Paola Mosconi
Banca IMI
The opinion expressed here are solely those of the author and do not represent in
any way those of her employers
Brigo, D. and Mercurio, F. Interest Rate Models Theory and Practice. With
Smile, Inflation and Credit, Springer (2006)
Brigo, D. (2009), Essex Lecture Notes, Unit 5
Outline
1 Introduction
Securitization
Of Models and Mathematicians
2 Dependence
Credit Correlation
Dependence in Reduced Form Models
Copula Function
3 CDOs: Stylized Facts
Stylized Facts
4 CDO Pricing
Monte Carlo Pricing
One Factor Gaussian Copula
Large Homogeneous Portfolio (LHP)
5 Market Quotation Standard
Compound Correlation
Base Correlation
6 Selected References
Paola Mosconi Lecture 6 5 / 69
Introduction
Introduction
In the period 1998 to 2007, the asset backed securities market increased expo-
nentially both in volume and diversity.
As the crisis unfolded in 2007/2008, such market came under substantial criticism
as some securitized products played a major role in the financial difficulties for
various reasons (see Prime Collateralized Securities):
badly underwritten products
opaqueness of structures
overleveraged issuances
...
Despite the low issuance and the modest take-up by investors, most European structured
finance products performed well throughout the financial crisis from a credit standpoint,
with low realized default rates.
According to Standard & Poors, the cumulative default rate on European consumer
related securitizations, between the start of the financial downturn in July 2007 and Q3
2013 has been only 0.05%.
Discussion Paper by the European Central Bank and the Bank of England (July, 2014).
A framework for securitisation is the first major building block of the EUs plan, launched
during 2015, to develop a fully functioning capital markets union by the end of 2019.
Developing a securitisation market will help create new investment possibilities and provide
an additional source of finance, particularly for SMEs and start-ups.
Figure: European securitization outstanding (left) and issuance (right). Source: BOE and
ECB (2014).
vs
Did a mathematical formula really blow up Wall Street? (Embrechts, 2009)
Dont blame the quants (Shreve, 2008)
Crash Sonata in D Major (Szego, 2009)
Credit Models and the Crisis, or: How I learned to stop worrying and love the
CDOs (Brigo et al, 2010)
Figure: Source: Recipe for Disaster: The Formula that Killed Wall Street. Wired
Magazine (2009).
Undoubtedly, Lis formula has severe flows of which mathematicians and quants
were well aware even before the crisis.
The most dangerous partMr. Li himself says of the model, is when people
believe everything coming out of it. [...] Very few people understand the essence
of the model. [...] Its not the perfect model.But, he adds: Theres not a better
one yet.
To Stanfords Mr. Duffie, The question is, has the market adopted the model
wholesale in a way that has overreached its appropriate use? I think it has.
Two ways:
1 The whole distribution is simulated (Monte Carlo)
2 Single name marginal distributions + dependence structure = copula
Outline
1 Introduction
Securitization
Of Models and Mathematicians
2 Dependence
Credit Correlation
Dependence in Reduced Form Models
Copula Function
3 CDOs: Stylized Facts
Stylized Facts
4 CDO Pricing
Monte Carlo Pricing
One Factor Gaussian Copula
Large Homogeneous Portfolio (LHP)
5 Market Quotation Standard
Compound Correlation
Base Correlation
6 Selected References
Paola Mosconi Lecture 6 17 / 69
Dependence Credit Correlation
Credit Correlation I
Default Correlation
Given a time horizon T (typically one year) the default correlation between two
names can be expressed in terms of:
their marginal default probabilities q1 = E[1{1 <T } ] and q2 = E[1{2 <T } ]
their joint default probability q12 = E[1{1 <T } 1{2 <T } ]
as follows:
q12 q1 q2
12 = p
q1 (1 q1 )q2 (1 q2 )
Credit Correlation II
Table: Asset correlation (estimated) vs. default correlation. Source: Frey et al (2001).
