Professional Documents
Culture Documents
◗ February 2005
◗
Quantitative Strategy
Credit
Research
Analysts
Julien Turc
This article introduces a new model for pricing CDOs. This research is targeted at
+ (33) 1 42 13 40 90 CDO managers, investors and relative value players who want to improve their
Julien.turc@sgcib.com
understanding of the relative value between various index tranches and bespoke
Philippe Very CDOs.
+ (33) 1 42 13 55 96
Philippe.very@sgcib.com
■ This model is an extension of the traditional one-factor Gaussian model. It makes
David Benhamou correlation a function of the economy. This is the so-called local correlation.
+ (33) 1 42 13 94 75
David.benhamou@sgcib.com ■ We exhibit a simple and approximate formula for deducing the local correlation
■ Therefore, we present a more robust process for fitting the local correlation curve
directly to market data. This process is based on a mapping of the state-of-the-
economy factor into a strike, thereby making local correlation a very concrete and
intuitive variable.
■ Incidentally, we show that the base correlation skew is not a straight line. On low
strikes, it must have the shape of a smile rather than that of a skew.
■ We exhibit a simple rule for switching from index to bespoke correlation: attribute
the same (local) correlation to strikes that have the same risk of being hit. We show
how to enforce that rule in practice.
■ We show the impact of the model to the marked-to-market and hedge ratios of a
bespoke CDO. Our model generates a more stable marked-to-market than the
traditional Gaussian model.
Bloomberg: SGCT
www.globalmarkets.sgcib.com
Pricing CDOs with a smile
The gaussian Copula model Whereas the Gaussian copula model has become the established way of pricing
does not fit to the whole
correlation products, the market has felt the need for a coherent framework in which
correlation smile
either index tranches or non standard CDOs such as bespoke single tranches, and
CDOs squared could be valued. The Gaussian copula model does not provide an
adequate solution for pricing simultaneously various tranches of an index, nor for
adjusting correlation against the level of credit spreads.
Recent research has explored some ways to account for the so-called correlation
smile. Among these attempts, the most successive ones make correlation a function
of the economy. This is a simple and powerful idea: default risks should be more
correlated in an adverse economic environment. Yet, these models typically involve a
wide range of hidden parameters and unrealistic assumptions on issuer-specific
correlation. Moreover, they still heavily rely on an abstract ‘state-of-the-economy’
factor that is not likely to appeal to practitioners.
Our model is an extension to Practitioners on their side have built rule-of-thumb techniques for pricing bespoke
the Gaussian copula model.
and exotic CDOs. In order to bridge the gap between quantitative research and
It is based on a local
correlation practical CDO management, we propose to apply simple ideas already widely used on
equity derivative markets. Equity derivative traders and structurers make extensive
use of so-called local volatility models, making equity volatility a continuous function
of stock prices themselves. These models make the link between Black & Scholes
implied volatility and local volatility, and require only minor adjustments to pricers
already build within the traditional Black & Scholes framework. This is the reason for
their popularity.
!A first part goes back to the definition of correlation in CDO products and the
model commonly used for analysing correlation. We introduce our local correlation
model and compare it further with local volatility models for equity derivatives.
! A second part presents a technique for deducing the local correlation from the base
correlation skew. We exhibit a simple formula under the assumption of a large and
homogeneous portfolio (the so-called large-pool model). This approach is dependent
on how the base correlation skew is extrapolated below the minimum strike traded on
the market.
! In a third part, we present a more robust process for fitting the local correlation
curve directly to market data. This process is based on a mapping of the state-of-the-
economy factor into a strike. We explain why correlation must have the shape of a
smile rather than that of a skew, at least on very low strikes.
! A fourth part compares the market practice for pricing and hedging CDOs to our
local correlation model. We give first numerical results in term of both marked-to-
market and hedge ratios.
Finally, an appendix provides detailed mathematical demonstrations for three major
results: the market law of losses, the relationship between local correlation and
implied losses, and the relationship between marginal compound and local
correlation. Reading and understanding the article does not require going through
this technical appendix.
2
Quantitative Strategy
The traditional one-factor Here is one way to formalise this model: we decided to use the same correlation ρ for
Gaussian copula model is
all names that solely depends on X , the level of the economy. Each name j in our
based on a state-of-the-
economy factor reference is described by its asset A j that is the weighted average of 2 Gaussian
variables: X the level of the economy and εj the factor specific to the jth name X
represents a systemic risk: the higher it is, the better the economy is. The weight is
the correlation ρ to the economy:
Aj = − X ρ + ε j 1 − ρ
For a given horizon, default occurs if the asset A falls below a certain threshold. This
threshold is calibrated to generate default probabilities coherent with current market
spreads.
As illustrated in our last article on CDOs, this standard Gaussian copula model hardly
provides a coherent pricing framework. Senior and high Mezzanine tranches trade at
level of spreads that imply a high correlation compared to Equity and low Mezzanine
pieces. The so-called correlation smile can be translated into a base correlation curve,
and this base correlation shows that, indeed, correlation increases when it comes to
pricing more extreme scenarios. The correlation curve is typically increasing, and
therefore it is better known as the skew.
Compound correlation
Compound correlation is the The compound correlation approach is the most natural way of pricing single
correlation directly implied tranche CDOs as each product is valued with a dedicated parameter. It is
from the spread of a CDO
designated in our article from July’04 as ‘implied correlation’, but we feel that the
‘compound’ term has become more of a market standard.
The compound correlation is defined by applying the standard Gaussian copula
model to a given CDO tranche – whether Equity, Mezzanine or Senior. It is defined
for any CDO tranche as the correlation that prices it at its market value.
