9. Dependence on the Backtesting Performance: In the current construction the CB is linked to
the performance of the portfolio backtesting for a selected set of 𝑁 representative
counterparties and is calculated historically by comparing the forecasted against the realized
exposure profiles for these counterparties.
10. Interpretation as IMM Model Corrector: The final number can be interpreted as a correction
factor to be applied to the IMM capital so as to account for the mis-specifications in the
determination of the 𝐸𝐸𝑃𝐸, i.e., the IM component of the 𝑅𝑊𝐴.
11. Importance of the CP Representative Portfolios: The capital buffer is entirely based on the
representative set but is then applied to the whole portfolio. Therefore, the representative
ness of the selected counterparties is a necessary feature to obtain a meaningful capital
buffer.
12. Definition of the Model Error Metric: For a given counterparty 𝑐 the following error metric is
defined:

255
 ∆𝐸𝐸𝑐 (𝑡1 , 𝑡2 ) = max(𝑀𝑇𝑀𝑐 (𝑡2 ), 0) − 𝐸𝐸𝑐 (𝑡1 , 𝑡2 )

where max(𝑀𝑇𝑀𝑐 (𝑡2 ), 0) is the realized exposure at 𝑡2 and 𝐸𝐸𝑐 (𝑡1 , 𝑡2 ) is the EE at 𝑡2 as
forecast at 𝑡1 .
13. 1Y Average of the Error Metric: The average of ∆𝐸𝐸𝑐 (𝑡1 , 𝑡2 ) over one year is given by

∆𝐸𝐸𝑐 (𝑡1 ) = 𝔼𝑡2 ∈[𝑡1 ,𝑡1 +1𝑌] [∆𝐸𝐸𝑐 (𝑡1 , 𝑡2 )]

and is the error over the whole horizon of the profile that is relevant for the 𝐸𝐸𝑃𝐸
calculation.
14. The Capital Buffer Backtesting Metric Setup: As discussed above the capital buffer should
be dependent on the portfolio backtesting performance.
15. Error Threshold Based Weight Function: This input enters into the calculation of the capital
buffer via the following weighting function:

max(𝑝𝑐 − 𝑝𝑙 , 0)
𝒯𝑐 = min ( , 1)
𝑝𝑢 − 𝑝𝑙

where 𝑝𝑐 is the p-value of the portfolio backtesting at

ℎ = 1𝑌

- the most relevant horizon for the regulatory capital – for the counterparty 𝑐 and 𝑝𝑙 and 𝑝𝑢
are the given lower and upper thresholds, i.e., 95% and 99%.
16. 𝒯𝑐 Behavior over the p-Range: 𝒯𝑐 is 0 below 𝑝𝑙 and increases monotonically from 0 to 1 in
the interval [𝑝𝑙 , 𝑝𝑢 ].
17. Backtest Range: Satisfactory to Sufficient: 𝑝𝑢 can be seen as the failure threshold for the
portfolio backtesting while 𝑝𝑙 is the lower bound of the 𝑝-values region where the
backtesting performance may be considered unsatisfactory.

256
18. Fluctuation of Capital Buffer Estimates: In practice the capital buffer is calculated by IMM
dealers on a quarterly basis and it is a desirable feature to not have large fluctuations in its
size.
19. Implicit Smoothening inside the p-Range: Defining the relevant buffer region as an interval
instead of as a binary threshold, the value of the buffer can smoothly account for potential
deteriorations of the backtesting performance, rather than jumping suddenly to a much higher
value if, e.g., a few top counterparties cross the failure threshold.
20. Time Horizon Capital Buffer: Having introduced 𝒯𝑐 one can calculate the buffer for a given
historical 𝑡𝑖 as a p-value weighted across the 𝑁 representative portfolios:

∑𝑁
𝑐=1 𝒯𝑐 ∆𝐸𝐸𝑐 (𝑡𝑖 )
𝐾(𝑡𝑖 ) =
∑𝑁
𝑐=1 𝐸𝐸𝑃𝐸𝑐 (𝑡𝑖 )

where the sum of the forecasted 𝐸𝐸𝑃𝐸’s in the denominator makes 𝐾(𝑡𝑖 ) a unit-less estimate
for the relative error in the 𝐸𝐸𝑃𝐸 framework.
21. Averaging the Capital Buffer Threshold: As a last step 𝐾(𝑡𝑖 ) can be averaged over the
available history with either overlapping or non-overlapping samples.
22. Caveat - Include Only Positive Averages: The final result for the capital buffer is given by

𝐶𝐵 = max(𝔼𝑡𝑖 [𝐾(𝑡𝑖 )], 0)

where the max(∙, 0) function ensures that only the positive corrections apply, i.e., the
negative corrections are not applied to the 𝑅𝑊𝐴.
23. Illustration of Backtesting with Capital Buffer: Anfuso, Karyampas, and Nawroth (2017)
provide a complete series of portfolio backtesting and capital buffer calculations for synthetic
MTM paths generated with Brownian motion for a portfolio of 20 independent
counterparties. This results in several observations.
24. CPT vs. CVM or AD: First the discriminatory power of CPT is lower because less
information about the MTM distribution is processed by the test in comparison with CVM or
AD.

257
25. Finiteness of the Capital Buffer: The capital buffer is finite even in the case of correct model
specification as a consequence of the finiteness of the history and of the max(∙, 0) function in

𝐶𝐵 = max(𝔼𝑡𝑖 [𝐾(𝑡𝑖 )], 0)

26. Full Cycle Capital Buffer Estimation: A full cycle calculation of the capital buffer shows that
the inequality of

𝑅𝑊𝐴𝑊𝑀 × (1 + 𝐶𝐵𝑊𝑀 ) ≥ 𝑅𝑊𝐴𝑅𝑀

is fulfilled and therefore the buffer accounts for the missing capital due to the use of the
wrong model.
27. Incorporating the Regulatory Upper Limit: While Anfuso, Karyampas, and Nawroth (2017)
do not include the upper bound, i.e., the mandatory regulatory capital upper bound, this can
be easily fixed by imposing

𝐸𝐸𝑃𝐸 = min(𝐸𝐸𝑃𝐸𝑇𝑂𝑇𝐴𝐿 , 𝐸𝐸𝑃𝐸𝑅𝐸𝐺𝑈𝐿𝐴𝑇𝑂𝑅𝑌 )

Conclusion

1. Framework for Basel III Backtesting: This chapter includes a complete framework to
backtest IMM according to the new Basel III guidelines.
2. Component of the Backtesting Scheme: The methodology includes a diagnosis of the models
contributing to the CCR exposure – RF’s, correlations, and portfolios – and a remedy for the
potential model deficiencies impact on the regulatory capital.

258
3. Applications of the Backtesting Framework: It can be seen that the very same framework can
be used as a template for:
a. The development phase and the criteria that a model should meet prior to regulatory
submission for the IMM waiver
b. The internal one-off validation of a given model
c. The periodic – e.g., quarterly – backtesting that the dealer should provide the
regulators
4. Dealer’s Internal Model Governance Scheme: A unified approach for validation and model
backtesting is strongly endorsed by Basel III and CRD4 as it greatly simplifies the internal
governance of the dealer.
5. Pre-requisite for IMM Model Approval: Additionally, in the new regulatory environment, the
model developers should account for the backtesting requirements at the earliest stage, since
strong backtesting performance is a key pre-requisite for IMM waiver approval.
6. Closed/MC/Rule Based Schemes: From a model perspective the methodology described
above can be applied equally to Monte Carlo, historical, or rule-based CCR engines.
7. Enhancement to the Underlying Model Dynamics: The illustrations were based on GBM
given their simplicity and relevance across the industry. Nevertheless, any other model can
be backtested following the same logical steps.
8. Availability of Historical Data as a Shortcoming: In all cases the main bottleneck for a sound
backtesting is the availability of sufficient historical data for statistically significant results.

References

• Anfuso, F., D. Karyampas, and A. Nawroth (2017): A Sound Basel III Compliant Framework
for Backtesting Credit Exposure Models
• Basel Committee on Baking Supervision (2010): Sound Practices for Back-testing
Counterparty Credit Risk Models

259
• Kenyon, C., and R. Stamm (2012): Discounting, LIBOR, and Funding: Interest Rate and
Credit Pricing Palgrave Macmillan
• Nawroth, A., F. Anfuso, and F. Akesson (2014): Correlation Breakdown and the Influence of
Correlations on VaR

260
 Initial Margin Backtesting Framework

Abstract

1. Mandatory Margins for OTC Transactions: The introduction of mandatory margining for
bilateral OTC transactions is significantly affecting the derivatives market, particularly in
light of the additional funding costs that financial institutions could face.
2. Initial Margin Forecast Models Backtest: This chapter details the approach by Anfuso, Aziz,
Loukopoulos, and Giltinan (2017) for a consistent framework, applicable equally to cleared
and non-cleared portfolios, to develop and backtest forecasting models for initial margin.

Introduction

1. BCBS-IOSCO Mandatory Margining Guidelines: Since the publication of the new Basel
Committee on Banking Supervision and the International Organizations of Securities
Commissions (BCBS-IOSCO) guidance for mandatory margining for non-cleared OTC
derivatives (Basel Committee on Banking Supervision (2015)) there has been a growing
interest in the industry regarding the development of dynamic initial margin models (DIM) –
see, for example, Green Kenyon (2015), Andersen, Pykhtin, and Sokol (2017b). By DIM
model this chapter refers to any model that can be used to forecast portfolio initial margin
requirements.