In the framework of reduced form (intensity) models, the default time is the first jump
of a Poisson process with intensity (t):
where is the intensity or hazard rate and represents an instantaneous credit spread.
1 = 1 1 1
1 (1 ) , 2 = 2 (2 ) , . . . , n = n (n )
1 put dependence in (stochastic) intensities of the different names and keep the of
different names independent
2 put dependence among the of different names and keep the intensities (either
stochastic or deterministic) independent. This is the approach currently used for
correlation products in the market
3 put dependence both among the and the intensities of different names
Linear correlation is not enough to express the dependence between two random variables.
In credit derivatives with intensity models, dependence must be introduced in the expo-
nential components of the Poisson processes for different names.
This is usually done by means of Copula functions.
Definition
Let (U1 , . . . , Un ) be a random vector with uniform margins and joint distribution
C (u1 , . . . , un ). C (u1 , . . . , un ) is the copula of the random vector.
Sklars Theorem
Let H be an n-dimensional distribution function with margins F1 , . . . , Fn . Then, there exists
an n-copula C (i.e. a joint distribution function on n uniforms) such that for all x Rn
Gaussian Copula
The Gaussian (or Normal) copula plays a central role in the modeling of credit dependence.
Gaussian Copula
It is obtained by using a multivariate normal distribution NRn with standard Gaussian
margins and correlation matrix R as multivariate distribution H:
Properties
No closed form expression, except for n = 2.
For n names, the correlation matrix R has n(n 1)/2 free parameters.
No tail dependence (upper/lower).
C (u, v ) = C (v , u) i.e. exchangeable copula.
F () = 1 e = U = = ln(1 U)
Goal
Model dependence across default times 1 , . . . , n , by introducing dependence directly
among standard uniforms!
1 = 1 1
1 ( ln(1 U1 )), . . . , n = n ( ln(1 Un ))
Example Reloaded
P(U1 u1 , U2 u2 ) = min(u1 , u2 )
Consider that:
Therefore:
q.e.d.
Outline
1 Introduction
Securitization
Of Models and Mathematicians
2 Dependence
Credit Correlation
Dependence in Reduced Form Models
Copula Function
3 CDOs: Stylized Facts
Stylized Facts
4 CDO Pricing
Monte Carlo Pricing
One Factor Gaussian Copula
Large Homogeneous Portfolio (LHP)
5 Market Quotation Standard
Compound Correlation
Base Correlation
6 Selected References
Paola Mosconi Lecture 6 29 / 69
CDOs: Stylized Facts Stylized Facts
The portfolio loss distribution is not symmetric but shows the following features:
(Synthetic) CDOs
Synthetic CDOs with maturity T are obtained by pooling together CDSs of different
names (up to n) with the same maturity and tranching the total loss:
n
X n
X
Loss(T ) = LGDi 1{i T } = (1 Reci ) 1{i T } (2)
i =1 i =1
along two attachment points A and B, with A < B. The protection seller pays the
protection buyer the cumulated tranched loss that exceeds A and does not exceed B.
The (percentage) tranched loss between attachment points A and B at time t is given
by:
1
LosstrA,B (t) :=
(Loss(t) A)1{A<Loss(t)B} + (B A)1{Loss(t)>B}
B A
1
= (Loss(t) A)1{A<Loss(t)} (Loss(t) B)1{Loss(t)>B} (3)
B A
1
(Loss(t) A)+ (Loss(t) B)+
=
B A
or, in a compact notation:
0 if Loss(t) < A
Loss(t)A
LosstrA,B (t) = BA
if A < Loss(t) B
1 if Loss(t) > B
The tranche [0, X ] absorbs the first losses up to X % of the total portfolio loss and it is
called equity tranche.
The corresponding tranched loss, according to (3) with A = 0 and B = X , is given by:
1
LosstrX (t) := Losstr0,X (t) = Loss(t) (Loss(t) X )+
X
The contract consists of two legs: the default leg and the premium leg.