3
Pricing CDOs with a smile
! This method helps to easily derive prices for non standard tranches as base
correlation curve is monotonic for standard strikes.
… but not for estimating ! The computations of second order greeks (gamma) is not reliable. A mezzanine
second-order risks tranche is viewed as a long/short position between two Equity pieces, both of them
priced with different assumptions. The model sometimes generates unstable P&L
simulations for delta-hedged mezzanine tranches. Based on our observations, this is
particularly true for the 6-9% 5y tranche.
Traders are confronted to a very difficult choice for hedging single tranche CDOs. If
base correlation has become the most natural way of analyzing the sensitivity to the
smile, both models may be used to hedge against spreads movements
It is also important to point that none of this method is coherent. Practitioners have
made adjustments to these methods to price non standard products, taking into
account the smile curve. But these methods may be not be adequate for very
leveraged products (CDOs squared for instance). As in equity derivative markets, new
quantitative frameworks are now needed to model the smile curve and price second
generation products.
Correlation could be a More recent research has focused on making correlation a random variable itself, for
function of the economy
example by making it a function of the economy. In mathematical terms, such models
are known as Gaussian mixtures. A popular assumption has been to change the beta
of each single issuer according to whether the state-of-the-economy factor is above a
given threshold. In this case, the model involves a wide range of parameters for each
issuer-specific thresholds, up and down correlations. It is then relatively simple to
specify a choice for these parameters in order to fit to the few quoted CDO tranches.
This last approach seems to us as the most promising. However, we think that this up
and down correlation structure is not fully realistic. More importantly, it relies on a
wide range of hidden parameters, thereby making the model quite unstable when it
comes to pricing derivative products such as CDO squared. We propose alternatively
to adopt a descriptive approach and to start from market observations to specify the
correlation model.
4
Quantitative Strategy
The key idea of our model is to make correlation a function of the economy. In this
view, it is similar to the up and down model mentioned above. There is a striking
difference, yet: correlation is supposed to be a smooth and homogeneous function of
the economy. A deterioration in economic fundamentals affects all issuers at the
same time and gradually.
Our model is based on a In mathematical terms, we decided to use the same correlation ρ for all names, and to
continuous and
make correlation a function ρ ( X ) of the economy:
homogeneous local
correlation
Aj = −X ρ(X ) + ε j 1 − ρ(X )
In the standard Gaussian copula framework, ρ is constant and does not depend on X .
We do not give any particular form to the local correlation function but we imply it
directly from spreads of liquid iTraxx tranches.
! The second part of this article presents one simple way for deducing local
correlation from base correlation. This first process requires to build first an
extrapolation for the base correlation skew. We derive a closed-form formula for local
correlation within the large-pool approximation.
! The third part details a process for fitting local correlation directly to market data.
This process is better in the sense that it does not require any arbitrary extrapolation
for the base correlation skew.
Local volatility models for equity derivatives… … and local correlation for CDOs
Dupire Large-pool
Interpolate Interpolate
formula formula
Fitting Fitting
Interpolate Interpolate
process process
5
Pricing CDOs with a smile
Local volatility models are Practitioners have coped with this flaw in the model by analysing equity options as
popular for pricing equity
combinations of standard calls and puts, and by pricing these elementary options
derivative products
based on an interpolation of the volatility smile. However, some exotic options such
as American or Bermudean ones resist to such decomposition attempts. Local
volatility models are a simple deviation to the B&S framework, where the volatility of
the stock is a function of the stock price itself (and time, eventually). A decrease in
equity value typically generates an increase in this local volatility, and this
mechanism makes it possible for the model to fit to the volatility smile. When it
comes to pricing exotic options, it is possible to apply the numerical schemes built
within the traditional B&S framework, with minor adjustments only. What is more
difficult here is estimating the local volatility function. There are typically two
alternative and opposite solutions for this:
Local volatility can be ! There is a smart formula (the so-called Dupire formula) for deducing local volatility
estimated in two different
directly from the volatility smile. This requires, as a first step, building a fair
ways
interpolation and extrapolation of the smile, which is no simple task.
! A less elaborate but more robust approach postulates a parametrisation for local
volatility, then fits local volatility directly to market data.
The Gaussian copula model is the equivalent of the Black & Scholes formula in the
CDO universe. The correlation smile is a proof that the popular Gaussian model is
flawed. The base correlation technique is an attempt to split CDOs into elementary
products. Yet, this technique is scarcely a coherent framework for standard bespoke
CDOs, let alone for complex derivative products such as CDO squared.
! In a first step, we focus on the market perception of the distribution of losses. CDO
spreads express a view on how default losses should be distributed. We show how to
adjust the distribution of losses generated by the Gaussian copula model by the
correlation skew in order to get a fair estimate of the distribution of losses implied by
CDO spreads.
! In a second step we make the link between implied distribution of losses and local
correlation. This could be done using a fully numerical technique, but we present a
useful numerical approximation for this. At the limit, on a fully diversified portfolio
with a large number of lines, there is a direct link between conditional default risk on
the one hand, and losses on the portfolio level on the other hand. For a given level of
losses, one can get an estimate of the conditional default risk that triggers such
losses, and ultimately find the local correlation that generates that conditional
default risk.
Therefore, from a given skew of base correlation, one can get an estimate of the
distribution of losses on the portfolio level and thereby an estimate of the local
correlation that generates this distribution.