261
2. Protection Afforded by BCBS-IOSCO: The business case for such a development is at least
two-fold. First, the BCBS-IOSCO IMR (B-IMR) rules are expected to protect against
potential future exposure at a high-level of confidence (99%) and will substantially affect
funding costs, XVA, and capital.
3. IM and VM Based Margining: Second, B-IMR has set a clear incentive for clearing;
extensive margining in the form of variation margin (VM) and initial margin (IM) is the main
element of the central counter-party (CCP) risk management as well.
4. IMR Impact on Bilateral + Cleared: Therefore, for both bilateral and cleared derivatives,
current and future IMR significantly affects the probability and the risk profile of a given
trade.
5. B-IMR Case Study - Performance Evaluation: This chapter considers B-IMR as a case study,
and shows how to include a suitably parsimonious DIM model on the exposure calculation. It
also proposes an end-to-end framework and also defines a methodology to backtest model
performance.
6. Organization of this Chapter: This chapter is organized as follows. First, the DIM model for
forecasting future IMR is presented. Then methodologies for two distinct levels of back-
testing analysis are presented. Finally, conclusions are drawn.

How to Construct a DIM Model

1. Applications of the DIM Model: A DIM model can be used for various purposes. In the
computation of the counter-party credit risk (CCR), capital exposure, or credit valuation
adjustment (CVA), the DIM model should forecast, in a path-by-path basis, the amount of
posted and received IM at any revaluation point.
2. Path Specific IMR Estimation: For this specific application, the key ability of the model is to
associate a realistic IMR to any simulated market scenario based on a mapping that makes
use of a set of characteristics of the path.

262
3. RFE Dependence on the DIM: The DIM model is a priori agnostic to the underlying risk
factor evolution (RFE) models to generate the exposure paths (as shall be seen, dependencies
may arise, if for example, the DIM is computed on the same paths that are generated for the
exposure).
4. Cross-Probability Measure IMR Distribution: It is a different story if the goal is to predict the
IMR distribution (IMRD) at future horizons, either in the real-world 𝑃 or the market-implied
𝑄 measures.
5. IMRD Dependence on the RFE: In this context, the key feature of the model is to associate
the right probability weight with a given IMR scenario; hence the forecast IMRD also
becomes a measure of the accuracy if the IMRD models (which ultimately determine the
likelihood of different market scenarios).
6. 𝑃 vs. 𝑄 Measure IMRD: The distinction between the two cases will become clear later on, in
the discussion of how to assess model performance.
7. ISDA SIMM BCBS IOSCO IM: The remainder of this chapter considers the BCBS-IOSCO
IM as a case study. For the B-IMR, the current industry proposal is the International Swaps
and Derivatives Association Standard Initial Margin Model (SIMM) – a static aggregation
methodology to compute the IMR based on first-order delta-vega trade sensitivities
(International Swaps and Derivatives Association (2016)).
8. Challenges with SIMM Monte Carlo: The exact replication of SIMM in a capital exposure or
an XVA Monte Carlo framework requires in-simulation portfolio sensitivities to a large set
of underlying risk factors, which is very challenging in most production implementations.
9. Andersen-Pykhtin-Sokol IM Proposal: Since the exposure simulation provides the portfolio
mark-to-market (MTM) on the default (time 𝑡) and closeout (time 𝑡 + 𝑀𝑃𝑜𝑅, where 𝑀𝑃𝑜𝑅
is the margin period of risk) grids, Andersen, Pykhtin, and Sokol (2017b) have proposed
using this information to infer path-wise the size of any percentile of the local
∆𝑀𝑇𝑀(𝑡, 𝑡 + 𝑀𝑃𝑜𝑅, 𝑃𝑎𝑡ℎ𝑖 ) distribution, based on a regression that uses the simulated
portfolio 𝑀𝑇𝑀(𝑡) as a regression variable.
10. Andersen-Pykhtin-Sokol Proposal Assumptions: The

∆𝑀𝑇𝑀(𝑡, 𝑡 + 𝑀𝑃𝑜𝑅) = 𝑀𝑇𝑀(𝑡 + 𝑀𝑃𝑜𝑅) − 𝑀𝑇𝑀(𝑡)

263
 distributed is constructed assuming that no cash flow takes place between the default and the
closeout. For a critical review of this assumption, see Andersen, Pykhtin, and Sokol (2017a).
11. Enhancing the Andersen-Pykhtin-Sokol Model: This model can be further improved by
adding more descriptive variables to the regression, e.g., values at the default time of the
selected risk factors of the portfolio.
12. Optimization: Re-using Exposure Paths: For the DIM model, the following features are
desirable. First the DIM should consume the same number of paths as the exposure
simulation, to minimize the computational burden.
13. DIM Optimization – B-IMR SIMM Reconciliation: Second, the output of the DIM model
should reconcile with the known IMR value for

𝑡=0

i.e.

𝐼𝑀(𝑃𝑎𝑡ℎ𝑖 , 0) = 𝐼𝑀𝑅𝑆𝐼𝑀𝑀 (0)

for all 𝑖.
14. Key Aspects of IOSCO/SIMM: Before proceeding, this section notes some of the key aspects
of the BCBS-IOSCO margining guidelines, and, consequently, of the ISDA SIMM Model
(International Swaps and Derivatives Association (2016)).
15. Andersen-Pykhtin-Sokol Proposal Assumptions: First, the 𝑀𝑃𝑜𝑅 for the IM calculation of a
daily margined counter-party is 10 𝐵𝐷. This may differ from the capital exposure
calculation, in which, for example

𝑀𝑃𝑜𝑅 = 20 𝐵𝐷

if the number of trades in the portfolio exceeds 5,000.

264
16. No Netting across the Asset Classes: Second, the B-IMR in the Basel Committee on Banking
Supervision (2015) prescribes calculating the IM by segregating trades from different asset
classes. Tis feature is reflected in the SIMM model design.
17. SIMM Methodology Market Volatility Independence: Finally, the SIMM methodology
consumes trade sensitivities as its only inputs and has a static calibration that is not sensitive
to market volatility.
18. Regression on the ∆𝑀𝑇𝑀 Distribution: For the IM calculation, the starting point is similar to
that of Andersen, Pykhtin, and Sokol (2017a), i.e.
a. A regression methodology based on path’s 𝑀𝑇𝑀(𝑡) is used to compute the moments
of the local ∆𝑀𝑇𝑀(𝑡, 𝑡 + 𝑀𝑃𝑜𝑅, 𝑃𝑎𝑡ℎ𝑖 ) distribution, and
b. ∆𝑀𝑇𝑀(𝑡, 𝑡 + 𝑀𝑃𝑜𝑅, 𝑃𝑎𝑡ℎ𝑖 ) is assumed to be a given probability distribution that
can be fully characterized by its first two moments – the drift and the volatility.
Additionally, since the drift is immaterial over the 𝑀𝑃𝑜𝑅 horizon, it is not computed
and set to 0.
19. Quadratic Regressor for Local Volatility: There are multiple regression schemes that can be
used to determine the local volatility 𝜎(𝑖, 𝑡) The present analysis follows the standard
American Monte Carlo literature (Longstaff and Schwartz (2001)) and uses a least-squares
method (LSM) with a polynomial basis:

𝑛
2 (𝑖, 2 (𝑖,
𝜎 𝑡) = 𝔼[∆𝑀𝑇𝑀 𝑡) | 𝑀𝑇𝑀(𝑖, 𝑡)] = ∑ 𝑎𝜎𝑘 𝑀𝑇𝑀𝑘 (𝑖, 𝑡)
𝑘=0

𝐼𝑀𝑅/𝑃,𝑈 (𝑖, 𝑡) = Φ−1 (0.99/0.01, 𝜇 = 0, 𝜎 = 𝜎(𝑖, 𝑡))

where 𝑅/𝑃 indicates received and poste, respectively. In this implementation, the 𝑛 in

𝑛
2 (𝑖, 2 (𝑖,
𝜎 𝑡) = 𝔼[∆𝑀𝑇𝑀 𝑡) | 𝑀𝑇𝑀(𝑖, 𝑡)] = ∑ 𝑎𝜎𝑘 𝑀𝑇𝑀𝑘 (𝑖, 𝑡)
𝑘=0

265
 is set equal to 2, i.e., a polynomial regression of order 2 is used.
20. Calculating the Unnormalized IM Value: The unnormalized posted and received
𝐼𝑀𝑅/𝑃,𝑈 (𝑖, 𝑡) and calculated analytically in

