Interest Rate Model Selection

A conscientious choice for mortgage investors.

Alexander Levin

ntensive developments in interest rate modeling have

I delivered a bold but confusing model selection choice

to financial engineers, risk managers, and investment
analysts. Do these modeling issues sound familiar?

• Should a mortgage bank assess interest rate risk

using the lognormal Black-Karasinski [1991]
model or the normal Hull-White [1990] model?
• Can a portfolio be hedged using different pric-
ing models for assets and derivatives?
• Is there any historical evidence that one model
is better than another?
• What does the market think about the interest
rate distribution? (It must have some idea, or how
would interest rate options be traded?)

We show that selecting the best term structure

model is becoming more of a conscientious task than a
matter of taste. Recent historical rates, the implied volatil-
ity skew for swaptions, and general volatility levels confirm
rate normalization and reject the idea of lognormality. We
propose valuing mortgages using the Hull-White [1990]
model, which can be quickly and accurately calibrated
to both the yield curve and the swaption volatility matrix.

ALEXANDER LEVIN leads

LOGNORMALITY: THE OLD DAYS
valuation model develop-
ment at Andrew Davidson & Those who read research on how interest rates per-
Co., Inc., in New York City. formed in the 1980s and the early 1990s are accustomed
[email protected] to the conjecture of lognormality. It is that interest rates

74 INTEREST RATE MODEL SELECTION WINTER 2004

are lognormally distributed; i.e., their logarithm is normally annualized deviations (right axis). The absolute histori-
distributed. The rates therefore cannot become negative, cal volatility seems to be very much independent of rates
and their randomness should be naturally and steadily in the left half of the chart (west of 10%). When the rates
measured by relative volatility. This conjecture underlies are in double-digits, the same absolute volatility measure
the validity and applicability of the Black-Scholes pricing grows with the rate level.
model to the interest rate option market. A good intro- Now, reading the historical labeled bars, we conclude
ductory treatment of the Black-Scholes model and the that the absolute volatility has become rate-independent
notion of Black volatility can be found in Hull [2000]. since the late 1980s. This conclusion is generally con-
Following Wilmott [1998], we will measure volatil- firmed by a similar analysis performed for the ten-year
ity by plotting the averaged daily increments versus the swap rate history dating back to 1989 (Exhibit 2).
rate level. We can collect all daily rate increments and store A weak or absent relation between absolute volatility
them in “buckets,” each bucket corresponding to some and rate level is a sign of normality rather than lognormality.
rate level. For example, a 7% bucket includes all daily It also prompts quoting rate uncertainty (and therefore option
increments when the rate was between 6.5% and 7.5%. prices) in terms of absolute volatility (such as 110 basis points)
After the data are collected, we average increments using rather than relative volatility (say, 20%). Recently, many bro-
the root mean square formula applied within each bucket, kers have begun communicating in exactly that way.
and then annualize them. Although the U.S. Treasury rates
are currently not the best benchmark for mortgages, they WHAT DOES THE SWAPTION MARKET THINK?
have the longest history (Exhibit 1).
In Exhibit 1, let us first disregard the bars and look Can we recover the rate distribution from the way
only at the line depicting historical volatility measured by interest rate options trade? A simple way is to measure the

EXHIBIT 1
Daily Volatility versus Level for 10-Year Treasury Rate

Observations Annualized
deviations, bp
1400 300
00-03
95-99
1200 90-94 250
85-89
80-84
1000
200

800

150

600

100
400

50
200

0 0
4 5 6 7 8 9 10 11 12 13 14
rate level, %

WINTER 2004 THE JOURNAL OF PORTFOLIO MANAGEMENT 75

EXHIBIT 2
Daily Volatility versus Level for 10-Year Swap Rate
Observations Annualized
deviations, bp
700 160
00-03
97-99 140
600
93-96
89-92 120
500

