The credit spread curve

I: Fundamental concepts, fitting, par-adjusted spread, and

expected return

Richard J. Martin∗
arXiv:2201.01330v1 [q-fin.PR] 4 Jan 2022

January 6, 2022

Abstract
The notion of a credit spread curve is fundamental in fixed income investing, but in practice
it is not ‘given’ and needs to be constructed from bond prices either for a particular issuer, or for
a sector rating-by-rating. Rather than attempting to fit spreads—and as we discuss here, the
Z-spread is unsuitable—we fit parametrised survival curves. By deriving a valuation formula
for a risky bond, we explain and avoid the problem that bonds with a high dollar price trade
at a higher yield or spread than those with low dollar price (at the same maturity point), even
though they do not necessarily offer better value. In fact, a concise treatment of this effect
is elusive, and much of the academic literature on risky bond pricing, including a well-known
paper by Duffie and Singleton (1997), is fundamentally incorrect.
We then proceed to show how to calculate carry, rolldown and relative value for bonds/CDS.
Also, once curve construction has been programmed and automated we can run it historically
and assess the way a curve has moved over time. This provides the necessary grounding for
econometric and arbitrage-free models of curve dynamics, which will be pursued in later work,
as well as assessing how the perceived relative value of a particular instrument varies over time.

Introduction
Credit investors need to answer a variety of questions about the bond markets in which they
operate, such as:

A1 How has a particular subset of the universe, e.g. BBB miners in the 5–10y maturity bucket,
traded over the last few years? Where is it now relative to history?

A2 Is the BBB mining curve flat or steep today by comparison with history?

A3 How have A/BBB vs BB/B miners traded in the last few years? The effect of investment
grade (IG) and high yield (HY) moving oppositely is often called compression/decompression.

A4 What is the influence of country spread on corporate spread, sector by sector (cf. [7])?

B1 How has the credit spread of a particular issuer varied over an extended period of time (several
years)?

B2 Has the curve of a particular issuer flattened or steepened of late?

∗
Dept. of Mathematics, Imperial College London, South Kensington, London SW1 2AZ, UK. Email:
[email protected]

1
 C How do we determine carry, rolldown and relative value of a bond? These are essential
ingredients in understanding how to evaluate expected return.

Questions marked ‘A’ are at broad sector-level and are essentially matters of data aggregation,
fitting or parametrisation, taking into account rating or maturity. They might therefore seem rather
trivial: why cannot we simply take a set of bonds with the desired characteristics—rating, maturity
range, etc.—and then compute the average spread (weighting by issue size and duration would be
common practice)? The problem with this is that over time bonds enter and exit the bucket, for
various reasons: new bonds being issued; existing ones being redeemed or called; ratings changing;
maturity steadily declining so that the bond moves into or out of the desired maturity range. This
will create jumps in the average spread. Also if the desired bucket is too narrowly defined, e.g. 5–7y
miners rated BB−, we might find no bonds at all. Finally, there is no law that states that bonds
should trade monotonically with credit rating, and in practice we can easily find examples where a
BBB name trades tighter than a BBB+ one (at the same maturity point), as the market spread is
in a sense current, and the rating may be out of date, or perceived as such. All this indicates that
we need to produce a parametrically-defined grid of curves that are of a sensible shape and vary
monotonically with rating (i.e. do not cross). Then the estimated spread for a particular rating
and tenor is a sort of weighted average of bonds of nearby rating and tenor.
By contrast ‘B’ refers to specific issuers. In CDS markets this question seems trivial because
a CDS curve is fundamental to the asset class; but in practice liquidity is concentrated at the 5y
point, so that at other maturities we only have indication levels used for marking traders’ books,
and is nonexistent beyond 10y. In fact, most issuers do not trade in CDS, and then one must refer
to a particular bond: then, the difficulty is that the bond ages over time and so part of its spread
variation results from reduction in maturity. In fact this calls for the same solution as above, but
with the simplification that we only want one curve, not one for each rating. While some issuers
come to market frequently, giving an almost continuous picture of the term structure, others may
only have one or two bonds outstanding.
This leads neatly to ‘C’, which refers to the analysis of a particular bond. It may be helpful to
define our terms:
• Carry is the profit and loss (PL) contingent on the yield of the bond remaining fixed. For a
par bond this is synonymous with the coupon, but for a bond trading > 100 the bond price
will decrease and for one trading < 100 it will increase as a result of the pull-to-par-effect. The
carry arises from two sources: a pure interest-rate component and a credit spread component.

• Rolldown is the PL contingent on yield change that arises a result of maturity reduction,
with the curve assumed to remain fixed. Usually curves are upward-sloping and so the yield
reduces, and the rolldown is positive, but when the curve is inverted the effect will be negative.
Carry and rolldown are often taken together, and represent the total PL arising from ageing
of the bond while the curve remains fixed.