𝑛
2 (𝑖, 2 (𝑖,
𝜎 𝑡) = 𝔼[∆𝑀𝑇𝑀 𝑡) | 𝑀𝑇𝑀(𝑖, 𝑡)] = ∑ 𝑎𝜎𝑘 𝑀𝑇𝑀𝑘 (𝑖, 𝑡)
𝑘=0

𝐼𝑀𝑅/𝑃,𝑈 (𝑖, 𝑡) = Φ−1 (0.99/0.01, 𝜇 = 0, 𝜎 = 𝜎(𝑖, 𝑡))

by applying the inverse of the cumulative distribution Φ−1 (𝑥, 𝜇, 𝜎) to the appropriate
quantiles; Φ(𝑥, 𝜇, 𝜎) being the probability distribution that models the local
∆𝑀𝑇𝑀(𝑡, 𝑡 + 𝑀𝑃𝑜𝑅, 𝑃𝑎𝑡ℎ𝑖 ).
21. Note on the Distributional Assumptions: The precise choice of Φ does not play a crucial role,
since the difference in the quantiles among the distribution can be compensated in calibration
be applying the appropriate scaling factors (see the α𝑅⁄𝑃 (𝑡) functions below). For simplicity,
in the below Φ is assumed to be normal.
22. Comparative Performance of the LSM: It is observed that the LSM method performs well
compared to the more sophisticated kernel methods such as Nadaraya-Watson, which is used
in Andersen, Pykhtin, and Sokol (2017a), and it has the advantage of being parameter free
and cheaper from a computational stand-point.
23. Applying 𝑡 = 0, 𝑀𝑃𝑜𝑅 and SIMM Reconcilers: The next step accounts for the

𝑡=0

reconciliation as well as the mismatch between SIMM and the exposure model calibrations –
see the corresponding items above.
24. De-normalizing using IM Scaling Parameters: These issues can be tackled by scaling
𝐼𝑀𝑅/𝑃,𝑈 (𝑖, 𝑡) with suitable normalization functions α𝑅⁄𝑃 (𝑡):

266
 𝐼𝑀𝑅/𝑃 (𝑖, 𝑡) = α𝑅⁄𝑃 (𝑡) × 𝐼𝑀𝑅/𝑃,𝑈 (𝑖, 𝑡)

10 𝐵𝐷
α𝑅⁄𝑃 (𝑡) = [1 − ℎ𝑅⁄𝑃 (𝑡)] × √ × [α𝑅⁄𝑃,∞ + (α𝑅⁄𝑃,0 − α𝑅⁄𝑃,∞ )𝑒 −𝛽𝑅⁄𝑃 (𝑡)𝑡 ]
𝑀𝑃𝑜𝑅

𝑀𝑃𝑜𝑅 𝐼𝑀𝑅𝑅⁄𝑃,𝑆𝐼𝑀𝑀 (𝑡 = 0)
α𝑅⁄𝑃,0 = √ ×
10 𝐵𝐷 𝑞(0.99⁄0.01 , ∆𝑀𝑇𝑀(0, 𝑀𝑃𝑜𝑅))

25. Differential Calibration for Posted/Received IM: In

10 𝐵𝐷
α𝑅⁄𝑃 (𝑡) = [1 − ℎ𝑅⁄𝑃 (𝑡)] × √ × [α𝑅⁄𝑃,∞ + (α𝑅⁄𝑃,0 − α𝑅⁄𝑃,∞ )𝑒 −𝛽𝑅⁄𝑃 (𝑡)𝑡 ]
𝑀𝑃𝑜𝑅

𝛽𝑅⁄𝑃 (𝑡) > 0

and

ℎ𝑅⁄𝑃 (𝑡) < 1

with

ℎ𝑅⁄𝑃 (𝑡 = 0) = 0

are four functions to be calibrated – two for received and two for posted IM’s. As will
become clearer later in this chapter, the model calibration generally differs for received and
posted DIM models.
26. Scaling IM using RFE MPoR: In

𝐼𝑀𝑅/𝑃 (𝑖, 𝑡) = α𝑅⁄𝑃 (𝑡) × 𝐼𝑀𝑅/𝑃,𝑈 (𝑖, 𝑡)

267
 10 𝐵𝐷
α𝑅⁄𝑃 (𝑡) = [1 − ℎ𝑅⁄𝑃 (𝑡)] × √ × [α𝑅⁄𝑃,∞ + (α𝑅⁄𝑃,0 − α𝑅⁄𝑃,∞ )𝑒 −𝛽𝑅⁄𝑃 (𝑡)𝑡 ]
𝑀𝑃𝑜𝑅

𝑀𝑃𝑜𝑅 indicates the 𝑀𝑃𝑜𝑅 relevant for the Basel III exposure. The ratio of 𝑀𝑃𝑜𝑅 to 10 𝐵𝐷
accounts for the VM vs. IM margin period, and it is taken as a square root because the
underlying models are typically Brownian, at least for short horizons.
27. Components of the α𝑅⁄𝑃 (𝑡) Term: In

𝑀𝑃𝑜𝑅 𝐼𝑀𝑅𝑅⁄𝑃,𝑆𝐼𝑀𝑀 (𝑡 = 0)
α𝑅⁄𝑃,0 = √ ×
10 𝐵𝐷 𝑞(0.99⁄0.01 , ∆𝑀𝑇𝑀(0, 𝑀𝑃𝑜𝑅))

𝐼𝑀𝑅𝑅⁄𝑃,𝑆𝐼𝑀𝑀 (𝑡 = 0) are the 𝐼𝑀𝑅𝑅⁄𝑃 computed at

𝑡=0

using SIMM; ∆𝑀𝑇𝑀(0, 𝑀𝑃𝑜𝑅) is the distribution of the 𝑀𝑇𝑀 variations over the first
𝑀𝑃𝑜𝑅; and 𝑞(𝑥, 𝑦) is a function that gives quantile 𝑥 for the distribution 𝑦.
28. 𝑡 = 0 chosen to match SIMM: The values of the normalization functions α𝑅⁄𝑃 (𝑡) at

𝑡=0

are chosen in order to reconcile 𝐼𝑀𝑅/𝑃 (𝑖, 𝑡) with the starting SIMM IMR.
29. Mean-reverting Nature of the Volatility: The functional form of α𝑅⁄𝑃 (𝑡) at

𝑡>0

is dictated by what is empirically observed, as is illustrated by Anfuso, Aziz, Loukopoulos,

and Giltinan (2017); accurate RFE models, in both 𝑃 and 𝑄 measures, have either a volatility

268
 term structure or an underlying stochastic volatility process that accounts for the mean-
reverting behavior to the normal market conditions generally observed from extremely low or
high volatility.
30. Reconciliation with Static SIMM Methodology: Since the SIMM calibration is static
(independence of market volatility for SIMM), the

𝑡=0

reconciliation factor is not independent of the market volatility, and thus not necessarily
adequate for the long-term mean level.
31. Volatility Reducing Mean-reversion Speed: Hence, α𝑅⁄𝑃 (𝑡) is an interpolant between the

𝑡=0

scaling driven by α𝑅⁄𝑃,0 and the long-erm scaling driven by α𝑅⁄𝑃,∞ , where the functions
𝛽𝑅⁄𝑃 (𝑡) are the mean-reverting speeds.
32. Estimating α𝑅⁄𝑃,∞ from the Long-End: The values of α𝑅⁄𝑃,∞ can be inferred by a historical
analysis of a group of portfolios, or it can be ad hoc calibrated, e.g., by computing a different
∆𝑀𝑇𝑀(0, 𝑀𝑃𝑜𝑅) distribution in

𝑀𝑃𝑜𝑅 𝐼𝑀𝑅𝑅⁄𝑃,𝑆𝐼𝑀𝑀 (𝑡 = 0)
α𝑅⁄𝑃,0 = √ ×
10 𝐵𝐷 𝑞(0.99⁄0.01 , ∆𝑀𝑇𝑀(0, 𝑀𝑃𝑜𝑅))

using the long-end of the risk-factor implied volatility curves and solving the equivalent
scaling equations for α𝑅⁄𝑃,∞ .
33. Interpreting the Haircut ℎ𝑅⁄𝑃 (𝑡) Term: As will be seen below, the interpretation of ℎ𝑅⁄𝑃 (𝑡)
can vary depending on the intended application of the model.