100
400
80
300
60

200
40

100 20

0 0
4 4.5 5 5.5 6 6.5 7 7.5 8 8.5 9 9.5 10
rate level, %

implied volatility skew, i.e., the dependence of the implied selected drift term), such as Black, Derman, and Toy (BDT)
Black volatility on the strike level. If market participants [1990] or Black and Karasinski (BK) [1991]. For g = 0, the
believe in lognormality, there would be little reason for absolute volatility is rate-independent and can lead to a nor-
the implied volatility to change with the option’s strike. mal model, such as Hull and White (HW) [1990]. If g =
A volatility skew testifies against lognormality by the very 0.5, we may have a popular family of square root models,
fact of its existence. such as the squared Gaussian model (SG) (see James and
To discuss a simple skew measurement method, let Webber [2000]), or the model of Cox, Ingersoll, and Ross
us first introduce a setup that generalizes many known and (CIR) [1985]. Any unnamed values for the CEV are cer-
popular single-factor models, a constant elasticity of vari- tainly possible, including negative values (hypernormality)
ance (CEV) model: and values exceeding 1 (hyperlognormality).
Blyth and Uglum [1999] propose a simple method
dr = ( Drift )dt + sr g dz (1) of recovering the most suitable CEV constant by just
looking at the observed swaption volatility skew. They
argue that, if a swap forward rate satisfies the random pro-
where r is some modeled rate, s is the volatility coeffi- cess (1), the skew should have the approximate form:
cient, and g is the CEV constant. As usual, t is time, z(t)
is the Brownian motion that disturbs the market, and the 1-g

exact specification of the drift term is not very important sK Ê Fˆ 2

(2)
ª
for our purposes.The CEV concept has no specific eco- sF Ë K¯
nomic meaning but can be viewed as a convenient way
to generalize and compare all known popular models.1 where sK and sF are the Black (i.e., proportional) volatil-
For g = 1, the absolute volatility is proportional to ities for the actual strike (K) and the at-the-money strike
the rate, and we have a lognormal model (with a properly (F), F is today’s forward rate, and K is the swaption strike.
76 INTEREST RATE MODEL SELECTION WINTER 2004
EXHIBIT 3
Implied Volatility Skew on 5-Year-into-10-Year Swap

18

Lognormal (CEV = 1)
17
Square-root (CEV = 0.5)
Normal (CEV = 0)
16 Optimal fit (CEV = 0.23)
Actual
Volatility

15

14

13

12
-250 -200 -150 -100 -50 ATM 50 100 150 200 250
Strike
*Source for actual volatility history: Bank of America; volatility for 200 ITM/OTM was not quoted.

Let us analyze the same set of CEV special values, options are traded with a square root volatility. This phe-
0, 1.0, and 0.5. If g = 1.0, there will be no skew at all: nomenon may be a combination of a slight theoretical
sK ∫ sF for any strike K. This is the Black-Scholes case. smile and the broker commission demand.
For g = 0, the skew has a functional form of inverse Exhibit 4 illustrates historical month-by-month
square root. For g = 0.5, it will have the shape of an inverse skew, suggesting that the normalization effect (g ª 0) has
fourth-degree root. It is worth mentioning here that each been observed since the beginning of 2001.
inverse root function is a convex one, so the theoretical This CEV analysis unambiguously rejects lognor-
skew should not be deemed a straight line (except when mality and reveals a more suitable model. Although the
g = 1). In fact, it should not be confused with a more best-fit CEV constant varies somewhat, any volatility
aggressive convex volatility smile that may or may not be model between the normal one and the square root seems
present in addition to the skew.2 to be a decent choice. Because of its analytical tractabil-
The object of our study—the five-year into ten-year ity and the recent CEV trend, we focus on the HW
swaption (5-into-10)—was selected with modeling volatil- model as the chief alternative to the BDT or BK models.
ity of mortgage rates and valuation of the prepayment
option in mind. Exhibit 3 depicts five skew lines plotted VOLATILITY INDEX
for three named CEVs, the actual volatility observations
averaged from January 1998 through May 2002, and the It is useful to design a market volatility index—a sin-
optimal fit line (g = 0.23) for the same period. The best gle number reflecting the overall level of swaption volatil-
CEV is therefore found to be generally between the nor- ity deemed relevant to the mortgage market. This gives us
mal case (HW model) and the square root case (SG or CIR a convenient way to communicate with practitioners and
models). It is also seen that low-struck options are traded compare models; it can also serve as a risk assessment tool.
with a close-to-normal volatility, while high-struck We designate a family of at-the-money swaptions

