ENPC CreditRisk CheatSheet
ENPC CreditRisk CheatSheet
ENPC CreditRisk CheatSheet
Abstract
This short document lists the main formulas, concepts and definitions of the class. Framed definitions starting with É are the key
concepts of the class that must be known. Æ are important information to keep in mind as general knowledge. Î refers to traps and
points of attention.
Î What is a risk-free rate?. the risk-free rate is usually considered as τ < t + dt | τ > t
the Constant Maturity Swap (CMS) price, for different maturities. For ex-
ample, in Europe, the 10 years, risk-free rate, would be the CMS 10y that The most applied reduced-form models assumes that the probability of de-
exchanges a fixed rate with EURIBOR 3M (ticker BBG being EUSA10Y). fault is the product of the infinitesimal time step with a so-called default
intensity (considered constant here):
Q(τ < t + d t | τ > t) = λ × d t
É Price of a bond – continuous rate and coupon. Considering a contin-
uous coupon, the formula for a bond of nominal 1, is: which is equivalent to say that: the random time of default follows an ex-
−r A T
ponential law of rate λ.
1−e Thus, the survival probability of the studied counterparty is:
B̄ A(0, T ) = 1 + (c − r A)
rA
Q(τ > t) = exp (−λt)
(only if we consider ri and sA as constant)
which is consistent with the no-arbitrage formula introduced earlier.
Î A chicken and egg problem. Note that this is a chicken and egg
problem: the spread is extracted from the price and the price is deduced
from the spread. In practice, the market, by buying and selling bonds, Æ Hazard rate modeling.
agrees to a price from which one can extract a spread to price new bonds
• Constant: Time homogeneous Poisson Process;
or credit derivatives.
• Deterministic: Time deterministic inhomogeneous Poisson Pro-
cess;
É The implied probability of default. The no-arbitrage assumption gives • Stochastic: Time-varying and stochastic Poisson Process as the
us: Cox, Ingersoll, Ross (CIR) model.
B̄ A(t, T )
Q(τ > T | τ > t) =
B(t, T )
where τ is the rv of the time of default. And thus, given the value of risky
bonds and risk-free bonds (B(t, T )):
É Implied probability of default taking into account recovery. Let R
A (T −t)
1 − P D = Q(τ > T | τ > t) = e−s be the recovery rate, the implied probability of default taking into account
recovery is:
B̄(0,T )
1 − B(0,T )
PD =
1−R
É The Expected Loss (EL). The Expected Loss (EL) on a credit exposure Moody’s S&P Fitch Rating description
can be split in three parts: Aaa AAA AAA Prime
Aa1 AA+ AA+
• PD: the Probability of Default (see above);
Aa2 AA AA High grade
• LGD: the Loss Given Default is equal to 1−R, where R is the recovery Aa3 AA- AA-
rate, that is the proportion of the exposure that the lender retrieves; A1 A+ A+
• EAD: the Exposure At Default, that can be fixed (for bullet bonds for A2 A A Upper medium grade
examplea ) or not (exposure of derivatives towards a counterparty A3 A- A-
for example). Baa1 BBB+ BBB+
Baa2 BBB BBB Lower medium grade
Resulting in: Baa3 BBB- BBB-
EL = E(EAD × 1{τ<M } × LGD) Ba1 BB+ BB+
Ba2 BB BB Non-investment grade / Speculative
= E(EAD) × E(1{τ<M } ) ×E(LGD)
|{z} | {z } Ba3 BB- BB-
assump.
PD B1 B+ B+
a
B2 B B Highly speculative
Bonds which notional is reimbursed at the maturity B3 B- B-
Caa1 CCC CCC
Caa2 CCC CCC Substantial risks
PRICING CDS Caa3 CCC CCC-
Lecture 1 Ca CC CC Extremely speculative
C C C Default imminent
C RD DDD In default
The value of the fixed leg of a CDS is:
É Transition matrices. In credit risk, a transition matrix, M t,t+1 = (mi j )i j ,
1 − e−(r+λ)T is a matrix where:
Fixed(0, T ) = s(0, T )
r +λ
mi j = P(Grade t+1 = j | Grade t = i)
É The sensitivity of a CDS. The sensitivity of the MtM is the risky dura-
tion, DV: Î Markovian property. The existence of such a generator matrix is
1 − e−(r+λ)(T −t) based on the assumption of the Markov propriety that states (in the dis-
DV(t, T, λ) = crete case) that:
r +λ
P(Grade t+1 = j | Grade t = i) = P(Grade t+1 = j | Grade t = i,
Grade t−1 = i t−1 , ..., Grade t−h = i t−h )
É Valuation of a CDS. Let s0 , be the spread in t = 0, and s t , the spread
today, in t . which is not observed in practice.