• Credit relative value (RV) expresses, in spread or price terms, the degree to which a bond
offers good value relative to its peers. This is often referred to as ‘rich/cheap to the curve’—
necessitating, of course, a curve.
At the risk of repetition: rolldown and RV require the construction of an issuer’s curve. (For carry
it is less obvious, but there is a subtlety that we will come to presently.) But this is in practice a
mythical beast, and needs to be built from available bond data.
Despite the fundamental nature of this subject, the literature, which stretches back some thirty
years, is disappointing. At one end of the spectrum are highly academic treatises such as [3] which

2
develop the subject from the perspective of complete markets and martingale pricing. Despite its
impression of intellectual depth, this paper is faulty in several respects. One objection is that credit
markets lack the necessary completeness for the mathematical grounding to be valid, but worse
than that is that the entire development follows from their incorrect eq.(1), pricing a contingent
claim in terms of a ‘default-adjusted discount rate’ R = r + hL, with r the riskfree short rate, h the
hazard rate and L the LGD1 . The incorrectness of this stems from the fact that the loss mechanism
is wrong: the discount factor is the exponential-integral of −Rt , which depends nonlinearly on L.
The financial interpretation of terms in L2 and higher (in the Taylor expansion around L = 0) is
that default may occur more than once in the given time interval, and that losses are in proportion
to the market value of the asset just before default. But this bears no resemblance to reality: a bond
can only default once, and the loss is a proportion of the par value, as that is the bankruptcy claim.
This error is fundamental because the coupon stream and principal payment are different claims and
require different discounting, and as we show later there is a tendency for premium bonds to trade
at a higher yield than discount bonds2 of the same maturity—we call this the par/non-par problem
from now on. The authors clearly do not recognise this, as they state at the beginning of their §4,
“The inability to separately identify [the hazard rate and LGD] using defaultable bond yields . . . ”
which as we have said is at variance with market pricing. Indeed, typical of this kind of discussion
(and there are plenty of other examples from academic institutions) is its lack of connection to
real-world examples from the debt markets. Lando [5] makes some more progress, to the extent of
valuing, at least in theory, the coupon stream and the recoverable portion of the principal when
default can occur at an unknown time, and goes on to discuss the simultaneous fitting of multiple
rating curves, and also rating dynamics via a transition matrix. The par/non-par problem is not
addressed, though, and there are theoretical and practical difficulties in fitting a transition matrix to
bond spreads. Many different matrices can give almost identical term structure [6], and resolution
of this ambiguity requires spread volatility information; and the process is quite computationally
intensive. We think a simpler and quicker idea is to write down directly a parametric form for the
survival curve in each rating state: that way, it is clear how the parameters directly relate to the
term structure and hence to the calibration data—which is not at all the case with rating transition
models.
At the other end of the spectrum lies the ‘applied’ literature, some of which is little more than a
plug for commercial software implementations, e.g. [2] which does little to address the main issues.
The ‘middle ground’ should be occupied by articles authored by quants with a solid technical
grounding and good experience of working in the markets [1]; these are probably the best sources
for the working quantitative analyst, but even so this paper achieves, in our view, more in less
space.
Here in basic terms is what we regard as the correct approach:

• We should discount cash-flows using riskfree discount factors and survival probabilities. Meth-
ods such as Z-spread are unsuited to bonds trading away from par, but in practice the effect
of recovery—which goes a long way to explain why high dollar price bonds trade at higher
spread—has not been carefully dealt with. Also, no bond spread definition is compatible with
CDS spread. Berd et al. [1] come to the same conclusion in an imprecise and roundabout
way. Van Deventer [2] uses a maturity-smoothed version of the Z-spread, and declares that
it has been successfully used since 1993: presumably he considers this to be some sort of
recommendation, but instead it indicates a basic lack of understanding. In fact, Figures 1,4
illustrate why such an approach is poor: the term structure would end up with a large kink
1
Loss given default, here as expressed as one minus recovery, 1 − ℜ.
2
A bond is said to be discount if its dollar price is below par, premium if above.

3
 at the 17y point, without economic justification as it is just a par/non-par effect.
• We do not need hazard rate models, and the existence of the survival function Q(T ) does
not need pseudo-academic grounding in arbitrage pricing theory (which is irrelevant to the
problems at hand). Instead we simply declare that the PV of a risky dollar occurring time T
from now is the product of a riskfree discount factor B(T ) and the survival function, hence
B(T )Q(T ).
• It is necessary to parametrise Q(T ) using an appropriate monotone-decreasing function and
we use eq.(9). This is done when we have one issuer and many bonds/CDS. It is the first
step in the so-called HJM approach [4, App.C]; the second is to write down the risk-neutral
dynamics, which we consider in forthcoming work.
• When there are too few bonds to fit a curve for an issuer, or where we want to think about
relative value, we index credit quality using ratings. We can then think about how a name
trades relative to its rating curve, and can use either the public rating or our opinion of
it, or conversely find its market-implied rating by seeing which curve prices it most closely.
When fitting multiple rating curves we need to do them all at once, which necessitates a more
general functional form for Q(T ), eq.(10) et seq., and we must be prepared for data that are
very ‘noisy’.
• In principle one can PV a bond/CDS without any computational short-cuts by discounting
the exact cash-flows using their exact dates. Our methods can be used this way. In practice,
though, one loses little and gains a much simpler implementation by assuming the coupons to
be continuously-paid3 . For CDS this is not a problem because the coupon is always 1% or 5%
and paid quarterly. In bond markets one does occasionally see some hoary specimens with
coupons > 10%, and when the issuer is distressed this assumption starts to be questionable.
However, in such cases the survival curve is likely to be highly idiosyncratic and determined
by the precise timings of the firm’s cashflows, so a generic model is of limited value.
• As intimated above we fit to prices (or CDS PVs). However, we want to plot the spread vs
maturity and this necessitates what we are calling the par-adjusted spread, which removes
the par/non-par effect.

1 Methods
The credit-riskiness of a bond is, for any performing credit, best encapsulated by a quantity known
as the spread which, loosely, indicates how much yield it has by comparison with a riskfree bond of
the same maturity. So this is what we are trying to model. (When a credit is distressed and unlikely
to make its payments, its yield becomes so high as to be meaningless and then the dollar price of the
bond is the most sensible metric. As we only interest ourselves in performing credit we ignore this
from now on.) Here is our development: we write down the valuation formula for a bond in terms
of survival probability and riskfree discounting; compare and contrast with Z-spread and carefully
explain the ‘par/non-par problem’; define par-adjusted spread; give examples of parametric fitting
of the survival curve; move to fitting multiple curves; explain how to calculate carry and rolldown
and relative value properly. But before progressing to the main course we need to munch through
one particular nettle: what is the riskfree rate?
3
van Deventer [2] makes a fuss about using exact rather than scheduled payment dates, but a moment’s thought
shows that a day or two’s discounting gives rise to a change in value that is negligible compared with the bid-offer of
the bond.