WINTER 2004 THE JOURNAL OF PORTFOLIO MANAGEMENT 77

and, assuming no mean reversion, optimize for the sin- market; the volatility index constructed for the Hull-
gle short rate volatility constant s (volatility index) best White model has gradually become the most stable one.
matching the swaptions’ volatility surface, on average. For example, the swap rate plunged a good 60% between
This volatility index is model-specific; unlike some other January 2000 and June 2003, but the absolute volatility
volatility indexes (such as the Lehman Brothers indexes), index barely changed. The two other models have pro-
it is not a simple average of swaption volatilities. The inter- duced volatility indexes that follow the rate level but in
nal analytics of each model (exact or approximate) are used the opposite direction (the lognormal model does by far
to translate the short rate volatility constant into swaption the worst job).
volatilities used for calibration. Interestingly enough, the squared Gaussian index was
Note that this constant-volatility zero mean rever- stable for most of 2003 and could handle the rate plunge
sion framework is used only to derive the volatility index. to a new historical low (2.9%) in June 2003. This con-
It is not recommended for actual option-adjusted spread firms that a square root volatility functional pattern may
valuation, where we strongly prefer the accuracy gained outperform others when rates are very low.
by optimizing a time-dependent volatility function s(t) These findings are consistent with the swaption skew
and a mean reversion constant. measures we have discussed. This is not a coincidence at
Exhibit 5 depicts the history of three volatility all. People who set the market for ATM swaptions are the
indexes (sigmas) computed from the beginning of 2000 same ones who trade out-of- and in-the-money options.
for the Hull-White normal model, the Black-Karasinski
lognormal model, and the squared Gaussian model. Each OTHER PROBLEMS WITH LOGNORMALITY
index is calibrated to the same family of equally weighted
ATM swaptions deemed relevant to the mortgage mar- Although the Black, Derman, and Toy [1990] and
ket: options on the two-year and the ten-year underly- the Black and Karasinski [1991] models have been the
ing swaps with expirations ranging from six months to ten bread and butter of option traders since they were devel-
years. We add for comparison a line for the seven-year rate oped, a full-scale implementation required for good mort-
level, and scale all four lines so that they start at 1.0. gage analytics is not a simple task.
Exhibit 5 confirms a spectacular normalization of the Short rate lognormal models are not analytically

EXHIBIT 4
Historical CEV Values

1 lognormality level
0.9
0.8
0.7
0.6
square-root level
0.5
0.4
0.3
0.2
0.1
0 normality level

-0.1
-0.2
Jan-98 Jul-98 Jan-99 Jul-99 Jan-00 Jul-00 Jan-01 Jul-01 Jan-02

78 INTEREST RATE MODEL SELECTION WINTER 2004

EXHIBIT 5
What Volatility Index is Most Stable?

2.5

7-yr swap
BK sigma
HW sigma
2
SqG sigma

1.5

0.5

synchronized start

0
0

3
0

3
0

3
00

01

02

03
-0

-0

-0

-0
-0

-0

-0

-0
l-0

l-0

l-0

l-0
n-

n-

n-

n-
pr

pr

pr

pr
ct

ct

ct

ct
Ju

Ju

Ju

Ju
Ja

Ja

Ja

Ja
O

O
A

tractable. For example, a Monte Carlo simulation, or any who think a 20% short rate volatility plugged into the
other forward sampling method employed as the primary model results in a 20% swaption volatility. Therefore,
mortgage pricing tool will simulate only the short rate pro- model calibration to the mortgage-relevant options (not
cess on its own. Analytics that would map this process into the options on the short rate) can be complicated.
long rate dynamics (a mortgage rate) simply do not exist
and need time-consuming numerical replacements. THE HULL-WHITE MODEL: AN OVERVIEW
Relative (Black) volatility is used for quotation
only; it is merely a price-volatility conversion tool, not The short rate in the HW [1990] model is driven
requiring any acceptance of the Black-Scholes model. As by a linear stochastic differential equation, which is a spe-
we have seen, volatility changes drastically with the level cial case of the CEV Equation (1):
of rates, and therefore with the expiration of traded
options. One constant number cannot describe the entire dr = a(t )(q (t ) - r )dt + s (t )dz (3)
universe of swaptions deemed relevant for mortgage pric-
ing. The BK model with constant volatility cannot be rec- where a(t) denotes mean reversion, and s(t) stands for volatil-
ommended, in view of a steep yield curve and a sharply ity; both can be time-dependent. Function q(t) is sometimes
inverse proportional volatility term structure. referred to as arbitrage-free drift. That is, by selecting a
Contrary to common belief, long rates in the BDT proper q(t), we can match any observed yield curve.
and the BK models are not lognormal, and generally are Since (3) is a linear differential equation disturbed
less volatile than short rates, even in the absence of mean by normally distributed Brownian motion, its output,
reversion. This may be an unpleasant surprise for those the short rate process, will also be normally distributed.