6
Quantitative Strategy
This implied distribution can For European CDOs (ie. CDOs that pay the premium whatever the outcome on Credit
be estimated by bringing an
Events), there is a smart formula for estimating directly the market-implied
adjustment to the Gaussian
copula model distribution of losses. This formula is a simple adjustment to the Gaussian copula
model (see Appendix A), and we refer to it as the ‘market law of losses’:
L( K ) = L( K , ρ KB ) − Skew( K ) * Rho( K )
where L(K) is the real probability that the cumulative loss of a given basket of entities
does not exceed the attachment point K for a given horizon, ρ KB is the base
correlation associated with the threshold K and L( K , ρ )
B
K is the naïve cumulative
loss function given by the Base correlation methodology. Skew(K) is the slope of the
correlation curve and Rho(K) is the sensitivity of the expected loss of the [0,K] equity
tranche to the change in base correlation.
The adjustment can be quite important. Any coherent model should verify this
equation. We will see in the next sections that this relationship may put some
constraints on the shape of the base correlation curve.
This relationship makes the link between the distribution of losses within the
Gaussian setting used for pricing CDOs, and the distribution of losses really implied
by the whole correlation structure. The Rho is numerically negative, given that an
increase in correlation tends to shift losses away from an Equity piece into more
Senior pieces. The correlation sensitivity is higher for small Equity pieces, and is
maximum for an attachment point around 3%. It converges to zero for very large
Equity pieces.
As highlighted in our article from July’04, the skew itself is positive, at least on the
level of strike quoted by the market, for attachment points above 3%.
Therefore:
The Gaussian copula model ! For a very high attachment point, the Rho nearly vanishes and the Gaussian model
is a valid approximation for is a very good approximation of the expected distribution of credit losses.
very large strikes only
Numerically, this holds for very large attachment points, far above the 12% or even
22% strikes traded on the market.
! For an intermediate attachment point, the Gaussian model is no longer a valid
assumption. The cumulative function of losses, L(K), is above the Gaussian
approximation, which means that the market expects the distribution of defaults to
be more concentrated towards limited losses than what is suggested by the Gaussian
model.
! It is extremely hard to extrapolate what is going on for low attachment points,
below 3%, and this will be analysed in the following of this article.
To illustrate the link between the Gaussian model and the full market-implied
distribution of losses, we have selected the 6% strike as a reference point. A first step
for modeling the distribution of losses would be to use the Gaussian model, with a
correlation given by the base correlation for the 6% strike. Our “market law of losses”
shows that this overestimates the risk that actual losses exceed the 6% threshold.
7
Pricing CDOs with a smile
Looking at the density of credit losses, the line generated by the Gaussian model
stands above the real density of market-implied losses, at least above the 4% strike.
This means that the Gaussian model overestimates the potential of losses above 4%
and underestimates the potential of losses below 4%.
7
95%
6
below this level
90% 5
density
4
85%
3
80% 2
1
75% 0
3% 4% 5% 6% 7% 8% 9% 3% 4% 5% 6% 7% 8% 9%
level of losses level of losses
Base correlation within the This looks counter-intuitive. Given that the smile prices in a higher probability for
Gaussian model
large defaults, one would expect the Gaussian model to actually underestimate losses
underestimates the risk of
limited losses just above a given strike, and not the opposite. Well, simply because the Gaussian
model fails to account for the sharp increase in spreads on high Mezzanine and Senior
tranches, one needs to rely on a high base correlation to fit to market data. This base
correlation is a global indicator. Locally, it is too high and leads to overestimating
losses just above a given threshold.
The ‘large pool’ approximation is often used by practitioners to get fast pricers for
correlation products. This method assumes that the underlying basket is large and
homogeneous enough to be considered as a perfectly diversified portfolio of identical
assets. It provides a sound numerical approximation for theoretical prices, and is an
ideal framework for understanding the behavior of pricing models.
The large pool assumption has a dramatic implication here: knowing the state of the
economy, the loss suffered on the portfolio is simply given by the expected loss on a
single company, handling the portfolio as such single issuer. This implies that,
conditional to the state of the economy, the loss suffered on a large and
homogeneous portfolio is given by the level of conditional default risk.
8
Quantitative Strategy
L-1 designs in mathematical terms the portfolio loss associated to a given confidence
level (a Value-at-Risk). So, the equation above simply states that in a specific state of
the economy, the level of losses on the portfolio is given by the conditional default
risk. Given that our previous section has shown how to estimate the distribution of
losses L, one can switch from base correlations to distribution of losses, and from L to
conditional default risk.
… and conditional default Obviously, correlation plays a key role in this adjustment. A high correlation
risk is linked to correlation
amplifies the impact of the economy on individual risks and therefore on individual
default thresholds. This role is a key point, and it shall be extremely useful to make
the link between conditional default risk and local correlation: for a given state of the
economy and a given level of conditional default risk, one can estimate the
correlation that generates this risk. Therefore, conditional default risk and correlation
are two equivalent variables.
N −1 ( p ) + x ρ ( x)
ε* =
1 − ρ ( x)
Local correlation can be derived from this relationship by finding the roots of a
second order polynomial equation. When there are two possible correlations, we have
selected the solution that makes the local correlation function continuous – if such a
choice is possible. Otherwise, we have selected the smaller of both solutions.
9
Pricing CDOs with a smile
First, for deriving the relationship between base correlation and portfolio losses, we
have made the hidden assumption that base correlation is the level of correlation that
fits to the value of the floating leg of a given Equity piece. This is not entirely true.
CDOs traded on the market (‘American CDOs’) generally decrease the premium paid to
investors after a Credit Event. Base correlations are defined as correlations that fit to
market spreads of traded CDOs, and not to their floating leg. There is a DV01 factor
between the floating leg and the spread of a CDO, and this DV01 is a constant only on
so-called European CDOs. So, our analysis holds for European CDOs.