4
1.1 Riskfree rate
A good part of the matter’s complexity predates the whole LIBOR/OIS situation by decades and
is rooted in whether the investor is ‘real-money’ or ‘levered’.
The vast majority of bond investing is done by real-money accounts such as pension funds,
insurance companies and sovereign wealth funds. For them, the relevance of a spread measure
is simply to quantify the excess yield over a Treasury bond of the same maturity (we are USD-
focused; in EUR we would use the German government yield). To hedge interest rate risk, they
short Treasuries or the futures. In that case the appropriate riskfree rate is simply the Treasury
yield, and indeed Duffie & Singleton say the same in [4, §7.2]. Aside from simplicity, this definition
confers another advantage: spreads will necessarily be positive.
As soon as we move to levered investing the position becomes more complicated. The correct
discounting rate is OIS and depends on the collaterisation of the borrower. However, even this is
not fully correct, because the economics may be more complicated: a hedge fund will typically have
a chunk of cash to invest and also a leverage facility, which it can use on demand. When operating
without leverage the riskfree rate is the same as for real-money firms, but when the leverage facility
is used, the funding rate and collateralisation come into play. To us it seems that the best way of
assessing bond spread is to use Treasury as the riskfree rate, and then when a firm needs to borrow
money the economics of the trade include the borrowing costs which have their own idiosyncrasies
and dynamics.
That said, one can use the swap rates and derive spreads to the swap curve, as is often done,
and end up with (usually) a slightly lower answer than the spread to Tsy. This is unlikely to cause
practical difficulties, but there is one thing to watch: the spread for the highest-grade issuers may
not be positive, as they may be a better credit risk than the banks. If, as here, spreads are required
to be positive, then these issuers will always appear expensive.

1.2 Valuation formulae

We write B(T ) for the riskfree discounting curve and Q(T ) for the survival curve. The PV of a
survival-contingent payment occurring at some future date T is B(T )Q(T ). This allows the coupon
stream to be valued as a sum X
cj B(Tj )Q(Tj ).
j

If, as previously mentioned, we approximate this as a continuous payment stream, then it is cΠ(T )
where Π is known as the RPV01 (risky PV01), where
Z T
Π(T ) = B(t)Q(t) dt.
0

The principal repayment is simply B(T )Q(T )—but only if there is no possibly of recovery. In reality
if default occurs we can present a claim for the principal amount and expect some proportion ℜ
(the recovery rate) to be honoured4 .
To value this extra amount, we divide the time frame [0, T ] into slices and note that the prob-
ability of default in precisely the interval [t, t + dt] is −dQ(t). Multiplying by B(t) and integrating
gives ℜ Ξ, where Z T
Ξ(T ) = − B(t) dQ(t).
0
4
Some of the older literature e.g. [3] discusses recovery of market value, recovery of Treasury, or recovery of
face/par value. We will not be sidetracked by this issue, as only the last of these is a good representation of what
really happens.

5
Both Π, Ξ are bilinear forms in5 B&Q. Adding the parts gives the model price of the bond:

Pb/100 = c Π(T ) + B(T )Q(T ) + ℜ Ξ(T ). (1)

This formula, and the reasoning behind it, emphasises an important distinction between the prin-
cipal and the coupon stream: the former is partly recoverable, the latter not. Hence the coupon
stream is a riskier claim on the firm than the principal repayment, and this is the root of the
par/non-par problem. Incidentally we said when discussing [3] that it makes no sense for the
valuation formula to be nonlinear in the LGD: clearly, (1) is linear in ℜ.

1.3 The problem with Z-spread

Compare (1) with the simple-minded price-yield relationship for a vanilla bond, which does not
recognise the concept of recovery:

1 − (1 + y/m)−mT
P/100 = c + (1 + y/m)−mT , (2)
y
with m the compounding frequency. This comes from the more general equation for the internal
rate of return (IRR) of a set of cashflows,
X
P/100 = Cj (1 + y/m)−mTj , (3)
j

in the specific case where the coupon amounts are c/m, plus the principal at time T , and summing
over the coupons as a geometric progression. (This is exact for mT an integer, and we use it
regardless.) The Z-spread sZ is obtained from (3) by replacing y with the sum of two parts, the
riskfree zero rate6 and a constant sZ , whose value is uniquely inferred from P .
The similarity between (1) and (2) is that (2) is a special case of (1) when B, Q are both
exponential functions. The difference is that (2) treats the coupons and principal as essentially the
same, but the market clearly does not, as we now explain.
If two bonds of the same maturity but different coupon trade at the same yield, then this means
that the discounting mechanism is the same for the principal as for the coupon, i.e. just using a
risky discount factor, and so the market is pricing zero recovery. Conversely if we consider that
recovery will be zero then the two bonds should trade at the same yield: the downside of buying
each is the same, and so the upside, as measured by the yield, should be too. We can formalise this
by saying that if the high-coupon bond trades at a higher yield then we should buy it and short
an equal cash value of low-coupon bond. This will guarantee a profit, if either the issuer survives
or else defaults with zero recovery; but if the recovery is intermediate there can be a serious loss,
e.g. if the bonds are trading at 130 and 85 when the trade is put on, and a little while later default
occurs with recovery 90%, then both legs of the trade lose money.
But in general the market does not price bonds this way, and so the high-coupon bond trades
at a higher yield. See for example Figure 1, illustrating a common effect. For more precise analysis
we need to find bonds of widely different coupon but almost identical maturity, but that situation
is uncommon. As bonds are nearly always issued at par, it only happens when the bonds were
issued a long time apart, with maturity dates that happen by chance to (almost) coincide, and
when yields have moved a long way in the interim.
5
Paying homage to a well-known British DIY chain.
6
The value z = z(T ) such that B(T ) = (1 + z/m)−mT .