The Present Value of the protection seller is:
P V (s0 , s t ) = DV(0, t, λ)(s0 − s t ) STATISTICAL TOOLS TO ASSESS CREDIT RISK
Lecture 2
Æ Classification trees. A classification tree is a model with a tree-like É Leland model. The Leland model is a structural model that assumes
structure. It contains nodes and edges/branches. It is grown through a that default occurs when the value of the assets of the firm goes under a
recursive binary splitting procedure: certain threshold K . The assets follows the following process:
• Step 1: We select the predictor X k and the cutting point s in or- dA t
der to split the predictor space into the two regions {X , X k < s} and = (r − δ)d t + σdWt
At
{X , X k ≥ s} that gives the greatest decrease in the criterion we want
to minimize (Gini index for example). where δ is the dividend rate.
• Step 2: We repeat step 1 but instead of splitting the whole predic-
Using Laplace transform, one can deduce the value of the debt, the
tor space we split one of the two regions identified at step 1. The
equity and the firm, today. Moreover, one can find the optimal amount
predictor space is now divided into three regions.
of capital in the balance sheet and the optimal coupon to pay in order to
• Step 3: The regions are then recursively split until no split can de- optimize the value of the equity of the firm.
crease the criterion.
Æ Pros and Cons of decision trees. Î Structural models limits. Structural models carry interesting eco-
nomic interpretations, but for single-name models they are two simplistic
• Pros: Can be easily interpreted and visualised, allows for non-
to be used for pricing or risk measurement. Statistical models are mostly
linear predictions.
used in the historical probability world and reduced-form models in the
• Cons: Exhibit high variance, can be biased when the sample is risk neutral world.
unbalanced.
Î Bagging and correlation. Bagging consists in training several trees on
bootstrapped samples from the original training set. The trees are then PORTFOLIO MODELS
Lecture 4
aggregated by choosing the most common prediction among the trees’
predictions or any suitable voting rule.
• The resulting classifier exhibits less variance than a single trees É Vasicek model. The Vasicek model gives the loss distribution of a port-
• The gain in variance is limited since the trees are correlated folio of defaultable assets.
Let us suppose we have a countable infinite number of bonds (loans, mort-
É Random Forests. A Random Forest algorithm is a bagging procedure gages, etc.) of equal nominal, same maturity, same probability of default at
with trees trained as follows: maturity (PD), and a same recovery rate (R). We assume that bond i de-
faults when the latent variable R i < s, where s is a common latent threshold
• A random sample is drawned from the training set for all bonds. Moreover, we assume that:
• At each knot the split is made using only a random subset of the p p
∀i ∈ N, Ri = ρ F + 1−ρ ei
predictors |{z} |{z} |{z}
corelation systemic idiosyncratic
Using only a subset of the predictors yields less correlated trees than a stan- factor factor factor
dard bagging algorithm.
with (ei )i∈N and F are standard normal variables, and thus (R i )i∈N are stan-
dard normal and correlated.
Æ Pros and Cons of Random Forest. We know that PD = Q(R i < s) = |{z}Φ (s) and thus:
• Pros: Exhibit less variance than basic trees or bagging and are Normal
less prone to overfitting. cdf
STRUCTURAL MODELS Note that L is conditioned by the value of F , the stochastic systemic factor.