269
34. ℎ𝑅⁄𝑃 (𝑡) for Capital/Risk Models: For capital and risk models, ℎ𝑅⁄𝑃 (𝑡) are two capital and
risk functions that can be used to reduce the number of back-testing exceptions (see below)
and ensure that the DIM model is conservatively calibrated.
35. ℎ𝑅⁄𝑃 (𝑡) for the XVA Models: For XVA pricing, ℎ𝑅⁄𝑃 (𝑡) can be fine-tuned – together with
𝛽𝑅⁄𝑃 (𝑡) - to maximize the accuracy of the forecast based on historical performance.
36. Lack of Asset Class Netting: Note that owing to the No netting across Asset Classes clause,
the 𝐼𝑀𝑅/𝑃,𝑥 (𝑖, 𝑡) can be computed on a stand-alone basis for every asset class 𝑥 defined by
SIMM (IR/FX, equity, qualified and non-qualified credit, commodity) without any additional
exposure runs. The total 𝐼𝑀𝑅/𝑃 (𝑖, 𝑡) is then given by the sum of the 𝐼𝑀𝑅/𝑃,𝑥 (𝑖, 𝑡) values.
37. Historical vs. Computed IM Calibrations: A comparison between the forecasts of the DIM
model defined in

𝑛
2 (𝑖, 2 (𝑖,
𝜎 𝑡) = 𝔼[∆𝑀𝑇𝑀 𝑡) | 𝑀𝑇𝑀(𝑖, 𝑡)] = ∑ 𝑎𝜎𝑘 𝑀𝑇𝑀𝑘 (𝑖, 𝑡)
𝑘=0

𝐼𝑀𝑅/𝑃,𝑈 (𝑖, 𝑡) = Φ−1 (0.99/0.01, 𝜇 = 0, 𝜎 = 𝜎(𝑖, 𝑡))

𝐼𝑀𝑅/𝑃 (𝑖, 𝑡) = α𝑅⁄𝑃 (𝑡) × 𝐼𝑀𝑅/𝑃,𝑈 (𝑖, 𝑡)

10 𝐵𝐷
α𝑅⁄𝑃 (𝑡) = [1 − ℎ𝑅⁄𝑃 (𝑡)] × √ × [α𝑅⁄𝑃,∞ + (α𝑅⁄𝑃,0 − α𝑅⁄𝑃,∞ )𝑒 −𝛽𝑅⁄𝑃 (𝑡)𝑡 ]
𝑀𝑃𝑜𝑅

𝑀𝑃𝑜𝑅 𝐼𝑀𝑅𝑅⁄𝑃,𝑆𝐼𝑀𝑀 (𝑡 = 0)
α𝑅⁄𝑃,0 = √ ×
10 𝐵𝐷 𝑞(0.99⁄0.01 , ∆𝑀𝑇𝑀(0, 𝑀𝑃𝑜𝑅))

and the historical IMR realizations computed with the SIMM methodology is shown in
Anfuso, Aziz, Loukopoulos, and Giltinan (2017) where alternative scaling approaches are
considered.

270
38. Criteria Utilized in the Comparison: A comparison is performed at different forecasting
horizons using 7 years of historical data, monthly sampling, and averaging among a wide
representation of single-trade portfolios for the posted and the received IM cases.
39. ℒ1 Error Metric Choice: For a given portfolio/horizon, the chosen error metric is given by
|𝐹𝑅⁄𝑃 (𝑡𝑘 +ℎ)−𝐺𝑅⁄𝑃 (𝑡𝑘 +ℎ)|
𝔼𝑡𝑘 [ ] where 𝔼𝑡𝑘 [∙] indicates an average across historical sampling
𝐺𝑅⁄𝑃 (𝑡𝑘 +ℎ)

dates – the definitions of 𝐹𝑅⁄𝑃 and 𝐺𝑅⁄𝑃 are contained below. Here and throughout this
chapter, 𝑡𝑘 is used in place of 𝑡 whenever the same quantity is computed at multiple sampling
dates.
40. Comparison of the Tested Universe: The tested universe is made up of 102 single-trade
portfolios. The products considered, always at-the-money and of different maturities, include
cross-currency swaps, IR swaps, FX options, and FX forwards – approximately 75% of the
population is made up of

∆=1

trades.
41. Calibrated Estimates of the Parameters: As is made evident by Anfuso, Aziz, Loukopoulos,
and Giltinan (2017), the proposed term structure of α𝑅⁄𝑃 (𝑡) improves the accuracy of the
forecast by a significant amount – they also provide the actual calibration used for their
analysis.
42. Conservative Calibration of the Haircut Function: Below contains further discussions on the
range of values that the haircut functions h𝑅⁄𝑃 (𝑡) are expected to take for a conservative
calibration of DIM to be used for regulatory exposure.
43. Comparison with CCP IMR: Finally, as an outlook, Anfuso, Aziz, Loukopoulos, and Giltinan
(2017) show the error metrics for the case of CCP IMR where the Dim forecasts are
compared against the Portfolio Approaches to Interest Rate Scenarios (Pairs: LCH.ClearNet)
and historical value-at-risk (HVaR; Chicago Mercantile Exchange) realizations.
44. Prototype Replications of CCP Methodologies: The realizations are based on prototype
replications of the market risk components of the CCP IM methodologies.

271
45. Universe Used for the CCP Tests: The forecasting capability of the model is tested separately
for Pairs and HVaR IMR as well as for 22 single-trade portfolios (IRS trades of different
maturities and currencies). The error at any given horizon is obtained by averaging among
22 × 2 cases.
46. Accuracy of the Proposed Scaling: Without fine tuning the calibration any further, the time-
dependent scaling α𝑅⁄𝑃 (𝑡) drives a major improvement in the accuracy of the forecasts with
respect to the alternative approaches.

How to Backtest a DIM Model

1. Assessing Model for Different Applications: The discussion so far has focused on a DIM
model for B-IMR without being too specific about how to assess the model performance for
different applications, such as CVA and margin valuation adjustment (MVA) pricing,
liquidity coverage ratio/net stable funding ratio (LCR/NSFR) monitoring (Basel Committee
on Capital Supervision (2013)), and capital exposure.
2. Estimating the IMR Distribution Accurately: As mentioned above, depending upon which
application one considers, it may or may not be important to have an accurate assessment of
the distribution of the simulated IM requirements value (IMRD).
3. Backtesting to measure DIM Performance: This chapter introduces two distinct levels of
backtesting that can measure the DIM model performance in two topical cases:
a. DIM applications that do not depend directly on the IMRD (such as capital exposure
and the CVA), and
b. DIM applications that directly depend on the IMRD (such as MVA calculation and
LCR/NSFR monitoring).
The methodologies are presented below, with a focus on the 𝑃-measure applications.

272
Backtesting DIM Mapping Functions (for Capital Exposure and CVA)

1. Review of the Monte-Carlo Framework: In a Monte-Carlo simulation framework, the

exposure is computed by determining the MTM values of a given portfolio on a large number
of forward-looking risk-factor scenarios.
2. Adequacy of Forecasts across Scenarios: To ensure that a DIM model is sound, one should
verify that the IM forecasts associated with the future simulation scenarios are adequate for a
sensible variety of forecasting horizons as well as initial and terminal market conditions.
3. Setting up a Suitable Backtesting Framework: A suitable historical backtesting framework so
as to statistically assess the performance of the model by comparing the DIM forecast with
the realistic exact IMR, e.g., in the case of B-IMR calculated according to the SIMM
methodology – for a representative sample of historical dates as well as market conditions
and portfolios.
4. Generic IMR of a Portfolio: Let us first define generic IMR of a portfolio 𝑝 as

⃗⃗⃗𝑔 = 𝑀
𝐼𝑀𝑅 = 𝑔𝑅⁄𝑃 (𝑡 = 𝑡𝛼 , 𝛱 = 𝛱(𝑝(𝑡𝛼 )), 𝑀 ⃗⃗⃗𝑔 (𝑡𝛼 ))

The terms are as follows.

5. Posted/Received IMR Computation Algorithm: The functions 𝑔𝑅 and 𝑔𝑃 represent the exact
algorithm used to compute the IMR for the posted and the received IM’s, respectively (e.g.,
such as SIMM for B-IMR, or in the case of the CCP’s, IM methodologies such as Standard
Portfolio Analysis of Risk (SPAN), Pairs, or HVaR).
6. Date of the IMR Valuation:

𝑡 = 𝑡𝛼

is the time at which the IMR portfolio 𝑝 is determined.

273
7. Portfolio Trade Population at 𝑡𝛼 : 𝛱(𝑝(𝑡𝛼 )) is the trade population of portfolio 𝑝 at time 𝑡𝛼 .
⃗⃗⃗𝑔 (𝑡𝛼 ) is a generic state variable that characterizes all of the
8. Market State Information at 𝑡𝛼 : 𝑀

𝑇 ≤ 𝑡𝛼

market information required for the computation of the IMR.

9. DIM Forecast of the Portfolio: Similarly, the DIM forecast for the future IMR of a portfolio
𝑝 can be defined as

𝐷𝐼𝑀 = 𝑓𝑅⁄𝑃 (𝑡0 = 𝑡𝑘 , 𝑡 = 𝑡𝑘 + ℎ, 𝑟⃗, ⃗⃗⃗𝐷𝐼𝑀 = 𝑀

𝛱 = 𝛱(𝑝(𝑡𝑘 )), 𝑀 ⃗⃗⃗𝐷𝐼𝑀 (𝑡𝑘 ))

The terms are as follows.

10. Posted/Received DIM Computation Algorithm: The functions 𝑓𝑅 and 𝑓𝑃 represent the DIM
forecast for the posted and the received IM’s, respectively.
11. Date of the DIM Forecast:

𝑡0 = 𝑡𝑘

is the date time at which the DIM forecast is computed.

12. Horizon of the DIM Forecast:

𝑡 = 𝑡𝑘 + ℎ

is the time for which the IMR is forecast – over a forecasting horizon

ℎ = 𝑡 − 𝑡0

274
13. Predictor Set of Market Variables: 𝑟⃗ - the predictor – is a set of market variables whose
forecasted values on a given scenario are consumed by the DIM models as input to infer the
IMR.
14. 𝑟⃗ as Simulated Portfolio MTM: The exact choice of 𝑟⃗ depends on the DIM model. For the
one considered previously, 𝑟⃗ is simply given by the simulated MTM of the portfolio.
⃗⃗⃗𝐷𝐼𝑀 (𝑡𝑘 ) is the generic state variable characterizing all the
15. Market State Information at 𝑡𝑘 : 𝑀

𝑇 ≤ 𝑡𝑘

market information required for the computation of the DIM forecast.