Schematically the cash flows of a CDOs contract can be summarized as follows:
where:
dLosstrA,B (t) is the tranched loss increment at time t
the outstanding notional is given by the survived positive (re-scaled) notional at
the relevant payment time:
A,B
A premium rate R0,T (0) fixed at time T0 = 0 is paid at times T1 , . . . , Tb = T from the
protection buyer to the protection seller. The rate is paid on the survived positive notional
at the relevant payment time. This notional decreases of the same amount as the tranched
loss increases, taking into account the recovery.
b
X
A,B
R0,T (0) D(0, Ti ) i [1 LosstrA,B (Ti )]
i =1
where i = Ti Ti 1 .
Once enough names have defaulted and the loss has reached A, the counts start. Each
time the loss increases, the corresponding loss change re-scaled by the tranche thickness
B A (i.e. dLosstrA,B (t)) is paid to the protection buyer, until maturity arrives or until the
total pool loss exceeds B, in which case the payments stop.
Z T b
X
ProtLA,B (0) = D(0, t)dLosstrA,B (t) D(0, Ti )[LosstrA,B (Ti ) LosstrA,B (Ti 1 )]
0 i =1
Assuming deterministic interest rates, the price of the tranche for the protection buyer
is:
TrancheA,B
0,T (0) = E[ProtLA,B (0)] E[PremLA,B (0)]
b
X
= P(0, Ti ) E[(LosstrA,B (Ti ) LosstrA,B (Ti 1 ))]
i =1 (5)
b
X
A,B
R0,T (0) P(0, Ti ) i [1 E(LosstrA,B (Ti ))]
i =1
where E[.] denotes the expectation under the risk neutral measure.
The premium rate that makes the contract fair at inception is therefore given by:
Pb tr tr
A,B i =1 P(0, Ti )[E[LossA,B (Ti )] E[LossA,B (Ti 1 )]]
R0,T (0) = Pb tr
(6)
i =1 P(0, Ti )i [1 E(LossA,B (Ti ))]
Outline
1 Introduction
Securitization
Of Models and Mathematicians
2 Dependence
Credit Correlation
Dependence in Reduced Form Models
Copula Function
3 CDOs: Stylized Facts
Stylized Facts
4 CDO Pricing
Monte Carlo Pricing
One Factor Gaussian Copula
Large Homogeneous Portfolio (LHP)
5 Market Quotation Standard
Compound Correlation
Base Correlation
6 Selected References
Paola Mosconi Lecture 6 38 / 69
CDO Pricing
CDO Pricing
Goal
The problem of pricing a CDO tranche or calculating its premium rate consists in
calculating the corresponding expected tranched loss.
Pricing a CDO by means of Monte Carlo goes through the following steps:
3 Consider N scenarios and for each scenario j, simulate the dependent default
times through the copula approach outlined in slide 27.
4 At any given time t and for any scenario j, compute the loss and tranched loss
given respectively by eq. (2) and eq. (3).
5 Average across all scenarios in order to find the expected tranched loss
E[LosstrA,B (t)] at the desired date t or set of dates Ti .
Monte Carlo pricing is conceptually straightforward, but has two main limitations:
Semi-Analytical Methods
In order to overcome the limitations of the Monte Carlo approach, alternative methods
have been proposed, which rely on the semi-analytical computation of the portfolio
loss. One popular method combines:
the One Factor Gaussian Copula approach to calculate the joint default probability
the Large Homogeneous Portfolio (LHP) approach to calculate the portfolio loss
A one factor copula model is a way of modeling the joint defaults of n different
names.
The structure for this model was suggested by Vasicek (1987) and it was first
implemented by Li (2000) and Gregory and Laurent (2005).
Idea
One Factor Gaussian Copula reduces the dimensionality of the problem,
increasing analytical tractability. For this reason, it has become a standard when
pricing CDOs and CDS Index tranches.