WINTER 2004 THE JOURNAL OF PORTFOLIO MANAGEMENT 79

Negative rates are not precluded. Although this fact is well along the curve. This feature agrees with the behavior of
known (but never met with enthusiasm among practi- absolute implied volatility for traded swaptions; it gener-
tioners), there are many advantages in the model that ally falls with the swap maturity. This observation there-
make up for this drawback. fore helps us calibrate mean reversion in the model.
The model is analytically tractable. For example, the If a = 0, function BT becomes identical to 1.0,
arbitrage-free function q(t) is expressed analytically through regardless of maturity T. This important special case,
a given forward curve. The average zero-coupon rates and called the Ho-Lee [1986] model, allows for a pure paral-
their standard deviations are also known for any maturity lel change in the entire zero-coupon curve (every point
and any forward time. (Many derivations of the HW and moves by the same amount). Such an opportunity can be
other Gaussian models can be found in Levin [1998].) advantageous for standardized risk measurement tests. No
Any long zero-coupon rate rT of arbitrary maturity other model allows parallel shocks to be mathematically
T is proven also to be normally distributed and linear in the consistent with its internal analytics.
short rate; volatilities are related as
Calibration to ATM Swaptions
Long-Rate Volatility 1 - e - aT
= ∫ BT (4) Because the standard deviation of any zero-coupon
Short-Rate Volatility aT
rate can be found explicitly for any bond maturity and any
at any time t. forward time, it can be directly compared with quoted
Function BT of maturity T plays an important role Black volatility. Although market swaps are coupon-bear-
in the HW model. It allows for calibrating the volatility ing instruments, zero-coupon volatility analysis remains
function s(t) to the option market. If mean reversion is quite accurate within the maturity range deemed relevant
positive, then BT < 1, and the model allows for quasi- for the mortgage market (up to ten years). Whether the
parallel shocks, with rate deviations gradually depressed model operates with a time-dependent volatility function

EXHIBIT 6
Calibrated Volatility Term Structure—May 2002

150
vol on 2-yr swap
140 vol on 10-yr swap
the short-rate caplet vol
calibrated s(t)
130

120
absolute quote, bp

110

100

90

80

70

60
0 12 24 36 48 60 72 84 96 108 120
expiry, months

80 INTEREST RATE MODEL SELECTION WINTER 2004

EXHIBIT 7
Comparative Calibration for the HW Model—5/13/2002

Volatility, bp (best fit) Mean Reversion (%) Accuracy (bp)*

123.1 (constant) Set to zero 12.4

145.9 (constant) 3.36 (best fit) 6.8

Time-Dependent Set to zero 7.8

Time-Dependent 2.05 (best fit) 3.6

* Effective RMSE measured across the volatility matrix.

s(t) or a constant volatility parameter, s(t) ∫ s = const, it Although many market participants perceive that
can be optimized to approximate the volatility matrix of caps and swaptions trade in unison, they may overlook an
traded ATM swaptions. This calibration procedure important modeling difference—the jumps. Long swaps
includes finding the best mean-reversion parameter a in are chiefly diffusive, and a model disturbed by a Brown-
the sense discussed above. ian motion [like z(t) in Equation (1)] makes sense. Short
Exhibit 6 plots the calibration results using a series rates combine continuous diffusion (small day-after-day
of ATM options on the two-year swap and on the ten- changes) with sudden regulatory corrections.
year swap as the input. The bars for six different expira- It is mathematically not very difficult to add jumps
tions show known volatility quotes converted into the to diffusion (Merton did it in 1976), but the equivalent
absolute form (i.e., the relative quote multiplied by the volatility term structure will become hump-shaped. Under
forward rate).3 a jump or a jump-diffusion disturbance, the short-dated
The overall effective error of calibration is just 3.6 Black volatilities come up considerably suppressed.
basis points of absolute volatility, as measured across the Exhibit 9 compares market volatilities for traded
swaption matrix. The mean-reversion parameter was caps (solid bars) with volatilities of caps derived from the
restricted to be a constant; its best value is found as a = swaption-fitted HW [1990] model. The model drastically
2.05%. For some particular applications (like standardized overstates short-dated cap volatilities, in both absolute and
risk tests), one may prefer a zero mean reversion, or a con- relative terms. As the cap’s maturity extends, swaptions and
stant volatility parameter. These restrictions generally caps seem to converge. Can the cap (rather than the
reduce the calibration accuracy as shown in Exhibit 7. swaption) volatility structure be plugged into a mortgage
Calibrated s(t) in Exhibit 6 is rather responsive to the pricing system? Perhaps so, if the system’s interest rate
slope and the shape of the input volatility structure. It falls model maintains the jump-diffusion setting. As develop-
sharply beyond the six-year horizon, perhaps as a result of ers of mortgage analytical systems traditionally do not do
market perception about current versus long-term volatil- this, the blind use of caps will understate volatility and
ity. During the stormy market of the first half of 2002, therefore the prepayment option value.4
short-dated options were indeed traded at unprecedented We prefer using the swaption market for bench-
absolute volatility levels of 125-140 basis points, well above marking volatility, especially for fixed-rate mortgages.
their long-term averages. This was not the case in the calm Valuation of adjustable-rate mortgages may need additional
August of 1998, just prior to the Russian crisis (Exhibit 8). attention in view of embedded reset caps.