16. Portfolio Trade Population at 𝑡𝑘 : 𝛱(∙) is defined as before.
17. Caveats around 𝑓𝑅 and 𝑓𝑃 : Despite being computed using the stochastic RFE models, 𝑓𝑅 and
𝑓𝑃 are not probability distributions, as they do not carry any information regarding the
probability weight of a given received/posted IM value. 𝑓𝑅⁄𝑃 are instead mapping functions
between the set 𝑟⃗ chosen as predictor and the forecast value for IM.
18. Confidence Level Based DIM Calibration: In terms of 𝑔𝑅⁄𝑃 and 𝑓𝑅⁄𝑃 one can define
exception counting tests. The underlying assumption is that the DIM model is calibrated at a
given confidence level (CL); therefore, it can be tested as a 𝑉𝑎𝑅(𝐶𝐿) model.
19. Model Conservatism Linked to CL: This comes naturally in the context of real-world 𝑃
applications, such as capital exposure or liquidity monitoring, where a notion of model
conservatism, and hence of exception, is applicable, since the model will be conservative
whenever it understates (overstates) posted (received) IM.
20. The Portfolio Backtesting Algorithm Steps: For a portfolio 𝑝, a single forecasting day 𝑡𝑘 , and
a forecasting horizon ℎ, one can proceed as follows.
21. 𝑡𝑘 Estimate of the Forecast Functions: The forecast functions 𝑓𝑅⁄𝑃 computed at time 𝑡𝑘 are
⃗⃗⃗𝐷𝐼𝑀 = 𝑀
𝑓𝑅⁄𝑃 (𝑡0 = 𝑡𝑘 , 𝑡 = 𝑡𝑘 + ℎ, 𝑟⃗, 𝛱 = 𝛱(𝑝(𝑡𝑘 )), 𝑀 ⃗⃗⃗𝐷𝐼𝑀 (𝑡𝑘 )) Note that 𝑓𝑅⁄𝑃 depends

exclusively on the predictor 𝑟⃗ –

𝑟⃗ = 𝑀𝑇𝑀

275
 for the case considered above.
22. Impact of the Horizon on Predictor/Portfolio: The realized value of the predictor

𝑟⃗ = 𝑅⃗⃗

is determined. For the model considered above, 𝑅⃗⃗ is given by the portfolio value 𝑝(𝑡𝑘 + ℎ)
where the trade population 𝛱(𝑝(𝑡𝑘 + ℎ)) at 𝑡𝑘 + ℎ differs from 𝑡𝑘 only because of portfolio
aging. Aside from aging, no other portfolio adjustments are made.
23. Forecast Received/Posted IMR Estimate: The forecast values for the received and the posted
IM’s are computed as

𝐹𝑅⁄𝑃 (𝑡𝑘 + ℎ) = 𝑓𝑅⁄𝑃 (𝑡0 = 𝑡𝑘 , 𝑡 = 𝑡𝑘 + ℎ, 𝑟⃗, ⃗⃗⃗𝐷𝐼𝑀 = 𝑀

𝛱 = 𝛱(𝑝(𝑡𝑘 )), 𝑀 ⃗⃗⃗𝐷𝐼𝑀 (𝑡𝑘 ))

24. Forecast of the Received/Posted IM Estimate: The realized values for the received and the
posted IM’s are computed as

⃗⃗⃗𝑔 = 𝑀
𝐺𝑅⁄𝑃 (𝑡𝑘 + ℎ) = 𝑔𝑅⁄𝑃 (𝑡 = 𝑡𝑘 + ℎ, 𝛱 = 𝛱(𝑝(𝑡𝑘 + ℎ)), 𝑀 ⃗⃗⃗𝑔 (𝑡𝑘 + ℎ))

25. Exception Case: 𝐹/𝐺 Mismatch Conservatism: The forecast and the realized values are then
compared. The received and the posted DIM models are considered independently, and a
backtesting exception occurs whenever 𝐹𝑅 (𝐹𝑃 ) is larger (smaller) than 𝐺𝑅 (𝐺𝑃 ). As discussed
above, this definition of exception follows from the applicability of a notion of model
conservatism.
26. Detecting the Backtesting Exception History: Applying the above steps to multiple sampling
points 𝑡𝑘 one can detect back-testing exceptions for the considered history.
27. Dimensionality Reduction for the Comparison: The key step is the estimate of the
posted/received IMR forecast, where the dimensionality of the forecast is reduced – from a

276
 function to a value – making use of the realized value of the predictor, and, hence, allowing
for a comparison with the realized IMR.
28. Determining the Test 𝑝-value using TVS: The determination of the test 𝑝-value requires
additional knowledge of the Test Value Statistics (TVS), which can be derived numerically if
the forecasting horizons are overlapping (Anfuso, Karyampas, and Nawroth (2017)).
29. Caveats behind Blind TVS Usage: In the latter situation, it can happen that a single change
from one volatility regime to another may trigger multiple correlated exceptions; hence the
TVS should adjust the beck-testing assessments for the presence of false positives.
30. Accuracy of the 𝛼𝑅⁄𝑃 (𝑡) Scaling: The single trade portfolios seen earlier have been tested by
Anfuso, Aziz, Loukopoulos, and Giltinan (2017) using the SIMM DIM models with the three
choices of scaling discussed earlier. The results confirm the greater accuracy of the term
structure scaling of 𝛼𝑅⁄𝑃 (𝑡).
31. Accuracy in the Presence of Haircut: In fact, for the same level of the haircut function

ℎ𝑅⁄𝑃 (𝑡 > 0) = ±0.25

positive/negative for posted/received – a much lower number of exceptions is detected.

32. Realistic Values for the Haircut: Anfuso, Aziz, Loukopoulos, and Giltinan (2017) also
observe that, in this regard, for realistic diversified portfolios and calibration targets of

𝐶𝐿 = 95%

the functions ℎ𝑅⁄𝑃 (𝑡) take values typically in the range of 10 − 40%.
33. Assumptions Underlying the Haircut Assumption: The range of values for ℎ𝑅⁄𝑃 (𝑡) has been
calibrated using

𝛽𝑅⁄𝑃 (𝑡) = 1

and

277
 𝛼𝑅⁄𝑃,∞ (𝑡) = 1

Both assumptions are broadly consistent with historical data.

34. IOSCO results in Over-collateralization: Note also that the goal of the BCBS-IOSCO
regulations is to ensure that the netting sets are largely over-collateralized as a consequence
of:
a. The high confidence level at which the IM is computed, and
b. The separate requirements for IM and VM.
35. Impact of Over-collateralization: Hence, the exposure generating scenarios are tail events,
and the effect on capital exposure of a conservative haircut applied to the received IM is
rather limited in absolute terms.
36. Over-collateralization Impact on Exposure: This issue is demonstrated by Anfuso, Aziz,
Loukopoulos, and Giltinan (2017) where the expected exposure (𝐸𝐸) at a given horizon 𝑡 is
shown as a function of ℎ𝑅 (𝑡) – the haircut to be applied to the received IM collateral – for
different distributional assumptions on ∆𝑀𝑇𝑀(𝑡, 𝑡 + 𝑀𝑃𝑜𝑅).
37. Distribution Dependence on Haircut Functions: In particular, they compute the expected
exposure for

ℎ𝑅 (𝑡) = 0

and

ℎ𝑅 (𝑡) = 1

indicating full IM collateral benefit or no benefit at all – and take the unscaled IM as the 99𝑡ℎ
percentile of the corresponding distribution. For different classes of the ∆𝑀𝑇𝑀 distribution,
the exposure reduction is practically unaffected up to haircuts of ≈ 50%.

278
Backtesting the IMRD for MVA and LCR/NSFR

1. MC Based DIM IMR Distributions: The same Monte Carlo framework can be used in
combination with a DIM model to forecast the IMD at any future horizon – implicit here are
the models in which the DIM is not always constant across the scenarios. The applications of
the IMRD are multiple.
2. Some Applications using the IMRD: The following are two examples that apply equally to
the cases of B-IMR and CCP IMR:
a. Future IM funding costs in the 𝑃 measure, i.e., the MVA
b. Future IM funding costs in the 𝑄 measure, e.g., in relation to LCR and NSFR
regulations (Basel Committee on Banking Supervisions (2013))
3. Numerically Forecasting the IMR Distributions: The focus here is on the forecasts on the 𝑃-
measure – tackling the case of the 𝑄-measure may require a suitable generalization of
Jackson (2013). The main difference with the backtesting approach discussed above is that
the new model forecasts are the numerical distributions of the simulated IMR values.
4. Scenario-specific IM Forecasting: These can be obtained for a given horizon by associating
every simulated scenario with its corresponding IMR forecast, computed according to the
given DIM model.
5. Posted/Received IMR Density CDF: Using the notation introduced previously, the numerical
representations of the received/posted IMRD cumulative density functions (CDF’s) of a
portfolio 𝑝 for a forecasting day 𝑡𝑘 and a horizon ℎ are given by