Second, the large-pool approximation used for making the transition from
distribution of losses into conditional default risk holds only on a large and
homogeneous portfolio. This is hardly the case for real-life CDOs. This approximation
is useful to get a full and intuitive understanding of the mechanism, and as we will
see later in this article, it can be used as a starting point for a more precise numerical
scheme.
Using the two-step approach detailed earlier in this article, one can take a base
correlation skew as an input and build an estimate of the local correlation. Oddly
enough, not all base correlation skews generate a continuous local correlation
function. The problems stems from the choice for extrapolated correlation on the
[0.3%] bucket.
A linear extrapolation fails For example, a linear extrapolation generates a discontinuous local correlation. The
to generate a continuous
last step in the estimation process makes the link between individual risk and local
local correlation
correlation, and involves solving a second-order equation. At some point when the
economy deteriorates, two possible correlations emerge for one level of the
conditional default risk. None of these correlations are compatible with the local
correlation estimated under a better economy. This hints that the linear extrapolation
is not a right choice below 3%.
10
Quantitative Strategy
40%
Base correlation
120%
35% 100%
30% 80%
25% 60%
40%
20%
20%
15%
0%
10% -20%
bad economic conditions
0% 2% 4% 6% 8% 10% -3.9 -3.1 -2.3 -1.5 -0.7 0.1 0.9 1.7 2.5 3.3
Strike
We have found empirical evidence that base correlations should be decreasing for
very low strikes to give a continuous local correlation curve. The next section
concludes with a full analysis and theoretical background for this pattern.
40%
Base correlation
120%
100%
35%
80%
30% 60%
40%
25%
20%
20% 0%
0% 2% 4% 6% 8% 10% -3.9 -3.1 -2.3 -1.5 -0.7 0.1 0.9 1.7 2.5 3.3
bad economic conditions
Strike
Local correlation should be For using the model in practice, we propose to go the other way round. Instead of
fitted directly to market data
extrapolating the base correlation skew and then building the local correlation, we
first specify a shape for the local correlation and then get CDO spreads generated by
this assumption. The local correlation is then adjusted numerically to fit to market
spreads on CDO tranches.
This approach avoids any trouble with discontinuous local correlations. It really
makes local correlation the centerpiece of the process. One may say that this merely
moves the extrapolation problem from one point to another. It is our belief that fitting
11
Pricing CDOs with a smile
One should give a concrete However, finding a correct specification for the local correlation function is far from
meaning to the state-of-the-
obvious. Local correlation is a function of the risk factor within the model, a factor
economy factor
that is interpreted as the state of the economy. It is at first sight difficult to interpret
this variable from the standpoint of a trader or of a portfolio manager. This is the
purpose of this third part to give a market interpretation to local correlation.
We propose a mapping of the First, we introduce a new notion, the marginal compound correlation. This is defined
economy into a strike
as the compound correlation of very tiny CDO pieces. Second, we exhibit a mapping of
the state-of-the economy factor into a strike, such that local correlation becomes
equivalent to marginal compound correlation. In other words, it is possible to
associate to each strike a specific state of the economy, and the local correlation for
this state of the economy is given by the compound correlation of a tiny CDO tranche
around the strike. Third, we use that mapping for building local correlation as a
function of attachment points instead of the economy. This provides a complete
interpretation of local correlation from a market standpoint, and leads the way for a
simple and intuitive process for fitting the model.
Such a tiny CDO tranche can be viewed as a product that triggers a 100%-loss should
default losses exceed the strike, and involves no protection payment otherwise. It
should be noted that the exercise probability of such a tranche can be expressed
through two different ways:
! It is the market-implied probability that default losses exceed the strike. As we have
seen earlier within this article, the ‘market law of losses’ makes it possible to estimate
this cumulative risk from market data.
! The marginal compound correlation is by definition the correlation that prices the
tiny CDO tranche at its market spread within the traditional constant-correlation
model.
Marginal compound Therefore, the market-implied cumulative function of losses is equal to the traditional
correlation prices in the
constant-correlation cumulative risk, using the marginal compound correlation as a
cumulative function of
losses correlation input. This is a very remarkable feature, because it provides a simple way
to express the ‘real’ distribution of losses using the traditional Gaussian model. It is
even simpler than our ‘market law of losses’, because it involves estimates neither of
the skew nor of the Rho. However, it relies on the compound correlation of a tiny CDO
tranche, a variable that can not be directly observed on the market.
Anyway, that is not a concern to us for the moment. The purpose of the analysis here
is making the link between local and marginal compound correlation. This is
something we shall do within the large-pool approximation. Under the assumption of
a large and homogeneous portfolio, there is a simple formula linking the cumulative
function of losses to the compound correlation (see for example Appendix B). The
situation is trickier when it comes to local correlation, but we have seen in the second
part of this article a two-step process for linking the cumulative function of losses to
12
Quantitative Strategy
local correlation. Comparing both expressions of the cumulative function, one can
prove (see Appendix C) that local correlation and marginal compound correlation are
two equivalent variables, provided each strike is mapped properly into a ‘state of the
economy’.
Using the right mapping… Mathematically, for each strike K, this state of the economy is such that N(xK) = L(K), or:
x K = N −1 o L( K )
… local correlation is a This is exactly the right mapping for comparing local to marginal compound
marginal compound
correlations:
correlation
ρ [x K ] = ρ KM
where ρM is the marginal compound correlation at a given strike.