6
 6

Yield (%)
3

0
0 5 10 15 20 25 30
(i)
8.125% ’24
6 4% ’24

5
Yield (%)

1
2014 2015 2016
(ii)

Figure 1: (i) COLOM bonds on 08-Apr-16: yield vs tenor. Red, yellow, black, green respectively denote
bonds trading above 120, around 110, around 100, and around 90 dollar price. (ii) Comparison of 8.125%
’24 vs 4% ’24 over time shows a consistent yield (spread) difference.

A good example is Colombia, which in 2004 issued 8.125% bonds due 21-May-24, and in 2014
issued 4% bonds due 26-Feb-24. The yield difference in Figure 1(ii) is reasonably static over time,
and cannot simply be ascribed to liquidity7 . We now value these bonds as of 08-Apr-16 using a
flat hazard rate Q(T ) = e−λT , with different assumptions about recovery. In each case the hazard
rate λ is adjusted so as to make the average pricing error ∆P = Pb − P zero. Table 2 shows the
results. For ℜ = 0 the model price is too low for the 4% bond and too high for the 8.125% bond, as
expected. As ℜ is raised the error reduces until at ℜ = 53.5% both match, and this is the implied
recovery. If we think that Colombia recovery should be lower than this figure, then we buy the
8.125% and enjoy the higher yield while it lasts; if higher, then we should buy the 4%. In the
limit of ℜ → 100%, which seems silly but could in principle happen in a ‘technical’ default, the 4%
7
It is a convenient fact that the author traded both of them over the period shown.

7
 Coupon(%) 4 8.125
Tenor(y) 7.88 8.11
Price 101.10 125.50
Yield(%) 3.98 4.36
ℜ ∆P ∆P λ
0% −1.45 +1.45 0.0277
20% −1.10 +1.10 0.0341
40% −0.56 +0.56 0.0442
53.5% 0.00 0.00 0.0551
60% +0.36 −0.36 0.0626
80% +2.22 −2.22 0.1065

Figure 2: Two COLOM ’24 bonds as of 08-Apr-16, valued with different recovery assumptions. ∆P is
the model price minus the market price (hence, positive means bond appears cheap).

become risk-free, so these are obviously the ones to buy.

When building curves we should not, therefore, simply use yield or yield-related spread measures
such as Z- or I-spread. Although it is too fiddly to attempt to infer the market-implied recovery
when we have many bonds of different maturity and dollar price, we should make a better attempt
than simply assuming the recovery to be zero. Otherwise, high dollar price bonds always look
cheap, when in fact much of the cheapness is illusory. In the case of the two COLOM ’24 bonds,
if we make the routine assumption of 40% recovery, we will have the 8.125% modelled about a
point cheap to the 4% bonds—still a gap, but only a little wider than the bid-offer spread (which
would typically be around 0.5–0.75 pts), and so the transaction cost of switching from the 4% to
the 8.125% would scarcely be worth the bother, particularly when we remember that high dollar
price bonds tend to be more difficult to trade. But if we assume ℜ = 0 then there appear to be
almost 3 pts of value in switching.

1.4 Comment on other spread measures

We have already discussed Z-spread. Other measures are available but all have their own problems;
we briefly discuss them now. One way in which spread definitions differ is in their assumptions
about the RPV01.
The benchmark spread or spread to Treasury (SoT) is the most rudimentary of the yield-based
measures, and is almost uinversal in USD markets. It is simply the yield difference between the
bond in question and the benchmark Tsy bond which is not maturity-matched. Thus bonds of
maturity 7–15y are all quoted off the 10y Tsy. This is helpful for pricing on the day of the trade
but it is a useless construct for analytical work. As a bond’s maturity moves down to around 7y,
it jumps to being quoted off the 5y Tsy, which typically has lower yield than the 10y, and so the
SoT suddenly jumps up even if the bond price has not moved!
The asset-swap spread of a fixed-rate bond can be thought of in a couple of equivalent ways.
One is that I can hand over the coupons and an upfront amount 1 − P/100 (if this is negative, I
receive money) to a swap counterparty in return for LIBOR plus a spread sA , which is the asset-
swap spread. Consequently, with R denoting the par swap rate, and Π◦B , Π◦F the fixed and floating
swap PV01s:
P
1− + (c − R)Π◦B
sA = 100 . (4)
Π◦F

8
 CDS
Z-spd

ASW

100ℜ PRF Price

Figure 3: Asset-swap spread, Z-spread and CDS spread (schematically) vs bond price. All are zero
when the bond price is PRF , the value of a bond with the same payment schedule but no credit
risk. Note the different asymptotes as P → 0.

This is linear in the bond price, because the numerator of the above equation is the riskiness of the
bond in price terms—it increases if P is lower or c is higher, and is zero if the bond trades flat to
LIBOR—but the denominator is the swap PV01, which has nothing to do with the bond’s credit
quality.
Finally the par CDS spread is the value of the running spread that makes a CDS contract
value to par. It is similar to the Z-spread in the sense that if the CDS trades off the same survival
curve as the bond (i.e. there is no basis between the two markets) then, as the survival probability
declines, and with it the bond price, the par spread increases and is a convex function of the bond
price. The difference is that the CDS spread hits ∞ when the bond price hits recovery; so typically
it exceeds the Z-spread.