Issues Related to Caps MBS VALUATION AND

RISK MANAGEMENT IMPLICATIONS
We have demonstrated that the HW model can be
calibrated to traded swaptions. Would it be even easier to Let us assume that the normal HW [1990] model
use caps? After all, the function we seek, s(t), is the short and the lognormal BK [1991] model are independently
rate volatility function, and derivatives on short rates calibrated to the ATM swaptions. They should value
(LIBOR) seem to be good candidates to examine volatility. ATM swaptions identically, but the volatility skew of the

WINTER 2004 THE JOURNAL OF PORTFOLIO MANAGEMENT 81

EXHIBIT 8
Calibrated Volatility Term Structure—August 1998
100
vol on 2-yr swap
vol on 10-yr swap
95
the short-rate caplet vol
calibrated st)
90

85
absolute quote, bp

80

75

70

65

60
0 12 24 36 48 60 72 84 96 108 120
expiry, months

normal model (curve g = 0 in Exhibit 3) is quite unlike under the three different term structure models. In each
the flat one for the lognormal model (g = 1). This means case the short rate volatility function is calibrated to ATM
the two models will value any option other than those swaptions.
employed for calibration differently. As one would expect, cuspy mortgages located at the
Since embedded mortgage options (prepayment, center of refinancing curve (“at-the-money” FNCL7)
ARM caps and floors, clean-up calls) are spread over time are valued in a very close OAS range by all three mod-
and instruments, changing from the BK model to the HW els. When the prepayment option is out of the money (the
model will generally result in a change of values. Even discount sector), this option will be triggered in a falling
more important, the interest rate sensitivity measures will rate environment. This sector therefore looks relatively rich
change visibly—as a direct result of different volatility spec- under the HW model, while the premium sector bene-
ifications. Under the BK model, every up move in rates fits from using this model. As one could expect, the SG
proportionally inflates the absolute volatility, thereby model produces valuation results that are between results
reducing the modeled value of the mortgage-backed of the HW and the BK models.
security. This can be considered an indirect (via volatil- Although most mortgage instruments will look
ity) interest rate effect, artificially extending the effective shorter under the HW model, there are some notable
duration of mortgages. exceptions. As we point out above, the primary divergence
Exhibit 10 shows a 0.4-year duration reduction of HW from BK is found in differing volatility models.
when moving from the BK model to the HW model, for Since mortgage interest-only (IO) classes and mortgage
the current-coupon agency MBS. This may considerably servicing rights (MSR) have drastically changing convexity
requantify the delta-hedging needs in secondary market- profiles, they will also have unsteady exposures to volatil-
ing and MBS portfolio management. ity, i.e., vega.
The table in Exhibit 10 summarizes comparative val- For example, vega is typically positive for an IO
uation results for 30-year fixed-rate agencies obtained taken from a premium pool (case 1), negative for that
82 INTEREST RATE MODEL SELECTION WINTER 2004
EXHIBIT 9
Implied Cap Pricing—May 2002