This is the cornerstone of our approach. For each state of the economy, local
correlation is the correlation of a tiny CDO tranche at a strike that is exactly the level
of losses on the portfolio under that state of the economy. For each strike, one can
also estimate the state of the economy that hits the strike, and associate to this
attachment point a local correlation.
13
Pricing CDOs with a smile
Local correlations mapped with CDO strikes look close to implied compound
correlations
50%
45%
40%
35% Local Correlation
30%
Base Correlation
25%
20%
Implied Compound
15% Correlation
10%
5%
0%
0% 2% 4% 6% 8% 10% 12%
The local correlation is generated from an extrapolated base correlation skew. The base correlation has been extrapolated using
second-order splines, under the constraint that derivatives on the 0% strike shall be zero. This constraint generates indirectly a
stabilisation of the local correlation close to the 0% strike.
Our process for estimating This process, albeit simple, is highly dependent on our specification for the base
local correlation…
correlation on low strikes. We have already discussed this issue. One smart solution is
to invert the process: specify a possible shape for the local correlation, then estimate
the parameters for local correlation directly on market data. This leads us to the
following steps:
All the results in the last section have been deduced under the large pool assumption.
It does not seem possible to build a simple one-to-one relationship between local and
marginal compound correlation in the general case. Nevertheless we think that the
mapping of the economy into a strike helps find a parametric local correlation in a
more natural and easy way. Moreover, our study gives a rough but good idea of the
slope of local correlation and highlights its proximity with the compound correlations
observed in the market. It seems to us that the finite size of the portfolio and the
dispersion effect introduced by the heterogeneousness of the entities do not create a
major change in the structure of the local correlation.
In other words, the mapping of the economy into a strike (step 3 within our
estimation process) is based on the direct formula highlighted in the previous
14
Quantitative Strategy
There are multiple choices ! The first part deals with very low strikes. Local correlation is supposed to decrease
for the shape of local
below the upper attachment point of the smaller liquid equity piece. Correlation is
correlation…
supposed to be linear on this side of the curve.
! The second part of the curve focuses on mezzanine and senior tranches, where four
instruments are quoted and may be used to calibrate the curve. We chose an
increasing second order polynomial to describe this part of the market. We also added
a constraint to our model to enforce the continuity of the curve.
… but the crucial point is the There are five market inputs: CDO spreads for 0-3%, 3-6%, 6-9%, 9-12% and 12-22%
mapping of the economy
tranches. Our choice for the local correlation involves four parameters. They are
into a strike
estimated by minimising the quadratic distance between model and market spreads.
Our four-parameter model fits well to tranched iTraxx prices and captures the slope
and convexity observed on the base correlation market. The model is easily fitted to
market data, because it is well specified and is close to empirical evidence. In
conclusion, the factor-to-strike transformation has a very substantial value added in
the estimation process.
60%
50%
40%
30%
20%
10%
0%
0% 5% 10% 15% 20% 25%
One could alternatively use a piecewise linear local correlation after the strike
transformation. Such an approach could fit any form of base correlation curve.
In conclusion, the study carried out under the large-pool approximation provides a
simple and efficient parametrisation of the local correlation. Our model successfully
fits to market prices and can now be used to derive valuations for more complex
products.
The model inherits from the Finally, it should be noted that even if estimating the local correlation function is a
numerical tractability of the
crucial and difficult point, pricing CDOs based on a given local correlation is
one-factor Gaussian
framework straightforward. The one-factor Gaussian copula model is famous for its numerical
tractability, and our local correlation model inherits from this attractive feature.
Making correlation a function of the economy is a very minor change to the powerful
numerical schemes that have been developed under the Gaussian model. Therefore,
15
Pricing CDOs with a smile
our approach avoids the burden of building once again a series of pricers for
correlation products.
Skew( K ) * Rho( K )
ρ KM ≈ ρ KB −
∂L
( K , ρ KM )
∂ρ
where L( K , ρ ) is the cumulative loss function used in the standard Gaussian copula
approach. This provides yet another formula for linking base to local correlation.
More interestingly, this formula highlights a few variables that determine the level of
local against base correlation:
! The Rho is the sensitivity of expected loss to the strike. As mentioned in the first
part of this article, it is numerically negative.
! The Skew is usually positive for sufficiently high strikes. At stake is how it should
be extrapolated on lower strikes.
There is a strike for which ! The cumulative function of losses in the traditional Gaussian model decreases with
the cumulative function of
correlation, at least for sufficiently high strikes. This is intuitive: a higher correlation
losses changes its
sensitivity to correlation… increases the risk of a catastrophe scenario, and therefore makes it less likely for
losses to remain limited. The point is that below a specific strike, this relationship
becomes inverted. A higher correlation increases the dispersion of potential losses
around the expected loss, and therefore increases the likelyhood of very limited
losses, too.
Below a specific strike, the risk of limited losses A skew without a smile generates a discontinuity in local
increases with correlation correlation
60%
loss to base correlation
150%
50%
Correlation
40%
100%
30%
20%
50% Change in
sensitivity 10%
0% 0%
0% 5% 10% 15% 20% 25% 0% 5% 10% 15% 20% 25% 30%
-50% Strike
Strike
The purpose of our analysis here is understanding how local correlation compares to
base correlation.
! For a tiny Equity tranche, with a strike close to zero, base and local correlation are
just equivalent variables
16
Quantitative Strategy
! For a very large strike, the Rho vanishes and therefore base and local correlation
become close to each other again.
Right in the middle, the relationship between both correlations depends on the sign of
the sensitivity of the cumulative loss to correlation. This sensitivity is the
denominator within the relationship above. For any correlation level, there is a strike
where the cumulative function of losses in the traditional Gaussian model is no
longer sensitive to correlation.