The One Factor Gaussian Copula approach goes through the following steps:
1 We consider n names and start from their default times, to which the copula will be
applied:
1 = 1 1
1 ( ln(1 U1 )), . . . , n = n ( ln(1 Un ))
2 We rewrite the definition of copula given by eq. (1) under the risk neutral measure
Q:
C (u1 , . . . , un ) = NRn (N 1 (u1 ), . . . , N 1 (un ))
(7)
= Q(X1 < N 1 (u1 ), . . . , Xn < N 1 (un ))
This entails the calculation of a n-dimensional integral!
3 Inspired by the works of Merton (1974) and Vasicek (1987), we set Ui = N(Xi ),
where the standard Gaussian variables Xi are expressed in terms of:
a systematic common factor Y N (0, 1)
a idiosyncratic term, specific to each name, i N (0, 1) and i.i.d.
p
Xi = i Y + 1 i i (8)
such that corr(Xi , Xj ) = i j
!
N 1 (ui ) i y (10)
=N
1 i
Substituting the single name probabilities (10) into eq. (9), the Gaussian Copula
C (u1 , . . . , un ), in the One Factor framework, can be expressed as one-dimensional
integral:
Z "Y n !#
N 1 (ui ) i y
C (u1 , . . . , un ) = N (y )dy (11)
i =1
1 i
where (y ) is the standard Gaussian probability density.
Dimensionality Reduction
Remark The market further reduces the dimensionality of the problem by introducing a
single value of correlation i = for quotation reasons.
Goal
The Large Homogeneous Portfolio (LHP) approach allows to derive closed form
expressions for the (expected) portfolio loss, Loss, and the (expected) tranched loss,
LosstrA,B .
Assumptions:
1 Gaussian Copula
2 homogeneity of the characteristics of names underlying the credit portfolio:
LHP: Homogeneity I
Homogeneity in recovery rates, default probabilities and correlations implies that:
LHP: Homogeneity II
In general, this integral has to be computed numerically. This is the reason why this
kind of approach is called semi-analytical.
K = k (1 Rec)
McGinty and Ahluwalia (2004) have exploited the third assumption of the LHP approach,
i.e. the infinite size of the portfolio.
We introduce the default rate (DR) of the pool, at a given time T , as the fraction of
defaulted names w.r.t. the total pool of names. Conditional on the systematic factor Y it
is given by:
n
1X
DRnT (Y ) = 1{i T |Y }
n i =1
Conditional on Y , defaults are i.i.d. variables with mean given by the conditional proba-
bility of default:
!
N 1 (p) y
p(Y ; ) := E[1{i T |Y =y} ] = Q{i T |Y =y} = N .
1
DRnT (Y ) p(Y ; )
n
Loss
T (Y ; ) = (1 Rec) p(Y ; ) (16)
and N2 (., ., r ) is the standard normal cumulative distribution function with correlation r .
Outline
1 Introduction
Securitization
Of Models and Mathematicians
2 Dependence
Credit Correlation
Dependence in Reduced Form Models
Copula Function
3 CDOs: Stylized Facts
Stylized Facts
4 CDO Pricing
Monte Carlo Pricing
One Factor Gaussian Copula
Large Homogeneous Portfolio (LHP)
5 Market Quotation Standard
Compound Correlation
Base Correlation
6 Selected References
Paola Mosconi Lecture 6 55 / 69
Market Quotation Standard
Market Quotes
Markets quote CDO tranches only for standardized pools of CDS on different names.
The most liquid indices are the DJi-TRAXX, involving 125 European names and the
DJCDX, involving 125 US names.
A,B
Premium rates R0,T (0) are quoted for the maturities
T = 3y , 5y , 7y , 10y
[0%, 3%] , [3%, 6%] , [6%, 9%] , [9%, 12%] , [12%, 22%] for DJi-TRAXX
and
[0%, 3%] , [3%, 7%] , [7%, 10%] , [10%, 15%] , [15%, 30%] for DJCDX
Implied Correlation
From premium rate quotes, it is market practice to derive the default correlation, which
goes by the name of implied correlation.