60 160

140
50

120

40

absolute volatililty, bp
100
relative volatility, %

Market quotes
Swaptions-fitted HW
30 80

60
20

40

10
20

0 0
12 24 36 60 84 120
maturity, months

stripped off a discount pool (case 2), and about zero when agement: one for the assets, and another for the hedge?
the pool’s rate is at the center of the refinancing curve Suppose a mortgage desk uses the BK model, while the
(cuspy premium, case 3). Therefore, the BK model will swaps desk trades with a skew. Unless the position is made
generally overstate the rate sensitivity for case 1, under- vega-neutral, differing volatility specifications in the mod-
state it for case 2, and be close to the HW model in case els may considerably reduce hedge efficiency.
3 (see Exhibit 11). Keep in mind that an IO value, con-
trary to a regular MBS, increases with rates. Curiously NEGATIVE RATES
enough, the value of an IO stripped off the current-
coupon pool is always found to be higher under the HW Knowing that interest rates have never been nega-
model than under the BK model, for all rate moves. tive in U.S. history, we should question what detrimen-
These interesting findings, although affecting valu- tal effects might occur upon use of the Hull-White [1990]
ation and delta-hedging, do not contradict what is well model. Some have tried to estimate the probability of get-
known; MBS stripped derivatives and MSRs are influ- ting negative rates in the model; this approach typically
enced greatly by prepayments and slightly by interest rate creates needless concern. Indeed, the odds of such an event
models, provided that models are calibrated to the same are far from infinitesimal, but how badly can that dam-
set of volatility benchmarks. The latter constraint is crit- age the value of an MBS?
ical. As Exhibit 11 shows, the static (zero-volatility) val- Answering this question may be simpler than it
uation profile differs considerably from the option-adjusted sounds. Consider a LIBOR floor struck at zero. This
one. Buetow, Hanke, and Fabozzi [2001] provide a good non-existent derivative will have a sure zero practical
reminder to practitioners who may underestimate the value, but not under the HW model. We have priced such
importance of model calibration. a hypothetical instrument using assumptions that inflate
Can two different rate models be used for risk man- its value: market as of May 2002 (low rates, high volatil-

WINTER 2004 THE JOURNAL OF PORTFOLIO MANAGEMENT 83

EXHIBIT 10
LIBOR OAS and Duration Profiles for New FNCLs—May 13, 2002

Libor OAS, bps Duration, yrs

40 7
Black-Karasinski

35 Hull-White
6
Squared Gaussian

30
5

25

Duration 4

20

3
15

Libor OAS
2
10

1
5

0 0
FNCL5 FNCL5.5 FNCL6 FNCL6.5 (CC) FNCL7 FNCL7.5 FNCL8 FNCL8.5

LIBOR OAS Effective Duration

HW SG BK HW SG BK
GNSF5 10.5 15.7 19.6 6.13 6.38 6.66
GNSF5.5 15.3 20.1 24.0 5.75 6.00 6.28
GNSF6 4.3 8.0 11.1 5.15 5.41 5.68
GNSF6.5 (CC) 6.4 8.4 10.0 4.28 4.50 4.74
GNSF7 5.9 5.7 5.1 3.20 3.38 3.56
GNSF7.5 7.0 5.1 2.8 2.29 2.43 2.55
GNSF8 21.6 18.2 14.6 2.02 2.14 2.22
GNSF8.5 28.4 24.0 19.6 1.77 1.87 1.92
FNCL5 20.6 25.2 28.7 5.82 6.06 6.30
FNCL5.5 25.3 29.5 33.1 5.47 5.70 5.94
FNCL6 16.4 19.7 22.6 4.86 5.08 5.31
FNCL6.5 (CC) 14.5 16.2 17.6 4.04 4.24 4.45
FNCL7 7.9 7.5 6.9 2.80 2.95 3.13
FNCL7.5 9.7 7.4 4.8 2.13 2.25 2.37
FNCL8 22.2 18.9 15.5 1.89 1.98 2.07
FNCL8.5 15.4 11.0 6.6 1.60 1.68 1.73

84 INTEREST RATE MODEL SELECTION WINTER 2004

EXHIBIT 11
Valuation Results for IO on Current-Coupon FNCL Pool
Duration,
Valuation Multiple Convexity
7 150