! Above that strike, the cumulative function of losses increases with correlation.
Given that the skew is usually positive at such levels of strike, the relationship above
implies that local correlation is lower than base correlation.
… and the skew must be flat ! Exactly at this strike, the denominator in the relationship moves to zero. If the skew
at this level of strike…
is not zero at this point, local correlation diverges to arbitrarily high numbers. This, in
turn, generates a discontinuity in the local correlation. Therefore, the skew must be
flat at the strike where the cumulative function changes its sensitivity to correlation.
… generating a smile in the ! Below this strike threshold, sensitivity becomes negative. Exactly at the threshold,
base correlation curve
local correlation stand s below base correlation. By continuity, it stays so below the
threshold, too. As a result, the skew must be negative if the relationship above is to be
met.
17
Pricing CDOs with a smile
technique quite reliable – provided the underlying portfolio is indeed an index with
liquid CDO tranches.
By contrast, our local correlation model is ideally designed for pricing CDO squared.
Fitting the model to a portfolio of inner CDOs and building adequate numerical
schemes for CDO squared is beyond the scope of this article, and will be the focus of a
further publication.
There has been some debate on how base correlation should be adjusted from an
index into a portfolio with a higher spread. Within our framework, it is more natural to
analyse the impact of a change in portfolio on the marginal compound correlation.
The transformation writes:
ρ M ( K ) = ρ Index
M
o L−Index
1
o L( K )
In the table below we give a list of the pricings of different baskets with different
portfolios using the base and the local correlation method. All the spreads from the
standard iTraxx portfolio (with an average premium of 40bp) have been shifted to
give the portfolios more or less risk.
18
Quantitative Strategy
Our local correlation model Results are significantly different, which is not surprising as methods have different
generates a more stable
definitions of subordination. The dynamics of subordination are less pronounced with
marked-to-market
the local correlation curve as it takes into account the convexity of the loss function.
Deltas computed by the local correlation model take into account this roll-down effect.
Below is a table which compares leverages computed with the compound, base and local
correlation approaches.
… leading to lower deltas We see that the results are substantially different. The gap between the compound and
the local correlation approach is due to the roll-down effect. This effect depends
directly on the sensitivity to the compound correlation but also on the slope of the
compound correlation curve and the sensitivity of the strike to changes in individual
spreads. We obtain:
∂K
∆ = ∆ Im plied + Rho * slope *
∂Spreads
! The correlation sensitivity (the Rho) tends to be negative on mezzanine and senior
tranches and positive on equity tranches.
! The slope of the compound correlation curve is positive for high mezzanine and
senior tranches and negative for equity pieces.
! The sensitivity of the strike to the market changes in spreads depends closely on
the definition of the subordination. As we have seen, market players today define the
subordination as the relative value of a strike to the portfolio expected loss whereas
our model uses directly the law of losses of the portfolio basket. This sensitivity is
negative for mezzanine and senior tranches and positive for equity pieces.
19
Pricing CDOs with a smile
Local volatility models Taking into account the roll-down on the smile, the leverage decreases after
exhibit a similar pattern
correction against the roll-down for mezzanine and senior tranches, and increases for
equity pieces. This adjustment is very similar to the adjustments made on equity
derivative products. Equity traders usually adjust the Black & Scholes delta by the
product of the vega of the option by the slope of the volatility smile. This usually
leads to lower hedge ratios compared to the standard B&S formula. In these markets,
the local volatility model manages to capture this roll-down effect on the volatility
curve.
For each day since July’04, we gathered the actual 3-6% spread and computed the
compound correlation of the tranche. We then computed a theoretical 3-6% spread
assuming that global CDS spreads had unchanged since July‘04. It gave us a 3-6% spread
adjusted for iTraxx movements. One would expect this adjusted 3-6% spread to be
negatively correlated with the index. The following graph shows quite the opposite.
195 48
190 46
185 44
42
180
40
175
38
170
36
165 34
160 32
155 30
7/2/2004 9/2/2004 11/2/2004 1/2/2005
The 3-6% spread is adjusted for changes in index spread. It only reflects changes in compound correlation.
Source: SG Credit Research
… but flows on the We think that this positive correlation has been heavily driven by flows on credit
correlation market tend to
structured markets. In the midst of a CDO issuance wave, exotic traders typically hedge
drive CDO spreads the other
way round their positions both buy selling protection on indices and on index mezzanine tranches,
putting pressure at the same time on index spreads and on tranche-specific tranches.
Nevertheless, this empirical evidence shows than deltas on mezzanine tranches are
underestimated by standard copula models and by our local correlation approach, too.
This phenomenon is well known on equity derivatives markets and traders actually
add a second adjustment to the Black and Scholes deltas to take into account the
negative correlation between stocks and implied at-the-money volatilities. This
adjustment, which increases the hedge ratios, is not factored in by the Dupire model
20
Quantitative Strategy
and this is, in the view of many equity derivatives traders, an important drawback of
the model.
A successful hedging Similarly our local correlation model does not take into account the historical
strategy should incorporate
correlation between compound correlations and spreads as their dynamic is simply
an adjustment against the
statistical link between not depicted. A new rule of thumb could however help build a rough adjustment.
indices and CDO correlation
∂K
∆ = ∆ Im plied + Rho * slope * + β * Rho
∂Spreads
Conclusion
Our local correlation model is a major breakthrough in modeling correlation
products. It provides a coherent framework for pricing various tranches on a given
portfolio. Moreover, we have shown how to build an estimate of the market-implied
distribution of losses from a basis correlation skew. This relationship proves to be
crucial when it comes to pricing bespoke CDOs. We exhibit a rule for switching from
index into bespoke correlation: attribute the same (local) correlation to strikes that
have the same risk of being hit.