Implied Correlation
Two types of correlation are adopted by the market:
1 compound correlation
2 base correlation
The bootstrapping procedure for a given maturity T and tranche [A, B] goes through the
following steps:
1 The equation used to bootstrap is given by (6), that we recall here:
Pb mkt
i =1 P (0, Ti )[E[LosstrA,B (Ti )] E[LosstrA,B (Ti 1 )]]
R AB,mkt = Pb mkt (0, T ) [1 E(Losstr (T ))]
i =1 P i i A,B i
2 under the Gaussian Copula assumption (e.g. in the LHP approximation) the
expected tranche loss is given by eq.s (19) and (18), i.e.:
1
E(LosstrA,B ) = Losstr, B Losstr, tr,
A,B = 0,B () A Loss0,A ()
B A
LGD
E(Losstr0,X ) = Losstr,
0,X () = N(D) + N2 D, N 1 (p), := f0X
GC
()
X
Compound Correlation I
Therefore, we solve recursively eq. (6), where the expected losses are given by:
Compound Correlation II
Figure: Example of compound correlation for the DJ-iTraxx. Source: Brigo and Mercurio
(2006).
Senior Tranches: high attachment points are reached when many defaults occur, i.e.
when default correlation is large (similar to historical correlation)
Mezzanine Tranches: spreads are low given the high demand for these tranches (this
implies low correlation)
Equity Tranche: large spreads are obtained from low correlation. This tranche is
impacted by every default and a large correlation would mean a low probability of a
single default (lower than historical correlation)
Base Correlation I
Base correlation is based on the assumption that each equity tranche [0, X ] is
characterized by a unique value of correlation 0X , implying that a tranche [A, B]
depends on two values of base correlation.
Therefore, we solve recursively eq. (6), where the expected losses are given by:
Base Correlation II
Figure: Example of base correlation for the DJ-iTraxx. Source: Brigo and Mercurio
(2006).
If the Gaussian Copula assumptions were consistent with market tranche prices, there
would be a unique Gaussian Copula model (i.e. unique correlation) consistent with the
market and no distinction between compound correlation and base correlation.
Outline
1 Introduction
Securitization
Of Models and Mathematicians
2 Dependence
Credit Correlation
Dependence in Reduced Form Models
Copula Function
3 CDOs: Stylized Facts
Stylized Facts
4 CDO Pricing
Monte Carlo Pricing
One Factor Gaussian Copula
Large Homogeneous Portfolio (LHP)
5 Market Quotation Standard
Compound Correlation
Base Correlation
6 Selected References
Paola Mosconi Lecture 6 67 / 69
Selected References
Selected References I
Bank of England and European Central Bank (2014), The Case of a Better
Functioning Securitisation Market in the European Union, Discussion Paper
Brigo, D., Pallavicini, A., and Torresetti, R. (2010). Credit Models and the Crisis, or:
How I learned to stop worrying and love the CDOs. Credit Models and the Crisis: A
journey into CDOs, Copulas, Correlations and Dynamic Models, Wiley, Chichester.
Frey, R., McNeil, A.J., and Nyfeler, M.A. (2001): Modeling Dependent Defaults:
Asset Correlations Are Not Enough!.
http://www.risklab.ch/ftp/papers/FreyMcNeilNyfeler.pdf
Laurent, J.P., and Gregory, J., (2005). Basket Default Swaps, CDOs and Factor
Copulas. Journal of Risk, Vol. 7, No. 4, 103-122
Hull, J., and White, A. (2000). Valuing Credit Default Swaps II: Modeling Default
Correlations, Journal of Derivatives, Vol. 8, No. 3
The Lehman Brothers Guide to Exotic Credit Derivatives (2003)
Selected References II