1.5 Par-adjusted spread

One might suppose, given that valuation of bonds can be done without reference to any kind of
spread measure (one needs only B, Q and a recovery assumption), that we can scrap the whole
idea of spread altogether. That causes a problem, however. It is more natural to think in spread
terms, and it is fundamental that we graph spread vs maturity (or duration). We have said that
the Z- and I-spreads are not right, but then what is the correct idea? We pursue this now.
From (1) we have, using the par CDS spread s(T ) = (1 − ℜ)Ξ(T )/Π(T ):
   
1 − B(T )Q(T ) − Ξ(T )
Pb/100 − 1 = c − −s(T ) Π(T ). (5)
Π(T )
| {z }
rb(T )

The LHS is the deviation from par of the (model) bond price. On the RHS the term in square
brackets is understood as follows: the coupon minus a sort of riskfree rate, to be explained presently,
minus the par CDS spread. This is intuitively reasonable because if the bond is trading at par
then its asset-swap spread (coupon minus swap rate) should equal the par CDS spread, to avoid
arbitrage. We now explain the second term, which we will write as rb(T ), in more detail: its

9
numerator is, using integration by parts,
Z T Z T
1 − B(T )Q(T ) + B(t) dQ(t) = − Q(t) dB(t),
0 0

which can be neatly recast as the parity equation

B(T )Q(T ) + Ξ(T ) + rb(T )Π(T ) = 1; (6)

and recalling the definition of the riskfree instantaneous forward rate f (t) = −B ′ (t)/B(t), we find
RT
f (t)B(t)Q(t) dt
rb(T ) = 0R T ,
0 B(t)Q(t) dt

the average of instantaneous riskfree forward rates weighted by the risky discount factor. In the
case of a ‘flat riskfree curve’, f is constant, and as Q then divides out, rb is just equal to f . When
Q ≡ 1, rb is just the riskfree par rate. Now using the actual bond price P we can write


P/100 − 1 = c − rb(T ) − s Π(T ), (7)

which defines s, and we call it the par-adjusted spread of the bond. Note that s − s(T ) is the basis
in spread terms between cash and CDS8 .
The par-adjusted spread of a CDS is just the par spread. If the CDS is quoted upfront u plus
a fixed running spread of c, then the par-adjusted spread is

s = c + u/Π,

which is the coupon plus the upfront converted into a running spread. If instead the CDS is quoted
as a spread se then it is only approximate to say that s equals se, though if the CDS happens to
be trading at par then this is exact. The correct argument9 is to convert the CDS into an upfront
u = (e e where Π
s − c)Π, e is the ISDA RPV01 calculated using a flat hazard curve derived from the
traded spread (by λ = se/(1 − ℜ)). Then the previous formula is used. Incidentally the Bloomberg
e This quantity is, in our
CDSW screen calculates a so-called price Pcds as Pcds /100 − 1 = (c − se)Π.
notation, just −u. Then, analogously with (7), we have

Pcds /100 − 1 = (c − s)Π(T ).

the main difference being the absence of the rb term, which refers to the funding of the bond, but
is not needed for a synthetic position. A different way of expressing this is the conversion of the
traded spread se into the par spread:

(s − c)Π = (e e
s − c)Π.

We are now in a position to write down the expression for the deviation in price terms between
the (par-adjusted) spread of a bond s or CDS and a curve giving s(T ), Π(T ):



∆P/100 = s − s(T ) Π(T ) = 1 − P/100 + c − rb(T ) − s(T ) Π(T ) (8)

with rb omitted forPa CDS. In fitting to price data, we can simply choose to minimise the total
squared deviation j (∆Pj )2 . In practice we will typically want to weight the errors by issue size10 .
8
The so-called negative basis trade is where the cash bond trades cheap to CDS: in our notation, s − s > 0 then.
9
This convention, Standard American or SNAC, has been in force since the spring of 2009, and is universal.
10
Technically, the amount outstanding in million USD. We assume this to be $1Bn for liquid CDS.

10
Further, an improvement over simple least
p squares to use a penalty that increases less rapidly than
quadratically for large deviations, e.g. 1 + (∆P )2 − 1. A fuller discussion of the general principle
is in [9, §15.7].
Returning to the two COLOM bonds, when ℜ = 53.5% we price both exactly, using the same
hazard rate, and the par-adjusted spreads are both 260 bp. For these assumptions neither bond
is ‘better’ than the other, as their maturities are virtually identical and they have the same par-
adjusted spread. But if we use ℜ = 0 we need different hazard rates to price the bonds: for the
4%, we have λ = 0.0256 and s = 258 bp, and for the 8.125%, we have λ = 0.0296 and s = 298 bp.
The difference in par-adjusted spread (40 bp) is similar to the yield difference. In conclusion, the
par-adjusted spread measure takes recovery and deviation from par into account.

500

450

400

350
Par−adj spd (bp)

300

250

200

150

100

50

0
0 5 10 15 20 25 30
(i)
500

450

400

350
Par−adj spd (bp)

300

250

200

150

100

50

0
0 5 10 15 20 25 30
(ii)

Figure 4: COLOM bonds on 08-Apr-16: par-adjusted spread and fitted curve. Parameters: (i) ℜ = 0,
a = 0.0099, b = 0.0621, c = 0.2; (ii) ℜ = 50%, a = 0.0168, b = 0.2727, c = 0.05.