If the local correlation model fits perfectly well to market data and provides an
attractive framework for pricing second generation products, it does not factor in all
the phenomena occurring on correlations markets. In particular, the empirical
correlation between credit spreads and compound correlations is a challenge to any
modeling attempt.
The main purpose of the model is to provide an integrated framework for pricing
CDOs, with a view on relative value. Further research will analyse the shape of local
correlation across various indices. Our aim is to exhibit common units and constants
across which various CDOs could be compared. Further research should focus on
correlation discrepancies between reference indices and bespoke CDOs, and the use
of statistical correlation matrices for refining pricing estimates.
21
Pricing CDOs with a smile
Appendix
Appendix A: the market law of losses
Defining LK the expected loss of the [0,K] equity tranche, L(K ) the cumulative loss
function.
K K
LK = ∫ xdL( x) + K (1 − L( K )) = K − ∫ L( x)dx
0 0
∂ LK
and therefore, = 1 − L( K ) .
∂K
K K
LK = ∫ xdL( x, ρ KB ) + K (1 − L( K , ρ KB )) = K − ∫ L( x, ρ KB )dx
0 0
and therefore:
∂ LK ∂ρ B K
∂L
= 1 − L( K , ρ KB ) − K ∫ ∂ρ ( x, ρ
B
)dx
∂K ∂K
K
0
∂L
K
= 1 − L( K , ρ KB ) − Skew( K , ρ KB ) ∫ ( x, ρ KB )dx
0
∂ρ
∂ LK K
∂L
but Rho( K , ρ KB ) =
∂ρ K
B
= − ∫
0
∂ρ
( x, ρ KB )dx
L( K ) = L( K , ρ KB ) − Skew( K , ρ KB ) * Rho( K , ρ KB )
Appendix B: the large pool approximation
Under the large pool assumption, all the names have an identical default probability
function p (t ) .
∑1{− X }
N
ρ ( X ) + ε j 1 − ρ ( X ) ≤ N −1 ( p (t ) ) ≤ K
1
L( K ) = Q (1 − δ )
N j =1
Conditionally, to each value of the systemic factor X we can apply the law of large
N −1 ( p ) + x ρ ( x)
Note ε ( x) =
*
numbers. the individual default threshold
1 − ρ ( x)
conditional to the economy.
22
Quantitative Strategy
{ (
L( K X ) ≅ 1 (1 − δ )Q − X ρ ( X ) + ε j 1 − ρ ( X ) ≤ N −1 ( p(t ) ) ≤ K ) }
(
= 1 N ε * ( X ) ≤ )
K
1−δ
= Q N (ε * ( X ) ) ≤
K
We therefore obtain L( K ) , and:
1−δ
K −1 K
L( K ) = Q ε * ( X ) ≤ N −1 ( ) = N o ε * o N −1 ( )
1−δ 1−δ
1 −1
We can invert this relation and get ε * ( x) = N −1 L o N ( x) which is the final
1− δ
result.
Let us note that in the case where ρ is a constant, the h function can be inverted. There
is then a simple formula for the cumulative function of the loss distribution:
N −1 (K /(1 − δ ) ) 1 − ρ − N −1 ( p (t ) )
L( K , ρ ) = N
ρ
Appendix C: the relationship between compound and local correlation
Define ρ KM as the marginal compound correlation. We have seen in the thord part of
this article that the market-implied cumulative function of losses can be estimated by
applying the constant-correlation model and the marginal compound correlation as
an input: L( K , ρ KM ) = L( K )
We can apply the large pool approximation on the standard static copula model and
obtain:
N −1 (K /(1 − δ ) ) 1 − ρ M − N −1 ( p (t ) )
L( K ) = N K
ρK M
We can on the other hand apply the change of variable x ⇔ N −1 o L( K ) in the large
pool approximation in the local correlation model (Appendix B):
[
N −1 ( p (t ) ) + N −1 o L( K ) × ρ N −1 o L( K ) ]=N −1
( K /(1 − δ ))
1 − ρ [N o L( K )]
−1
L( K ) = N
[ ]
N −1 (K /(1 − δ ) ) 1 − ρ N −1 o L( K ) − N −1 ( p (t ) )
for all K,
ρ N −1
o[L ( K ) ]
23
Pricing CDOs with a smile
March 2001 How does the market reward the risks of default or rating downgrade ?
April 2001 A rule-of-thumb for valuing Telecom coupon step-ups
May 2001 What are the incentives to use internal ratings in the new Cooke ratio ?
June 2001 Understanding the dynamics of Credit Spreads
July 2001 A smile model for valuing capital securities
September 2001 Looking for value on the European Credit Market
September 2001 Pricing put options indexed on ratings
October 2001 Why can a credit curve invert ?
November 2001 Some applications of a firm value model
January 2002 Beyond volatility and correlation
February 2002 Cash-flow CDOs – highly enhanced assets
March 2002 A spread forecasting model
April 2002 CDOs as balance sheet management tools
May 2002 How to optimize a credit portfolio
June 2002 An analysis of CDO downgrades
July 2002 Backtesting systematic Rich/Cheap strategies
September 2002 Firm models and trading srategies (1)
October 2002 Firm models and trading strategies (2)
December 2002 2002: of useful models
March 2003 Forecasting individual corporate credit spreads
April 2003 Arbitrage between CDS and bonds
May 2003 Beta-neutral relative value analysis
June 2003 Firm models with event risk
July 2003 Hedging bonds using CDS and taking advantage of convexity
August 2003 Pricing the basis
September 2003 Arbitrage between credits and equities – new strategies, more opportunities
October 2003 Kohonen maps: going beyond classification for enhanced management strategies
December 2003 Forecasting the iBoxx Diversified CDS index
January 2004 Backtesting basis trades: a case study
March 2004 Back to basics: which spread for measuring credit ?
