Professional Documents
Culture Documents
Mark Davis
Department of Mathematics
Imperial College London
London SW7 2AZ
www.ma.ic.ac.uk/∼mdavis
1
Plan for the day
2
.
3
Agenda
• Credit Ratings
4
1 Credit Rating
Rating agencies (Moody’s-KMV, S&P, Fitch) assign credit ratings (AAA, AA,. . .)
to firms and transactions on the basis of detailed case-by-case analysis. They
also compile statistics of changes of rating and defaults.
Charts show
These are obtained by a ‘cohort analysis’: start with (say) all the AA-tated
firms on 1 January 1981 ..
5
Moody's Cumulative Default Probabilities
25%
20%
15% Aaa
A1
Baa1
Ba1
10% B1
5%
0%
1 2 3 4 5 6 7 8 9 10 11
Years
6
S&P 1-year rating transition matrix
7
Is the rating transition process Markovian? Let Mk denote the empirical
k-year transition matrix. If the process is Markov, we expect to find
Mk = (M1 )k .
In fact this is not at all accurate. There is a ‘momentum effect’: firms that have
been recently downgraded are more likely to be downgraded again than other
firms in the same rating category. David Lando suggests a Markov model with
additional ‘hyperstates’ A*, BBB* etc. Downgrade probabilities are higher in
A* than in A. A company that is downgraded moves first to A* and then, after
some time, to A.
8
1.1 Credit Default Swaps
Reference
bond/issuer
9
If we have CDS rates πk for maturities Tk , k = 1, . . . , m and a family of distri-
butions {Fθ , θ ∈ Rm } then we can determine the ‘implied default distribution’
Fθ̂ . Example: m = 1 and Fθ (t) = e−θt .
Moral: CDS rates determine the risk-neutral marginal default time distri-
bution for the reference issuer.
Note: Selling credit protection is (nearly) equivalent to buying the reference
bond with borrowed funds:
10
1.2 Collateralized Debt Obligations (CDO)
Senior, L+x,
Bond (80%, AAA)
Portfolio SPV
Mezzanine, L+y,
(8%, BBB)
Equity, 12%
residual receipts
Investors subscribe $100 to SPV which purchases bond portfolio. SPV issues
rated notes to investors. Coupons paid in seniority order.
11
Synthetic CDO
12-100%
Counter- x
party SPV
3-12%
y
0-3%
z
Single-name Tranche
CDSs CDSs
12
The design process for cash-flow CDOs
Cash Flow CBO
Senior, L+x,
Bond (80%, AAA)
Portfolio SPV
Mezzanine, L+y,
(8%, BBB)
Equity, 12%
residual receipts
• Set the size of the senior tranche as big as possible while satisfying expected
loss constraints needed to secure AAA credit rating (see below).
13
1.3 The rating process: Moody’s Binomial Expansion Technique
Start with a portfolio of M bonds, each (for simplicity) having the same notional
value X. Each issuer is classified into one of 32 industry classes. The portfolio
is deemed equivalent to a portfolio of M 0 ≤ M independent bonds, each having
notional value XM/M 0 . M 0 is the diversity score, computed from the following
table.
14
Diversity score table:
15
Diversity score example:
M = 60 bonds.
Meaning: 12 cases where the firm is the only representative of its industry
sector, 12 pairs of firms in the same sector, etc.
Diversity score = 45.
The effect of reduced diversity is to push weight out into the tail of the loss
distribution, as the chart shows.
16
Binomial loss distributions
0.18
0.16
d = 60
0.14 d = 45
d = 30
0.12
0.1
0.08
0.06
0.04
0.02
0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Fraction of portfolio
17
“Loss” in rated tranches
Suppose the coupon in a rated tranche is c and the amounts actually received
are a1 , . . . , an . Then the loss is 1 − q where
a1 a2 an
q= + + ∙∙∙ + .
1+c 1+c 1+c
Note that q = 1 (loss zero) when ai = c, i < n and an = 1 + c.
Moody’s rates tranches on a threshold of expected loss.
18
Expected Loss in Senior Tranche
0.05
0.045
0.04
0.035
0.03
0.025
0.02
0.015
0.01
0.005
0
30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Diversity Score
19
1.4 CreditMetrics
• There are N industry sectors and each obligor i has a weight N -vector wi
such that wi,j represents the participation of obligor i in sector j.
20
• The return for obligor i is
N
X
ri = wi,0 ri,0 + wi,j Rj
j=1
where Rj is the normalized return for sector index Ij and ri,0 is an idiosyn-
cratic factor.
21
AA
A Obligor 2
BBB
B BB BBB A AA
Obligor 1
This procedure gives the joint transition probabilities for all obligors. (In
the figure, Obligor 1 starts at BBB, obligor 2 at A.)