Lecture 3
É Merton model. The Merton model is a structural model that assumes É Copulas. A copula C , is a function ([0; 1]d → [0; 1]) that allows to
that the default occurs when the value of the firm (Vt ) is inferior to D the model dependencies:
amount of debt of the firm, at maturity T of the debt. In this model, the
∀(x 1 , ..., x d ) ∈ Rd , F (x 1 , ..., x d ) = C (F1 (x 1 ), ..., Fd (x d ))
value of the firm follows the following process:
Sklar’s theorem asserts that from any continuous multivariate distribution
d Vt
= r d t + σdWt G , a copula can be deduced with the following formula:
Vt
∀(u1 , ..., ud ) ∈ [0; 1]d , C(u1 , ..., ud ) = F (F1−1 (u1 ), ..., Fd−1 (ud ))
Using option value theory after noticing being a shareholder is like having
a European call option of strike D, maturity T , on Vt , one can find the value where F1 , ..., Fd are the marginal distributions of the d−dimensional distri-
of the equity, of the debt and thus, the spread of the firm and its probability bution F .
of default.
Æ The Gaussian copula.. In the Gaussian case, we have: Î Implicit correlation – Base correlation. Let us consider the tranche
of an ABS, with attachment point A, detachment point D and spread
∀(u1 , ..., ud ) ∈ [0; 1] , d
CRN (u1 , ..., ud ) = ΦR (Φ −1
(u1 ), ..., Φ −1
(ud )) s[A;D] .
• Implicit correlation: the implicit correlation of the considered
tranche, is the correlation ρ [A;D] to put into the pricing model to
Î Why and how to use copula? find s[A;D] ;
• Why? Thanks to copula, one can design complex dependencies • Base correlation: the base correlation of A (and of D) is the im-
between the defaults (tail dependencies, asymmetric dependen- plicit correlation ρ [0;A] (and ρ [0;D] ) to put into the pricing model
cies, etc.). to find the spreads of the – theoretical – equity tranches s[0;A] (and
s[0;D] ).
• How? Most copula usage requires numerical simulations: (i) the
copula is simulated to get uniform observations whose structure One can show that:
of dependence is the one of the copula, (ii) by composing each J V [0;D] (ρ [0;D] ) − J V [0;A] (ρ [0;A] )
uniform simulation by the considered univariate distribution, one s[A;D] =
DV [0;D] (ρ [0;D] ) − DV [0;A] (ρ [0;A] )
gets the required marginal distribution with the structure of de-
pendence of the copula. ensuring a bijective relationship between base correlations and spreads
The copulas most used are Gaussian, Student, Archimedean, Clayton, of CDO tranches.
Franck, Gumbel copulas.
Î Different kinds of credit enhancement. There are several way to
improve the credit profile of an ABS / a tranche of an ABS (CDO):
• Excess spread: the received rate is higher than the served one;
FTD, CLN, ABS, CDO, CSO • Overcollateralization: the face value of the underlying loan port-
Lecture 4 folio is larger than the security it backs;
• Monolines and wrapped securities: CDS on the underlying as-
sets are bought to monolines in order to cover the losses.
É Spread of the tranche of an ABS. To assess the spread of a CDO, that É Collateralized Debt Obligation Square (CDO2). A CDO square is an
is a tranche of an ABS, let us consider the case of a given tranche whose at- ABS whose assets are CDO, mostly mezzanine tranches of several RMBS.
tachment point is A and detachment point D. We denote L t the distribution
[A;D]
loss of the underlying portfolio in t , and L t the distribution loss on the
considered tranche.
The cash-flows for the buyer of the tranche are fixed and equal to:
REGULATORY APPROACH TO CREDIT RISK
Lecture 5
Z T
[A;D]
J F [A;D] (0, T ) = s[A;D] e−r t (D − A − E L t )d t
0
É Credit Risk Weighted Assets (RWA). Risk Weighted Assets
| {z } are computed by weighting assets following a regulatory formula
DV [A;D] (0,T ) fCredit RWA (PD, LGD, EAD).
[A;D] [A;D] Adding Credit RWA, to Market RWA and Operational RWA, we get the RWA
where E L t = E(L t )
of the banks, and can thus compute the capital ratio:
The cash-flows for the seller of the tranche are variable and equal to:
capital
capital ratio =
Z T
[A;D] −r T [A;D]
J V [A;D] (0, T ) = E L T e +r e−r t E L t dt RWA
0
And thus, the spread of the tranche is: Î Different approaches. There are three approaches to compute credit
J F [A;D] (0, T ) RWA:
s[A;D] = • The Standard Approach: the RWA is the product of a given Risk
DV [A;D] (0, T )
Weigh multiplied by EAD;
NB: for the loss distribution, one can use any portfolio model, Vasicek’s
being the one used most often. • IRBF (Internal Rating Based Foundation): the bank can esti-
mate PD; LGD and EAD depends on regulatory formulas that de-
pend on the considered asset;
• IRBA (Internal Rating Based Approach): PD, LGD and EAD are
computed by internal models of the bank.
where, T1 / r1 , T2 / r2 and D/ r D is the proportion / the cost (spread) of Tier Æ Use of counterparty risks measurement.