#{𝑣 ∈ 𝕍 | 𝑣 ≤ 𝑥}
𝐶𝐷𝐹𝑅⁄𝑃 (𝑥, 𝑡𝑘 , ℎ) = ∀ 𝑟⃗𝜔 ∈ Ω
𝑁𝕍

𝕍 = {𝑓𝑅⁄𝑃 (𝑡0 = 𝑡𝑘 , 𝑡 = 𝑡𝑘 + ℎ, 𝑟⃗, ⃗⃗⃗𝐷𝐼𝑀 = 𝑀

𝛱 = 𝛱(𝑝(𝑡𝑘 )), 𝑀 ⃗⃗⃗𝐷𝐼𝑀 (𝑡𝑘 ))}

6. Terms of the CDF Expression: In

279
 #{𝑣 ∈ 𝕍 | 𝑣 ≤ 𝑥}
𝐶𝐷𝐹𝑅⁄𝑃 (𝑥, 𝑡𝑘 , ℎ) =
𝑁𝕍

𝑁𝕍 is the total number of scenarios. In

𝕍 = {𝑓𝑅⁄𝑃 (𝑡0 = 𝑡𝑘 , 𝑡 = 𝑡𝑘 + ℎ, 𝑟⃗, ⃗⃗⃗𝐷𝐼𝑀 = 𝑀

𝛱 = 𝛱(𝑝(𝑡𝑘 )), 𝑀 ⃗⃗⃗𝐷𝐼𝑀 (𝑡𝑘 )) ∀ 𝑟⃗𝜔 ∈ Ω}

𝑓𝑅⁄𝑃 are the functions computed using the DIM model, 𝑟⃗𝜔 are the scenarios for the predictor
– the portfolio MTM values in the case originally discussed, and Ω is the ensemble of 𝑟⃗𝜔
spanned by the Monte Carlo simulation.
7. Suitability of IMRD for Backtesting: The IMRD in this form is directly suited for historical
backtesting using the Probability Integral Transformation (PIT) framework (Diebold,
Gunther, and Tay (1998)).
8. Forecasting Horizon PIT Time Series: Referring to the formalism described in one can
derive the PIT time series 𝜏𝑅⁄𝑃 for a portfolio 𝑝 for a given forecasting horizon ℎ and
backtesting history ℋ𝐵𝑇 as:

⃗⃗⃗𝑔 = 𝑀
𝜏𝑅⁄𝑃 = 𝐶𝐷𝐹 (𝑔𝑅⁄𝑃 (𝑡 = 𝑡𝑘 + ℎ, 𝛱 = 𝛱(𝑝(𝑡𝑘 + ℎ)), 𝑀 ⃗⃗⃗𝑔 (𝑡𝑘 + ℎ)) , 𝑡𝑘 , ℎ) ∀ 𝑡𝑘 ∈ ℋ𝐵𝑇

9. Samples from the Actual IMR Algorithm: In the expression for 𝜏𝑅⁄𝑃 above, 𝑔𝑅⁄𝑃 is the exact
IMR algorithm for the IMR methodology that is to be forecast – defined as

⃗⃗⃗𝑔 = 𝑀
𝐼𝑀𝑅 = 𝑔𝑅⁄𝑃 (𝑡 = 𝑡𝛼 , 𝛱 = 𝛱(𝑝(𝑡𝛼 )), 𝑀 ⃗⃗⃗𝑔 (𝑡𝛼 ))

and 𝑡𝛼 are the sampling points in ℋ𝐵𝑇 .

10. Probability of 𝑡𝑘 -realized IMR: Every element of the PIT time series 𝜏𝑅⁄𝑃 corresponds to the
probability of the realized IMR at time 𝑡𝑘 + ℎ according to the DIM forecast built at 𝑡𝑘 .
11. Backtesting of the Portfolio Models - Variations: As discussed extensively in Anfuso,
Karyampas, and Nawroth (2017) one can backtest 𝜏𝑅⁄𝑃 using uniformity tests. In particular,

280
 analogous to what was shown in Anfuso, Karyampas, and Nawroth (2017) for portfolio
backtesting in the context of capital exposure models, one can use test metrics that do not
penalize conservative modeling – i.e., models overstating/understating posted/received IM.
In all cases the appropriate TVS can be derived using numerical Monte Carlo simulations.
12. Factors affecting the Backtesting: In this setup the performance of a DIM is not done in
isolation. The backtesting results will be mostly affected by the following.
13. Impact of 𝑟⃗ on Backtesting: As discussed earlier, 𝑟⃗ is the predictor used to associate an IMR
with a given scenario/valuation time point. If 𝑟⃗ is a poor indicator for the IMR, the DIM
forecast will consequently be poor.
14. Mapping of 𝑟⃗ to IMR: If the mapping model is not accurate, then the IMR associated with a
given scenario will be inaccurate. For example, the models defined in

𝑛
2 (𝑖, 2 (𝑖,
𝜎 𝑡) = 𝔼[∆𝑀𝑇𝑀 𝑡) | 𝑀𝑇𝑀(𝑖, 𝑡)] = ∑ 𝑎𝜎𝑘 𝑀𝑇𝑀𝑘 (𝑖, 𝑡)
𝑘=0

𝐼𝑀𝑅/𝑃,𝑈 (𝑖, 𝑡) = Φ−1 (0.99/0.01, 𝜇 = 0, 𝜎 = 𝜎(𝑖, 𝑡))

𝐼𝑀𝑅/𝑃 (𝑖, 𝑡) = α𝑅⁄𝑃 (𝑡) × 𝐼𝑀𝑅/𝑃,𝑈 (𝑖, 𝑡)

10 𝐵𝐷
α𝑅⁄𝑃 (𝑡) = [1 − ℎ𝑅⁄𝑃 (𝑡)] × √ × [α𝑅⁄𝑃,∞ + (α𝑅⁄𝑃,0 − α𝑅⁄𝑃,∞ )𝑒 −𝛽𝑅⁄𝑃 (𝑡)𝑡 ]
𝑀𝑃𝑜𝑅

𝑀𝑃𝑜𝑅 𝐼𝑀𝑅𝑅⁄𝑃,𝑆𝐼𝑀𝑀 (𝑡 = 0)
α𝑅⁄𝑃,0 = √ ×
10 𝐵𝐷 𝑞(0.99⁄0.01 , ∆𝑀𝑇𝑀(0, 𝑀𝑃𝑜𝑅))

include scaling functions to calibrate the calculated DIM to the observed

𝑡=0

281
 IMR. The performance of the model is therefore dependent on the robustness of this
calibration at future points in time.
15. RFE Models used for 𝑟⃗: The models ultimately determine the probability of a given IMR
scenario. It may so happen that the mapping functions 𝑓𝑅⁄𝑃 are accurate but the probabilities
for the underlying scenarios for 𝑟⃗ are misstated, and, hence, cause backtesting failures.
16. Differential Impact of Backtesting Criterion: Note that
a. The choice of 𝑟⃗, and
b. The mapping

𝑟⃗ → 𝐼𝑀𝑅

are also relevant to the backtesting methodology discussed earlier in this chapter.
RFE models used for 𝑟⃗, however, are particular to this backtesting variance, since it
concerns the probability weights of the IMRD.

Conclusion

1. Framework to Develop/Backtest DIM: This chapter has presented a complete framework to

backtest and develop DIM models. The focus has been on B-IMR and SIMM, and the
chapter has shown how to obtain forward-looking IM’s from the simulated exposure paths
using simple aggregation methods.
2. Applicability of the Proposed Model: The proposed model is suitable for both XVA pricing
and capital exposure calculations; the haircut functions in

282
 10 𝐵𝐷
α𝑅⁄𝑃 (𝑡) = [1 − ℎ𝑅⁄𝑃 (𝑡)] × √ × [α𝑅⁄𝑃,∞ + (α𝑅⁄𝑃,0 − α𝑅⁄𝑃,∞ )𝑒 −𝛽𝑅⁄𝑃 (𝑡)𝑡 ]
𝑀𝑃𝑜𝑅

can be used to either improve the accuracy (pricing) or to ensure the conservatism of the
forecast (capital).
3. CCR Capital using DIM Models: If a financial institution were to compute CCR exposure
using internal model methods (IMM), the employment of a DIM could reduce the CCR
capital significantly, even after the application of a conservative haircut.
4. Over-collateralization inherent in Basel SA-CCR: This should be compared with the
regulatory alternative SA-CCR, where the benefits from over-collateralization are largely
curbed (Anfuso and Karyampas (2015)).
5. Backtesting Methodology to Estimate Performance: As part of the proposed framework, this
chapter introduced a backtesting methodology that is able to measure model performance for
different applications of DIM.
6. Agnosticity of DIM to the Underlying IMR: The DIM model and the backtesting
methodology presented are agnostic to the underlying IMR algorithm, and they can be
applied in other contexts such as CCP IM methodologies.