11
1.6 Hazard curve model: Single-name case
In general a constant forward hazard rate will not be sufficient to capture the term structure
accurately, so we need to fit a curve. We first deal with the case where we have one name, and
enough bonds to make it sensible to fit a curve. If this is not so then we need to fit multiple names
and maturities in the sector and, in effect, interpolate based on internal or external credit rating.
This is covered by the next subsection. After that we can always adjust one parameter so as to
exactly fit the model curve to a given bond.
A model that seems to offer the right amount of flexibility, while giving a sensible shape for
maturity T → ∞, is
d a + bcT
Q(T ) = (1 + cT )(b−a)/c e−bT , − ln Q(T ) = (9)
dT 1 + cT
(the second expression being the forward hazard rate). It is necessary and sufficient for all three
parameters to be positive, and they have a convenient interpretation: a, b are respectively the
forward hazard rates for T → 0, T → ∞, and c influences the shape of the curve between these
limits. Typically a < b, resulting in an upward-sloping spread curve, but for dubious credits we will
have the opposite. We suggest restricting the range of the time-scaling coefficient c to [0.05, 0.2]
(where T is understood to be in years); in fact, little is lost by fixing c, thereby reducing the number
of free parameters to two. The same idea is often used when calibrating the Nelson-Siegel model
[8], which has one more degree of freedom than (9).
A criticism of the above is that it cannot fit a humped forward hazard curve, as it has too few
parameters. But parsimony confers two advantages: robustness and explainability. The need for
robustness is amply demonstrated later on. The curve’s shape is determined by two influences: the
perceived credit quality of the issuer over different time horizons, and supply and demand in the
market. It is unlikely that either can give rise to highly nuanced curves with subtle shapes (except
as we have said for distressed credits): for one thing, there is simply too much uncertainty in the
future profitability and leverage of the issuer. Our view is that in general anything more complex
than a simple upward- or downward- sloping curve is likely to be overfitting. The results in [2]
show kinks for which no economic explanation is offered.
Figure 4 shows the results for the entire set of COLOM bonds, using (i) ℜ = 0 and (ii) ℜ = 50%.
Note how the 10.375% ’33 (145 dollar price, seen on the graph at T ≈ 17y), lies a long way off
the curve when ℜ = 0, for the same reason that it does in Figure 1(i), but not when ℜ = 50%.
Indeed, the fit is generally better in (ii), probably because the latter is a more realistic recovery
assumption, and hence a better approximation to the way the market works.

1.7 Hazard curve model: Multiple ratings

This is the more general form of the model. We now want to fit many ratings at once, and allow
the parameters a, b to be rating-dependent:
d aj + bj cT
Qj (T ) = (1 + cT )(bj −aj )/c e−bj T , − ln Qj (T ) = (10)
dT 1 + cT
It is a useful feature of the way that credit markets seem to work that, roughly, spreads vary in
geometric progression across the linear rating scale

AAA=1, AA+=2, AA=3, . . . , BBB=9, . . . , CCC=18.

This is helpful because when we come to fit to data we cannot have dozens of parameters: a
reasonable number is between five and ten. We suggest the following, which uses seven:

12
 • Short- and long-term forward hazard rates a, b for rating AA (=3)

• Short- and long-term forward hazard rates a, b for rating BBB (=9)

• Short- and long-term forward hazard rates a, b for rating B (=15)

• Shape parameter c, assumed equal across all ratings

This allows us to capture global movements of all curves up and down, steepening/flattening, and
compression (i.e. high yield spreads decrease and investment grade increase) or decompression. In
between these three ratings, we use logarithmic interpolation; outside, we use logarithmic extrap-
olation.
How much dispersion is there when we fit the curves? It is indeed naive to suppose that bonds
will line up nicely with their appropriate rating curves, but is the deviation quite small, or does
a plot of spread vs maturity, using different colours for different ratings, look more like a swarm
of multicoloured bees emerging from a hive? Those with market experience will be more likely to
take the latter view, and indeed this is close to reality. The same point was made in relation to
risk-neutral calibration of credit migration models, in [6]. See Figure 5 for the mining sector; other
sectors are similar.

B−
B
500 B+
BB−
BB
BB+
400 BBB−
BBB
BBB+
Spread (bp)

A−
300 A
bonds

200

100

0
0 2 4 6 8 10 12 14 16 18 20

Figure 5: Fitted spread curves for the mining sector, Jan’18. Bonds shown in dots using same colour
scheme as rating curves. Note the dispersion around the fitted curves.

We have already suggested that the market does not always respect external ratings, as it is
providing a constantly-evolving view of credit risk. Some of the dispersion that is found can simply
be attributed to external ratings being very out of date. A more refined approach is to use the same
rating scale for our internal ratings, and then when we are doing the fitting we use those instead.
For example, in recent years PEMEX was externally rated BBB− while trading at 400+bp spread;
this level, and the credit metrics, were more consistent with BB−. By using that as the rating we
will typically get a better fit (lower fitting residual) and big issuers that are ‘clearly misrated’ will
not distort the fitting procedure. The advantage of using the same rating scale is that where we
are agnostic about the rating—and we cannot have an opinion on every issuer under the sun—we
can simply use the external one.

13
1.8 The need for variable recovery
We have explained why it is necessary to use a positive recovery rate rather than assuming it to be
zero. However, making a simple 40% recovery assumption, in line with CDS single-name and index
pricing, is not ideal either. This is because the model will be used to assess relative value. For
bonds trading at a low dollar price, a 40% recovery assumption will make them appear expensive
regardless of the model credit rating. To take a specific example, a 1y bond with a high market-
implied default probability of 60%, assuming ℜ = 40%, will have a price P ≈ 64, but there are
plenty of bonds trading lower than that. Arguably models of this sort are no longer useful once a
bond becomes distressed, as all bonds will trade at or near recovery, which in turn is completely
idiosyncratic, but some improvement on an artificial 40% seems desirable. One solution is to use
recovery that declines with rating, and a useful construction is to make the recovery 70 − 3r%,
with r the linear rating. Thereby BBB− has 40% recovery, AA has 61%, BB has 34%, B has 25%.
There is a case for having sovereigns a little higher than this.