6
Multiple
100

5
Black-Karasinski
Convexity 50
4 Hull-White
Static Valuation

3
0

-50
1
Duration

0 Vega > 0 Vega < 0 -100

dn 200 dn 150 dn 100 dn 50 base up 50 up 100 up 150 up 200
premiums rate moves discounts

ity), with model mean reversion set to zero. The volatil- and from the insightful comments of James Barrett, Robert Lan-
ity function s(t) is calibrated and extrapolated beyond the dauer, and Yung Lim. The volatility skew data used in the anal-
ten-year expiration. ysis are courtesy of Craig Lindemann and Krystn Paternostro. The
The value of a zero-struck floor is found to be author thanks Jay DeLong and Steven Heller who incorporated
insignificant for the average life range relevant to mort- the new library of interest rate models into the OAS products.
Initial AD&Co. internal publication would not have been pos-
gage pricing (up to ten years). Thus, for the ten-year non-
sible without the diligent production efforts of Ilda Pozhegu.
amortizing floor, the value is 7 basis points, which is All interest rate models discussed in the paper are included
equivalent to a 0.8 basis point error in the OAS. The error in AD&Co’s VectorTM suite of analytical models. They can
grows with the horizon; the 30-year floor is priced at 35 operate with time-dependent or constant volatility calibrated to
basis points, which would lead to 2.5 basis points of spu- an arbitrary family of ATM swaptions.
rious OAS. 1
The CEV model is analyzed in many published sources,
We can conclude that the Hull-White [1990] model including Wilmott [1998] and James and Webber [2000].
2
is rather harmless. It will not lead to sizable mispricing even This smile can be explained by the jumpy nature of the
in the worst mortgage-irrelevant case. This conclusion, underlying rate. For example, LIBOR caps are traded with a
however, certainly merits periodic review. considerable volatility smile because these rates are subject to
regulatory interference. Swap rates are much more diffusive than
jumpy, so swaption smiles are much less pronounced.
ENDNOTES 3
This method of conversion is more accurate in match-
ing option values under the lognormal and normal versions of
This article comes from several Andrew Davidson Co.
Black-Scholes than the formal variance match.
publications. It has greatly benefited from joint research with 4
A recent complex three-factor jump-diffusion develop-
Andrew Davidson (many views are as much his as the author’s)
ment by Chan et al. [2003] is a rare exception.

WINTER 2004 THE JOURNAL OF PORTFOLIO MANAGEMENT 85

REFERENCES Hull, J. Options Futures, and other Derivative Securities, 4th ed.
Englewood Cliffs, NJ: Prentice-Hall, 2002.
Black, F., E. Derman, and W. Toy. “A One-Factor Model of
Interest Rates and its Application to Treasury Bond Options.” Hull, J., and A. White. “Pricing Interest Rate Derivative Secu-
Financial Analysts Journal, January/February 1990, pp. 33-39. rities.” Review of Financial Studies, 3 (1990), pp. 573-592.

Black, F., and P. Karasinski. “Bond and Option Pricing when James, J., and N. Webber. Interest Rate Modeling: Financial Engi-
Short-Term Rates are Lognormal.” Financial Analysts Journal, neering. New York: John Wiley & Sons, 2000.
July/August 1991, pp. 52-59.
Levin, A. “Deriving Closed-Form Solutions for Gaussian Pric-
Blyth, S., and J. Uglum. “Rates of Skew.” Risk, July 1999, pp. ing Models: A Systematic Time-Domain Approach.” Interna-
61-63. tional Journal of Theoretical and Applied Finance, 1 (1998), pp.
349-376.
Buetow, G., B. Hanke, and F. Fabozzi. “Impact of Different
Interest Rate Models on Bond Value Measures.” The Journal Merton, R. “Option Pricing when the Underlying Stock
of Fixed Income, December 2001, pp. 41-53. Returns are Discontinuous.” Journal of Financial Economics, 5
(1976), pp. 125-144.
Chan, Y. K., R. Bhattacharjee, R. Russell, and M. Teytel. “A
New Term-Structure Model Based on the Federal Funds Tar- Wilmott, P. Derivatives. New York: John Wiley & Sons, 1998.
get.” Mortgage Securities, Citigroup, May 2003.

Cox, J.C., J.E. Ingersoll, and S.A. Ross. “A Theory of the Term To order reprints of this article, please contact Ajani Malik at
Structure of Interest Rates.” Econometrica, 53 (March 1985), pp. [email protected] or 212-224-3205.
385-408.

Ho, T.S.Y., and S.-B. Lee. “Term Structure Movements and Reprinted with permission from the Winter 2004 issue of The
Pricing Interest Rate Contingent Claims.” Journal of Finance, 41 Journal of Portfolio Management. Copyright 2004 by Institutional
(December 1986), pp. 1011-1129. Investor Journals, Inc. All rights reserved. For more information
call (212) 224-3066. Visit our website at www.iijournals.com.

86 INTEREST RATE MODEL SELECTION WINTER 2004