1 capital, Tier 2 Capital and debt in the liabilities of the bank and r the • Pricing: to take into this risk when pricing a derivative (Counter-
risk-free rate. party Value Adjustment – CVA);
• Risk management: for internal and regulatory purposes (SIMM,
Î WACC and hurdle rate.. The WACC is the hurdle rate of bank activi- KCVA, PFE, etc. – see below).
ties, that is, the minimum return necessary to create value.
É Exposures – EE, PFE, MPFE, EPE, EEPE. V (t) denotes the market
É Economic Value Added (EVA). The Economic Value Added for a bank value of a derivative, at time t . Counterparty exposure is equal to
is: E(t) = V (t)+ = max(0, V (t)).
É Risk Adjusted Returns On Risk Adjusted Capital (RARORAC). The Effective Positive Exposure (EPE).
Risk Adjusted Return On Risk Adjusted Capital for a bank is: Z t
1
RARORAC = RAROC − k EPE(t) = EE(s)ds
t 0
Another definition of the RARORAC is:
Expected Effective Positive Exposure (EEPE).
Allocated Economic Capital
RARORAC = RAROC − k × Z t
Used Economic Capital 1
EEPE(t) = max(EE(h))ds
t 0 h<s
É Cost of capital – A CAPM approach. The Capital Asset Pricing Model PFE is then considered as the EAD in the regulatory formula used to com-
states that the average return of the stock i , E(ri ), follows: pute RWA.
E(ri ) = βi (E(r M ) − r f ) + r f
É CVA – Credit Value Adjustment.
σ
where βi = ρi,M σ i
M
, r M is the return of the whole stock market, σ M is the Z T
volatility of the whole stock market, σi is the volatility of the stock i , ρi,M
CVA = (1 − R) EQ (E(s))dSC (s)
is the correlation between the returns of the market and the ones of i , r f is 0
the risk-free rate. Where SC is the survival function of the counterparty.
É Cost of capital – The Gordon-Shapiro Approach. The Gordon-Shapiro Æ Two ways to mitigate counterparty risk:
approach states that the valuation of a firm, P , is a function of the expected
• Netting: in presence of multiple trades with a counterparty, net-
growth g of its cash-flows (Dt ) and the expected return k of the sharehold-
ting agreements allow, in the event of default of one of the coun-
ers:
∞ terparties, to aggregate the transactions before settling claims
Dt D1
( E(t) = (ΣV (i))+ 6= Σ(V (i))+ ).
X
P= =
t=1
(1 + k) t k −g
• Collateralizing: collateral is a property or other assets that a
with Dt = Dt−1 × (1 + g) counterparty offers as a way for the counterpary to secure the
exposure. É IRC – Incremental Risk Charge The IRC is a risk metrics that captures
risk due to adverse rating migrations on vanilla credit securities such as
Î Other counterparty risk metrics. There are several other counter-
bonds and CDS on corporates and sovereigns within the trading book.
party risk metrics such as:
• SIMM: SIMM stands for Standard Initial Margin Model and is
used to compute the initial margin of non-cleared derivatives;
• KCVA is a VaR estimate of the CVA required by Basel III that im-
pose a capital charge. É CRM – Comprehensive Risk Measure
Æ IRC and CRM, credit market risk metrics: IRC and CRM are credit • The CRM is a risk metrics that, as the IRC, captures the risks due to
market risk metrics that are not counterparty risk metrics as they mea- adverse rating migrations on credit securities. It applies to credit cor-
sure the credit risk due to the potential default(s) of the underlying refer- relation portfolios (CDO, CLO, CBO, etc.) within the trading book.
ence(s) on derivative products. With the VaR and the stressed VaR, these • The CRM also captures risks due to credit spread, recovery rates and
measure are part of the market RWA. base correlations variations.