April 2004 A hedging model for capital structure arbitrage
Juiy 2004 Pricing and hedging of correlation products
July 2004 The new iTraxx indices
September 2004 CDS vs Stock – the quest for the optimum hedge ratio
October 2004 Empirical evidence on the BFM model
November 2004 Introducing the CDS Curves Monitor
February 2005 The relative value of EDS against credit and equity derivatives
February 2005 Pricing CDOs with a smile
24
Quantitative Strategy
Credit Research
HEAD OF RESEARCH Benoît Hubaud (33) 1 42 13 61 08 benoit.hubaud@sgcib.com
(44) 20 7676 7168
FINANCIALS European banks & Insurance Robert Montague (44) 20 7676 7062 robert.montague@sgcib.com
Richard Thomas (44) 20 7676 7030 richard.thomas@sgcib.com
Matthieu Loriferne (44) 20 7676 7167 matthieu.loriferne@sgcib.com
CORPORATES
Autos Pierre Bergeron (33) 1 42 13 89 15 pierre.bergeron@sgcib.com
Automobile, Autoparts, Transportation Stéphanie Herrault (33) 1 42 13 63 11 stephanie.herrault@sgcib.com
STRATEGY
Credit Suki Mann (44) 20 7676 7063 suki.mann@sgcib.com
Guy Stear, CFA (44) 20 7676 7158 guy.stear@sgcib.com
The information herein is not intended to be an offer to buy or sell, or a solicitation of an offer to buy or sell, any securities and including any expression of opinion, has been obtained
from or is based upon sources believed to be reliable but is not guaranteed as to accuracy or completeness although Société Générale (“SG”) believe it to be fair and not misleading or
deceptive. SG, and their affiliated companies in the SG Group, may from time to time deal in, profit from the trading of, hold or act as market-makers or act as advisers, brokers or bankers
in relation to the securities, or derivatives thereof, of persons, firms or entities mentioned in this document or be represented on the board of such persons, firms or entities. Employees
of SG, and their affiliated companies in the SG Group, or individuals connected to then, other than the authors of this report, may from time to time have a position in or be holding any of
the investments or related investments mentioned in this document. Each author of this report is not permitted to trade in or hold any of the investments or related investments which are
the subject of this document. SG and their affiliated companies in the SG Group are under no obligation to disclose or take account of this document when advising or dealing with or for
their customers. The views of SG reflected in this document may change without notice. To the maximum extent possible at law, SG does not accept any liability whatsoever arising from
the use of the material or information contained herein. This research document is not intended for use by or targeted at private customers. Should a private customer obtain a copy of
this report they should not base their investment decisions solely on the basis of this document but must seek independent financial advice.
IMPORTANT DISCLOSURES: Please refer to our website: http:\\www.globalmarkets.sgcib.com
This publication is issued in France by or through Société Générale ("SG”) which is regulated by the AMF (Autorité des Marchés Financiers).
Notice to UK Investors: This publication is issued in the United Kingdom by or through Société Générale ("SG"). All materials provided by SG Commodity Research and SG Global
Convertible Research are produced in circumstances such that it is not appropriate to characterize them as impartial as referred to in the Financial Services Authority Handbook. However,
it must be made clear that all research issued by SG will be fair, clear and not misleading. In the United Kingdom, SG Cowen & Co. is a Trading Name of SG. SG is a Member of the
London Stock Exchange.
Notice to US Investors: This report is issued solely to major US institutional investors pursuant to SEC Rule 15a-6. Any US person wishing to discuss this report or effect transactions in
any security discussed herein should do so with or through SG Americas Securities, LLC or SG Cowen & Co., LLC to conform with the requirements of US securities law. SG Americas
Securities, LLC, 1221 Avenue of the Americas, New York, NY, 10020. (212) 278-6000. SG Cowen & Co. LLC, 1221 Avenue of the Americas, New York, NY, 10020. (212) 278-4000. Some
of the securities mentioned herein may not be qualified for sale under the securities laws of certain states, except for unsolicited orders. Customer purchase orders made on the basis of
this report cannot be considered to be unsolicited by SG Americas Securities, LLC or SG Cowen & Co., LLC and therefore may not be accepted by SG Cowen investment executives unless
the security is qualified for sale in the state.
Analyst Certification: Each author of this research report hereby certifies that (i) the views expressed in the research report accurately reflect his or her personal views about any and all
of the subject securities or issuers and (ii) no part of his or her compensation was, is, or will be related, directly or indirectly, to the specific recommendations or views expressed in this
report.
Notice to Japanese Investors: This report is distributed in Japan by Société Générale Securities (North Pacific) Ltd., Tokyo Branch, which is regulated by the Financial Services Agency of
Japan. The products mentioned in this report may not be eligible for sale in Japan and they may not be suitable for all types of investors.
Notice to Australian Investors: Société Générale Australia Branch (ABN 71 092 516 286) (SG) takes responsibility for publishing this document. SG holds an AFSL no. 236651 issued
under the Corporations Act 2001 (Cth) ("Act"). The information contained in this newsletter is only directed to recipients who are wholesale clients as defined under the Act.
hhtp://www.sgcib.com. Copyright: The Société Générale Group 2005. All rights reserved.