References

• Andersen, L., M. Pykhtin, and A. Sokol (2017a): Re-thinking Margin Period of Risk eSSRN.
• Andersen, L., M. Pykhtin, and A. Sokol (2017b): Credit Exposure in the Presence of Initial
Margin eSSRN.
• Anfuso, C., D. Aziz, K. Loukopoulos, and P. Giltinan (2017): A Sound Modeling and
Backtesting Framework for Forecasting Initial Margin eSSRN.
• Anfuso, C., and D. Karyampas (2015): Capital Flaws Risk 27 (7) 44-47

283
• Anfuso, C., D. Karyampas, and A. Nawroth (2017): A Sound Basel III Compliant
Framework for Backtesting Credit Exposure Models eSSRN.
• Basel Committee on Banking Supervision (2013): Basel III: The Liquidity Coverage Ratio
and Liquidity Risk Monitoring Tools
• Basel Committee on Banking Supervision (2015): Margin Requirements for Non-centrally
Cleared Derivatives
• Diebold, F. X., T. A. Gunther, and A. S. Tay (1998): Evaluating Density Forecasts with
Applications to Financial Risk Management International Economic Review 39 (4) 863-883
• Green, A. D., and C. Kenyon (2015): MVA by Replication and Regression arXiV
• International Swaps and Derivatives Association (2016): ISDA SIMM Methodology
• Jackson, L. (2013): Hedge Backtesting for Model Validation Risk 25 (9) 64-67
• Longstaff, F., and E. Schwartz (2001): Valuing American Options by Simulation: A Simple
Least-Squares Approach Review of Financial Studies 14 (1) 113-147

284
 CCP and SIMM Initial Margin

Initial Margin

1. Initial Margin as Portfolio VaR: Initial margin is initiated as VaR for the derivative portfolio.
VM represents P and IM uses P&L in some holding period.
2. Bilateral IM as Parametric VaR: Bilateral IM uses the parametric VaR where the parameters
were calibrated to the historical data. Bilateral IM is called the Standard Initial Margin Model
(SIMM).
3. CCP IM using Historical VaR: On the other hand, the Central Counter-party (CCP) uses
historical VaR. CCP also demands other extra pool of assets to cover losses in the event of
multiple members’ default.
4. Bilateral CCP VM/IM Methodology:

Metric Bilateral CCP

Variation Margin MTM - -

Initial Margin VaR SIMM VaR/CVaR

Default Fund Stress VaR - Systemic Risk

5. Bilateral CCP/Holding Period Horizon: The holding period is defined as 10 days for bilateral
and 5 days for CCP.

285
6. VaR Estimation Time Period Horizon: For SIMM, the scenario considered is the recent 3
years plus the common stress period (30 Aug 2008 to 29 November 2008), and the parameter
called Risk Weight is calibrated. CCP directly uses the rate shift or the spread shift data for
the reference period. For IRS cases, LCH uses 10 years (2500 days), JSCC uses 5 years
(1250 days), and CME uses similar reference period (1260 + stress).
7. SIMM/ES VaR Confidence Level: The confidence level used for SIMM is 99% (i.e., its
equivalent calibrated level), 99.7% expected shortfall (ES) for LCH, 99% ES for JSCC, and
99.7% VaR for CME.
8. Expression for VaR/Confidence Shortfall: VaR:

ℙ[𝑉(𝑡 + 𝛥𝑡) − 𝑉(𝑡) ≤ 𝑇ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑] = 𝐶𝑜𝑛𝑓𝑖𝑑𝑒𝑛𝑐𝑒𝐿𝑒𝑣𝑒𝑙

Expected Shortfall (CVaR):

𝔼[𝑉(𝑡 + 𝛥𝑡) − 𝑉(𝑡) ≤ 𝑇ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑] = 𝐶𝑜𝑛𝑓𝑖𝑑𝑒𝑛𝑐𝑒𝐿𝑒𝑣𝑒𝑙

CCP IM

1. Regulation Mandated Clearing of Derivatives: At the September 2009 Pittsburgh summit, it

was agreed that G20 should clear certain OTC derivatives transactions at CCP’s by the end
of 2012 at the latest.
2. IRS/FX - LCH/JSCC/CME: LCH SwapClear, Japanese Securities Clearing Corporation
(JSCC), and Chicago Mercantile Exchange (CME) clear not only Interest Rate Swaps, but
also other FX products.
3. CDS: ICE/JSCC - US/EU Sovereigns: In the CDS market, ICE Clearing is dominant for
clearing US, European, and Sovereign CDS names.

286
4. Bond/Repo/Equity/Futures: Other securities such as Bond and Repo or Bond Futures and
Equity Futures are also clearable.
5. Components of the IRS IM: Base IM is ES, and Liquidity Add-On is what is embedded as
the additional factor of concentration for the DV01. Other add-ons are the CCP-specific
model costs (e.g., LCH specific).
6. CDS - ICE and JSCC:

𝐼𝑀 = 𝐵𝑎𝑠𝑒𝐼𝑀 + 𝐵𝑖𝑑𝑂𝑓𝑓𝑒𝑟 𝐴𝑑𝑑𝑂𝑛 + 𝑆ℎ𝑜𝑟𝑡𝐶ℎ𝑎𝑟𝑔𝑒 𝐴𝑑𝑑𝑂𝑛 + 𝑂𝑡ℎ𝑒𝑟 𝐴𝑑𝑑𝑂𝑛

7. Components of the CDS IM: 𝐵𝑖𝑑𝑂𝑓𝑓𝑒𝑟 𝐴𝑑𝑑𝑂𝑛 is the transaction cost on a specific name’s
CS01. Short Charge is the jump-to-default (JTD) charge for selling the position, which is
typically applied on the one shortest name only. Other Add-On includes such factor as
Recovery Rate Add-On.
8. Bond/Equity and Futures - LCH, JSCC, and CME: There is a standard method the CME
developed in 1988 called SPAN@ (Standard Portfolio Analysis of Risks). It uses basic
representative scenarios – 16 of them – and the netting among intra-month in the same
security and the security futures and the netting among the different securities are considered.

Interest Rate Swap Methodology

1. Base IM: Base IM has 1250 to 2500 scenarios in which the rate absolute/relative shift 𝑅(𝑡) is
used. The absolute shift causes the rate to behave like a normal distribution. Relative shift,
however, results in a log-normal distribution.
2. Adjusting Historical Shifts using EWMA: The relative shift was originally used in all CCP’s,
but after the onset of the negative interest rates, those shifts originally taken from historical
data were adjusted by exponentially weighted moving average (EWMA).

287
 𝑅𝑁 (𝑡) = 𝑅𝑁 + 𝑆(𝑡)

3. Calculation of Rate Shift Adjustment: Suppose that the rate shift at 𝑡 is calculated using the
LCH method below. It is adjusted by the ratio of volatility.

𝑁 = 𝑇𝑜𝑑𝑎𝑦

𝑡 = 𝑆𝑐𝑒𝑛𝑎𝑟𝑖𝑜 𝐷𝑎𝑡𝑒

𝜎𝑁 1
𝑅(𝑡) = {𝑍(𝑡 + 5) − 𝑍(𝑡) ∶ 𝑡 ∈ 𝑇} ⟹ 𝑆(𝑡) = {𝑅(𝑡) ∙ ( + 1) ∙ ∶ 𝑡 ∈ 𝑇}
𝜎𝑡 2

4. Application of EWMA Decay Factor: EWMA uses the decay factor 𝜆 to calculate the
volatility. It represents how much the older volatilities affect the next volatility.

𝜎 2 (𝑡 + 1) = 𝜆[𝜎 2 (𝑡) + (1 − 𝜆)2 𝑅 2 (𝑡 + 1)]

5. CCP Rate Shift Generation Settings:

JSCC LCH CME

IR Change Measure Absolute Absolute Relative (+4%)

FX Change Measure Relative Relative Relative

EWMA Decay Factor 0.985 0.992 0.970

Scenario 1250 2500 1260

Confidence Level ES of Worst 12 ES of Worst 6 99.70%

288
6. Rate Shift Methodology used by CME: A different approach is taken by CME, which is to
shift the rate by some offset 𝛼.

𝑍(𝑡 + 5) + 𝛼
𝑅(𝑡) = {ln ∶ 𝑡 ∈ 𝑇}
𝑍(𝑡) + 𝛼

7. Liquidity Margin: Liquidity Margin is the add-on term for the concentration position on a
specific tenor bucket 𝑏 and instrument 𝑖 - 𝑃𝑉01(𝑏, 𝑖) The correlation among the instruments
and the tenor buckets are handled by using either a correlation matrix or a similar
methodology – the function below is denoted as LM in the later sections

𝐿𝑀 = 𝐹𝑢𝑛𝑐𝑡𝑖𝑜𝑛 (𝑃𝑉01(𝑏, 𝑖))

Interest Rate Swap Calculation

1. Scenario-Specific Trade Level-IM: The base IM can be calculated with re-gridded delta
multiplied by the scenario’s 𝜔 – official CCP is based on full revaluation, but the gammas
are typically not large.
2. Scenario P&L Linear in Delta: As scenario P&L is a linear function of delta, the Base IM
numbers can be incremented in the accumulator once the scenarios are fixed.
3. Liquidity Margin Dependent on 𝑃𝑉01(𝑏, 𝑖): On the other hand, Liquidity Margin uses re-
gridded delta flows called 𝑅𝑒𝑝𝐹𝑙𝑜𝑤𝑠 with which the portfolio’s 𝑃𝑉01(𝑏, 𝑖) is calculated.
4. Portfolio Base IM/Liquidity Margin:

𝐵𝑎𝑠𝑒 𝐼𝑀(𝑃𝑜𝑟𝑡𝑓𝑜𝑙𝑖𝑜, 𝜔) = ∑ 𝐵𝑎𝑠𝑒 𝐼𝑀(𝑇𝑟𝑎𝑑𝑒𝑗 , 𝜔)

𝑗

289
 𝐿𝑖𝑞𝑢𝑖𝑑𝑖𝑡𝑦 𝑀𝑎𝑟𝑔𝑖𝑛 (𝑃𝑜𝑟𝑡𝑓𝑜𝑙𝑖𝑜) = 𝐿𝑀 (∑ 𝑅𝑒𝑝𝐹𝑙𝑜𝑤𝑠𝑗 )
𝑗

Credit Default Swap Methodology

1. Base IM: Base IM is called Spread Response Requirement in ICE’s Terminology, and it
represents the log returns of the EOD credit spread resulting from the time series analysis
from April 2007.
2. Scaling/Flattening/Steepening/Tightening/Inverting/Widening: The result of the log returns
are molded into 6 shapes, scaling/flattening/steepening/tightening or
scaling/flattening/inverting/widening scenarios.