1.9 Computational matters

The expressions for Π, Ξ involve integrals, which we in practice approximate using the trapezium
rule: in detail,
N
X B(tj−1 )Q(tj−1 ) + B(tj )Q(tj )
Π(tN ) ≈
2
j=1
N
X B(tj−1 ) + B(tj )

Ξ(tN ) ≈ Q(tj−1 ) − Q(tj )
2
j=1
N
X
Q(tj−1 ) + Q(tj )
rb(tN )Π(tN ) ≈ B(tj−1 ) − B(tj )
2
j=1

and these approximations exactly satisfy the parity equation (6).

1.10 Relative value, Carry, and Rolldown

We now turn to problem C as identified at the outset. Here is a subtle matter worthy of attention.
Suppose a bond trades at a significant spread to its corresponding curve. We can say that this gives
rise to extra carry by comparison with bonds that lie on the curve, but also that the bond offers
relative value, by being ‘cheap to the curve’. However, in quantifying the total return we must be
careful not to double-count. An obvious way to avoid this is to say that over a period [0, ∆t] the
bond earns extra carry than a bond on the curve, but the relative value component is obtained at
the end of the period. Loosely this means that the relative value is the spread difference multiplied
by the forward duration11 , i.e. the duration as seen at time ∆t, not as seen today: otherwise we
will double-count.
We write c′ = c − rb for a bond, and c′ = c for a CDS, so that the formulae below apply equally
to both. Reminder: s = par-adjusted spread, sb = model par spread, se = traded CDS spread.
• Credit carry is the change in value, plus accrued coupon, in the event that the spread remains
unchanged. Over a period ∆t this amounts to


c′ ∆t + (s − c′ ) Π(T ) − Π(T − ∆t) (11)
11
Ambiguous as the forward RPV01 is commonly understood to mean Π(T ) − Π(∆t), whereas we mean Π(T − ∆t)
here.

14
 and the last term is the pull-to-par. Hence the carry is not simply s ∆t. We said earlier
that this calculation, despite being superficially trivial, requires a model curve, and now the
reason becomes clear: we need Π(T ).

• Rolldown is the effect of the spread changing from rolling down the model curve. This is
the spread change multiplied by the RPV01 on the future date, assuming that all curves are
unchanged:

sb(T ) − sb(T − ∆t) Π(T − ∆t). (12)

• RV is the effect of moving towards the model curve at the end of the time period, and so is


s − sb(T ) Π(T − ∆t). (13)

The expression uses sb(T ) not sb(T − ∆t), as the latter would incorrectly double-count the
rolldown effect.

• Total return is the sum of all three and hence is

c′ ∆t + (s − c′ )Π(T ) − sb(T − ∆t) − c′ Π(T − ∆t). (14)

Notice that there is a delicate cancellation of terms en route when the components are added,
giving this neat result.

In fact, the end result can be derived in a different way, at least for CDS, which makes it clear
that it is correct. Consider the PL arising from selling protection (today) when the traded spread
is se0 and unwinding it later when it is se1 . As usual the coupon is c. This is12

c′ ∆t + (e e ) − (e
s0 − c)Π(T e − ∆t).
s1 − c)Π(T (15)

Earlier we showed that the par spread s is related to the traded spread se by (s − c)Π = (e e
s − c)Π.
So (15) corresponds exactly to (14), the idea being that the par spread is s today and moves to the
value on the model curve, which is sb(T − ∆t), over the period [0, ∆t].
There is a different way of doing it, giving the same total total return as before, but subdividing
differently:

• Carry differs by using the model spread sb = sb(T ):

c′ ∆t + (b
s − c′ ) Π(T ) − Π(T − ∆t) (11a)

• Rolldown is the same as above, (12).

• RV is the effect of moving towards the model curve at the beginning of the time period, and
so is

s − sb(T ) Π(T ). (13a)

It is easily seen that the sum of (11a,12,13a) is the same as that of (11,12,13), i.e. (14).
We have talked about total return, rather than expected return. This is because the above
development has assumed that the curve remains unchanged and the bond spread moves towards
it over the time [0, δT ]. A theory of expected return must capture more than this, as there are
other considerations:
12 e is the ISDA ‘flat-hazard-curve’ RPV01.
Recall that Π

15
 • The bond may default or credit migration may occur to a different rating state. But as we
have curves for each rating, this is easily captured by summing over all rating transitions,
with transition probabilities taken from the appropriate row of the chosen transition matrix,
for which see [6]. It is a reasonable way of thinking about the probability of large and sudden
price changes.

• Even if no transition occurs, the bond may not converge all the way to the curve off which
we think it should trade. This can be captured by multiplying the RV component by a
number between 0 and 1. But if this is done the two total return calculations given above,
i.e. (11a,12,13a) vs (11,12,13), will no longer be the same.

• Curve dynamics. This will be pursued in further work.