22
2 Joint distributions, Hazard Rates and Copulas
23
For random variables τ1 , τ2 ≥ 0 with joint density f (t1 , t2 ) the marginal and
joint distributions are
Z tZ ∞
F1 (t) = f (u, v)dv du
0 0
Z ∞Z t
F2 (t) = f (u, v)dv du
0 0
Z t1 Z t2
F (t1 , t2 ) = f (u, v)dv du,
0 0
while the survivor function is
Z ∞Z ∞
G(t1 , t2 ) = f (u, v)dv du
t1 t2
24
Let τmin = min(τ1 , τ2 ), τmax = max(τ1 , τ2 ). Then
t+dt
t
t t+dt
25
Suppose τ1 occurs first. The conditional density of τ2 is then
f (τ1 , t)
R∞
τ1 f (τ1 , v)dv
for t ≥ τ1 , so that the new hazard rate is
f (τ1 , t)
h2 (t) = R ∞ ,
t f (τ1 , v)dv
while if τ2 occurs first the hazard rate switches to
f (t, τ2 )
h1 (t) = R ∞ .
t f (u, τ2 )du
The hazard rate process is therefore
h(t) = h0 (t)1(t<τmin ) + h2 (t)1(τmin =τ1 ) + h1 (t)1(τmin =τ2 ) 1(τmin ≤t<τmax )
26
2.1 Copula-based calibration
(i) Back out marginal default distributions F1 (t), F2 (t), . . . from credit spreads
or CDS rates.
27
In more detail:
(i) In a CDS contract, protection buyer pays regular premiums π until min(τ, T )
where T is the contract expiry time and τ the default time of the Reference
Bond. Protection seller pays (1 − R)1(τ <T ) at next coupon date after τ , where
R is the recovery rate. If F is the risk-neutral survivor function of τ , the ‘fair
premium’ π is determined by
n
X n
X
πp(0, ti )F (ti ) = π × CV01 = (F (ti−1 ) − F (ti ))(1 − R)p(0, ti ).
i=1 i=1
28
A standard procedure is to take
Z t
Fθ (t) = exp − h(s)ds
0
where m
X
h(s) = θi 1]Ti−1 ,Ti ] (s).
i=1
Then θi , θ2 , . . . are determined recursively using π1 , π2 , . . ..
Note: in a multivariate setting there is a different parametrization Fj,θj for
each issuer.
29
NB: Care must be taken that the covariance matrix
1 ρ11 ρ12 . . .
Σ= ρ
21 1 ρ 23 . . .
.. .. .. ..
. . . .
is actually non-negative definite. If the ρij are obtained from sample estimates,
it may not be!
To generate X, take X = AZ where A is the Cholesky factorization of Σ
(i.e. Σ = AAT ) and Zi are independent N (0, 1). In the 2-dimensional case we
p
can simply take X1 = Z1 , X2 = ρ12 Z1 + 1 − ρ212 Z2 .
(iv) Default times τi are given by
−1
τi = Fi,θ i
(Ui ).
30
2.2 Simultaneous calibration: The Diamond Model
An obvious drawback of copula methods is: how do you choose the copula? An
alternative is to think in terms of ‘infection’.
A def, B non-def
1
h1 ah2
A non-def A def
0 3
B non-def B def
h2 ah1
2
A non-def, B def
31
• Hazard rate of remaining issuer increases by a factor a after first default.
32
Marginal distributions
h2 e−ah1 t
−(h1 +h2 )t −(h1 +h2 −ah1 )t
F1 (t) = 1 − e − 1−e
h1 + h2 − ah1
h1 e−ah2 t
−(h1 +h2 )t −(h1 +h2 −ah2 )t
F2 (t) = 1 − e − 1−e
h1 + h2 − ah2
Double Default
h2 e−ah1 t
−(h1 +h2 )t −(h1 +h2 −ah1 )t
FDD (t) = 1 − e − 1−e
h1 + h2 − ah1
h1 e−ah2 t
−(h1 +h2 −ah2 )t
− 1−e
h1 + h2 − ah2
Calibration
Joint calibration to credit spreads/CDS rates for issuers A and B, for given
‘enhancement’ parameter a. (Time-varying h1 , h2 required for term structure
of credit spreads)
33
Continuous-time CDS premium πi on asset i is determined by
Z T Z T
−rt
πi e (1 − Fi (t))dt = (1 − R) e−rt fi (t)dt,
0 0
where r is the riskless rate and fi (t) = dFi (t)/dt. From the model, we find
I2
π1 = (1 − R)
I1
where (with m(α, T ) = α1 (1 − e−αT ))
The first default time τmin = τ1 ∧ τ2 is exponential with rate (h1 + h2 ). Hence
the FTD premium is
πFTD = (1 − R)(h1 + h2 ).
34
Calibrated parameters h1, h2, and First-to-Default premium
0.035
0.03
0.025
h1
h2
0.02 FTD
0.015
0.01
0.005
0
1 2 3 4 5 6 7 8
enhancement factor a
35
Generators and Backward Equations
(as in Black-Scholes).
36
For the diamond model, xt ∈ {0, 1, 2, 3} and the generator can be expressed in
matrix form as
−(hA + hB ) hA hB 0
0 −βhB 0 βhB
A= .
0 0 −αhA αhA
0 0 0 0
37