𝑅𝑇,𝑆𝐶𝐸𝑁𝐴𝑅𝐼𝑂 = 𝑅𝑇 ∙ 𝑒 𝑅𝑖𝑠𝑘𝐹𝑎𝑐𝑡𝑜𝑟∙𝑆ℎ𝑎𝑝𝑒𝐹𝑎𝑐𝑡𝑜𝑟(𝑇)

3. Base IM as the Greatest Loss Scenario: Tightening and widening risk factors are described
for each reference credit. From those spread responses, the greatest loss scenario is chosen as
the base IM.
4. JSCC CDS Clearing IM Methodology: JSCC uses the relative shift without any EWMA
adjustment from the last 750 days.
5. Bid-Offer Margin: Bid-offer margin is the transaction cost associated with unwinding CDS
trades. It is surveyed on tenor buckets and reference credits.
6. Short Charge (JtD Charge): Short charge is the loss given default risk if the reference credit
of the CDS trades defaults at the same time as the Clearing Member’s (CM) default.
7. Notional Decomposition for Index Trades: For the index trades, the notional is decomposed
into each constituent’s amount.

290
 𝑆ℎ𝑜𝑟𝑡 𝐶ℎ𝑎𝑟𝑔𝑒 = max(𝐽𝑡𝐷𝑗 )

8. RR, Basis, and IR Margins: ICE provides other factors including Recovery Rate Risk, Basis
Risk, and IRS Risk Margins.

SIMM

1. SIMM Bilateral Initial Margin Specifications: Bilateral Initial Margin was introduced in
BCBS 226 and 261, and SIMM is used as the ISDA agreed methodology. The final rule was
published for JFSA, CFTC, USPR, and ESA in 2015 and 2016. The calculation method is
updated quiet often as new risk factors are introduced.
2. SIMM Structure: SIMM is the summation of the product IM’s, ranging from IR&FX, Credit,
Equity, and Commodity.

𝑆𝐼𝑀𝑀 = 𝑆𝐼𝑀𝑀𝑅𝐴𝑇𝐸𝑆𝐹𝑋 + 𝑆𝐼𝑀𝑀𝐶𝑅𝐸𝐷𝐼𝑇 + 𝑆𝐼𝑀𝑀𝐸𝑄𝑈𝐼𝑇𝑌 + 𝑆𝐼𝑀𝑀𝐶𝑂𝑀𝑀𝑂𝐷𝐼𝑇𝑌

For a product class IM, six risk classes are used: Interest Rate, Credit (Qualifying), Credit
(Non-qualifying), Equity, Commodity, and FX. There is a correlation matrix among the risk
factors.
3. Delta Margin:

𝐷𝑒𝑙𝑡𝑎𝑀𝑎𝑟𝑔𝑖𝑛 = √∑ 𝐾𝑏 2 + ∑ ∑ 𝛾𝑏𝑐 𝑆𝑏 𝑆𝑐 + 𝐾𝑅𝐸𝑆𝐼𝐷𝑈𝐴𝐿

𝑏 𝑏 𝑐≠𝑏

where 𝐾𝑏 is the bucket margin, 𝑆𝑏 has the sign with floored and the weighted sensitivity
(𝑊𝑆𝑘,𝑖 ). 𝑊𝑆𝑘,𝑖 is the delta multiplied by the risk weight.

291
 𝑊𝑆𝑘,𝑖 = 𝑅𝑊𝑘 𝑠𝑘,𝑖 𝐶𝑅𝑏

𝐾 = √∑ 𝑊𝑆𝑘 2 + ∑ ∑ 𝜌𝑘𝑙 𝑓𝑘𝑙 𝑊𝑆𝑘 𝑊𝑆𝑙

𝑘 𝑘 𝑙≠𝑘

where

min(𝐶𝑅𝑘 , 𝐶𝑅𝑙 )
𝑓𝑘𝑙 =
max(𝐶𝑅𝑘 , 𝐶𝑅𝑙 )

and

𝑆𝑏 = max (min {∑ 𝑊𝑆𝑖,𝑘 , 𝐾𝑏 } , −𝐾𝑏 )

𝑖,𝑘

4. Vega Margin: Vega Margin uses the similar calculation as Delta Margin. The vega risk is
computed as the product of the volatility and the vega.

𝜕𝑉𝑖𝑗
𝑉𝑅𝑖𝑘 = 𝐻𝑉𝑅𝑐 ∑ 𝜎𝑗𝑘
𝜕𝜎
𝑗

5. Curvature Margin: Curvature margin uses the gamma inferred from the vega. It can be
represented by the scaling function 𝑆𝐹(𝑡𝑗𝑘 ) multiplied by the volatility and the vega.

𝜕𝑉𝑖
𝐶𝑉𝑅𝑖𝑘 = ∑ 𝑆𝐹(𝑡𝑗𝑘 )𝜎𝑗𝑘
𝜕𝜎
𝑗

292
MVA

1. Margin Value Adjustment (MVA) Definition: The formulation of the IM cost can be cast as
the equation below:

𝑀𝑉𝐴 = ∫ 𝑓(𝑡)𝐼𝑀(𝑡)𝐷𝑓 (𝑡)𝑑𝑡

where 𝑓(𝑡) is the funding spread and 𝐼𝑀(𝑡) – the forward IM – uses forward risks where the
future delta and vega are implied from the spot IM with an approximation for the future delta
as

𝑇−𝑡
𝐷𝑒𝑙𝑡𝑎(𝑇) = 𝐷𝑒𝑙𝑡𝑎(𝑡) ×
𝑇

Vega works as

𝑇−𝑡
𝑉𝑒𝑔𝑎(𝑇) = 𝑉𝑒𝑔𝑎(𝑡) × √
𝑇

2. Funding Rate for the MVA: If the collateral rate is the repo rate, then the funding spread is
the spread between the funding rate and the repo rate.
3. Swaption: As an example, for a swaption using a normal Black Scholes, the formula can be
written as a function of the forward (𝐹), strike (𝐾), volatility (𝜎), and expiry (𝑡𝑒𝑥 ). Forward
Risk (delta):

𝐹−𝑘
Φ( ) ~ 𝐶𝑜𝑛𝑠𝑡𝑎𝑛𝑡 𝑖𝑓 𝑡 𝑤𝑜𝑢𝑙𝑑 𝑛𝑜𝑡 𝑐ℎ𝑎𝑛𝑔𝑒
√𝜎(𝑡𝑒𝑥 − 𝑡)

293
 Volatility Risk (Vega):

𝐹−𝑘 𝑡𝑒𝑥 − 𝑡
√(𝑡𝑒𝑥 − 𝑡) 𝑓 ( ) ~ 𝐶𝑜𝑛𝑠𝑡𝑎𝑛𝑡 × √
√𝜎(𝑡𝑒𝑥 − 𝑡) 𝑡𝑒𝑥

4. 5Y - 10Y 10 billion Yen ATM Payer Swaption:

𝑅𝑖𝑠𝑘𝐼𝑅𝐶𝑢𝑟𝑣𝑒 ~ 𝐸𝑥𝑒𝑟𝑐𝑖𝑠𝑒𝑃𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦 × 𝐷𝑉01 × 20

for 5Y risk weight or

𝑅𝑖𝑠𝑘𝐼𝑅𝐶𝑢𝑟𝑣𝑒 ~ 𝐸𝑥𝑒𝑟𝑐𝑖𝑠𝑒𝑃𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦 × 𝐷𝑉01 × 22

for 15Y risk weight.

𝑅𝑖𝑠𝑘𝐼𝑅𝑉𝑜𝑙 ~ 𝑉𝑒𝑔𝑎 × 𝑁𝑜𝑟𝑚𝑎𝑙 𝑉𝑜𝑙 × 21

and the forward profile is based on the above approximation.

Summary

1. CCP IM and SIMM: CCP IM uses a historical scenario in its VaR – or CVaR – calculation,
while SIMM uses parametric VaR. Other add-ons such as the Liquidity Margin are not small
in the CCP IM.

294
2. MVA Estimation: Margin Valuation Adjustment can be done by taking two factors – the
funding cost and the forward IM profile. The forward IM profile could be further enhanced
depending on the trade activity and the optimization activity.

295