1.11 Extension to EM bonds

An effect which has prevailed in markets for many years has been that EM issuers of the same rating
tend to trade wider than DM ones (at the same maturity point). The wider the sovereign spread,
the more pronounced is the difference. Arguably, if the bond rating incorporates the sovereign,
indutrsy-specific and issuer-specific features, this effect should not occur, but it still does. One
way of analysing it is to incorporate it into the multi-curve fitting method in a rather obvious way,
altering eq.(8) to


∆Pj /100 = 1 − Pj /100 + cj − rbj,ℓ(j)(Tj ) − sℓ(j)(Tj ) − αssov
j (Tj ) Πℓ(j) (Tj ); (16)

where ssov (T ) is the par CDS spread of the sovereign at the associated maturity point and the
subscript j indexes the bond and ℓ(j) its corresponding curve. The coefficient α must be constrained
to lie between 0 and 1, and typically one finds, on fitting the curves and α, that α is a little below
0.5 on average. A refinement is to notice that highly-rated EM bonds have less connection to the
sovereign spread than lower-rated ones, and this effect is easily incorporated into the model.
There is a distinction between this and the work presented in [7], which attempts a more
fundamental approach based on treating the EM bond as a modified first-to-default basket of a
standalone issuer and its sovereign. There, the inputs to the model are the standalone rating (as
determined by fundamental analysis) and the sovereign spread; here they are the combined rating
(which is the external rating) and sovereign spread.

1.12 Examples and discussion

Our original questions A,B can now be answered and we show sample results by means of figures.
First, Q. A1/2/3: how has a particular curve changed over time? Figure 6(i) shows, for EM
banks, the 5y and 10y points on the A rated curve over time, and hence the level and steepness,
while (ii) shows the 5y point for different ratings. As a second example, Figure 7 shows the mining
sector. Notice how the curve flattened during the 2016 commodities crisis.
Next, Q. B1: how rich or cheap is a bond to its curve? Figure 8 shows the RV for BSANCI ’22
(a Chilean bank) over a period of a few years. Notice after some initial volatility, when spreads were
higher, how the relative value settles down to quite a tight range about zero, as might reasonably
be expected.

16
 250 600
A 10y BB 5y
A 5y BBB 5y
A 5y
500
200

400
Spread (bp)

Spread (bp)
150

300

100
200

50
100

0 0
2015 2016 2017 2018 2015 2016 2017 2018

(i) (ii)

Figure 6: EM banks curve (adjusted by sovereign), 2015–2018. (i) Different points on A curve. (ii) 5y
points on different curves.

400 1000
A 10y BB 5y
A 5y 900 BBB 5y
350 A 5y
800
300
700
250
Spread (bp)

Spread (bp)

600

200 500

400
150
300
100
200
50
100

0 0
2015 2016 2017 2018 2015 2016 2017 2018

(i) (ii)

Figure 7: Mining curve, 2015–2018. (i) Different points on A curve. (ii) 5y points on different curves.

17
 60
BSANCI ’22

40

20
RV (bp)

−20

−40
2016 2017 2018

Figure 8: BSANCI ’22 relative value (+ means cheap) to the A+ banks curve, over a period of a few years.

18
Conclusions and further thoughts
We have shown how to fit a survival curve to bond/CDS data, while avoiding the use of Z-spread,
which as we have explained does not deal with non-par bonds properly. We have deliberately chosen
a parsimonious model, and have found that keeping the number parameters low gives robust fitting.
Although the following goes beyond the scope of the current paper, it is appropriate to remark
on the general statistical properties of curves and relative value. It is easy to fall into the trap of
making unwarranted generalisations, but the following principles should be borne in mind:
• Sector spreads, and the market as a whole, typically exhibit momentum in the short to
medium term—as evidenced by long rallies after crises small and large—but long-term mean
reversion. The latter effect is explained by market risk premium being assumed to revert,
and is evident from graphs of bond spreads over a couple of decades.
• The issuer RV ‘mean-reverts until it doesn’t’. A standard pastime of bond investors is buying
bonds they consider ‘cheap to the curve’ and selling those that are ‘rich’. The reason that
mean reversion does not occur immediately is that there is no consensus as to where the curve
is at any moment; the work here shows how to do it in a way that we consider to be superior.
This method of investing works well until an accident befalls a particular credit. Then it will
trade wider and wider and always appear cheap, as the rating typically moves later than the
price. When this happens, the RV is likely to exhibit short-term momentum, as the move
away from its original curve tends not to occur in one big jump. Accordingly, the role of a
fundamental analyst is to determine, when an issuer starts to trade cheap, whether this is
transitory or instead because of some irreparable problem. Any competent analyst, trader or
PM should be able to come up with plenty of examples of both.

References
[1] A. M. Berd, R. Mashal, and P. Wang. Defining, estimating and using credit term structures.
Part I. Technical report, Lehman Bros., 2004.
[2] D. van Deventer, K. Imai, and M. Mesler. Credit spread fitting and modelling. In Advanced
Financial Risk Management, volume 3, 2012.
[3] D. Duffie and K. J. Singleton. Modeling term structures of defaultable bonds. Technical report,
U. of Stanford, 1997.
[4] D. Duffie and K. J. Singleton. Credit Risk: Pricing, Measurement and Management. Princeton,
2003.
[5] D. Lando. On Cox processes and credit-risky securities. Rev. Deriv. Res., 2:99–120, 1998.
[6] R. J. Martin. Credit migration: Generating generators. RISK, 34(3), 2021. arXiv:2006.11146.
[7] R. J. Martin, T. Uzuner, and Y. Ma. Emerging market corporate bonds as first-to-default
baskets. RISK, 31(7), 2018. arXiv:1804.09056.
[8] C. R. Nelson and A. F. Siegel. Parsimonious modeling of yield curves. J. of Business, 60(4):473–
489, 1987.
[9] W. H. Press, B. P. Flannery, S. A. Teukolsky, and W. T. Vetterling. Numerical Recipes in C++.
Cambridge University Press, 2002.

19