10 Credit risk

10.1 Credit risky instruments

10.2 Measuring credit quality

10.3 Structural models of default

10.4 Bond and CDS pricing in hazard rate models

10.5 Pricing with stochastic hazard rates

10.6 Affine models

© QRM Tutorial Section 10

What is credit risk?
“Credit risk is the risk of a loss arising from the failure of a coun-
terparty to honour its contractual obligations. This subsumes both
default risk (the risk of losses due to the default of a borrower or a
trading partner) and downgrade risk (the risk of losses caused by a
deterioration in the credit quality of a counterparty that translates
into a downgrading in some rating system). ”
Obligor = a counterparty who has a financial obligation to us; for
example, a debtor who owes us money, a bond issuer who promises
interest, or a counterparty in a derivatives transaction.
Default = failure to fulfill that obligation, for example, failure to repay
loan or pay interest/coupon on a loan/bond; generally due to lack of
liquidity or insolvency; may entail bankruptcy.

© QRM Tutorial Section 10

A crucial risk category
Credit risk is omnipresent in the portfolio of a typical financial institution.
A portfolio of loans or (corporate) bonds is obviously affected by credit
risk.
Credit risk accompanies any OTC (over-the-counter) derivative transac-
tion such as a swap, because the default of one of the parties involved
may substantially affect the actual pay-off of the transaction.
There is a specialized market for credit derivatives, such as credit default
swaps.
Credit risk relates to the core activities of most banks but is also highly
relevant to insurance companies: Insurers are exposed to substantial
credit risk in their investment portfolios and counterparty default risk in
their reinsurance treaties.

© QRM Tutorial Section 10

Credit risk management: A range of tasks
An enterprise needs to determine the capital it requires to absorb losses
due to credit risk.
Portfolios of credit-risky instruments should be well diversified and
optimized according to risk-return considerations.
Institutions need to manage their portfolio of traded credit derivatives,
which involves pricing, hedging and managing collateral for such trades.
Financial institutions need to control the counterparty credit risk in their
trades and contracts with other institutions. This has particularly been
the case since the 2007–2009 financial crisis.

© QRM Tutorial Section 10

10.1 Credit risky instruments
10.1.1 Loans
May be categorized into: retail loans (to individuals and small or medium-
sized companies), corporate loans (to larger companies), interbank loans
and sovereign loans (to governments).
In each of these categories there may be a number of different products.
For example, retail customers may borrow money using mortgages
against property, credit cards and overdrafts.
A sum of money, known as the principal, is advanced to the borrower for
a particular term in exchange for a series of defined interest payments,
which may be at fixed or floating interest rates. At the end of the term
the borrower is required to pay back the principal.

© QRM Tutorial Section 10.1

 A useful distinction to make is between secured and unsecured lending.
If a loan is secured the borrower has pledged an asset as collateral for
the loan. In a mortgage the collateral is a property. In the event of
default, the lender may take possession of the asset to mitigate the loss.
In an unsecured loan the lender has no such claim on a collateral asset.

10.1.2 Bonds
Bonds are publicly traded securities issued by companies and governments
which allow the issuer to raise funding on financial markets.
Bonds issued by companies are corporate bonds and bonds issued by
governments are known as treasuries, sovereign bonds or, particularly in
the UK, gilts (gilt-edged securities).
The security commits the bond issuer (borrower) to make a series of
interest payments to the bond buyer (lender) and pay back the principal
at a fixed maturity.
© QRM Tutorial Section 10.1.2
 The interest payments, or coupons, may be fixed at the issuance of
the bond (so-called fixed-coupon bonds). Alternatively, there are also
bonds where the interest payments vary with market rates (so-called
floating-rate notes).
The reference rate for the floating rates is often a LIBOR rate (London
Interbank Offered Rate).
There are also convertible bonds which allow the purchaser to convert
them into shares of the issuing company at predetermined time points.

Risks faced by bondholders

A bond holder is subject to a number of risks, particularly interest-rate
risk, default risk, downgrade risk and spread risk.
Changes in the term structure of interest rates affect the value of bonds.
As for loans, default risk is the risk that promised coupon and principal
payments are not made.
© QRM Tutorial Section 10.1.2
 Downgrade risk is the risk that the bond loses value because the issuer’s
credit rating is lowered.
Historically government bonds issued by developed countries have been
considered default-free; for obvious reasons, after the European debt
crisis of 2010–2012, this notion was called into question.
Spread risk is a form of market risk that refers to changes in credit
spreads. The credit spread of a defaultable bond measures the difference
in the yield of the bond and the yield of an equivalent default-free bond.
An increase in the spread of a bond means that the market value of the
bond falls, which is generally interpreted as indicating that the financial
markets perceive an increased default risk for the bond.

© QRM Tutorial Section 10.1.2

10.1.3 Derivative contracts subject to counterparty risk
A substantial part of all derivative transactions is carried out over the
counter and there is no central clearing counterparty such as an organized
exchange to guarantee the fulfilment of the contractual obligations.
These trades are subject to the risk that a contracting party defaults
during the transaction, thus affecting the cash flows that are actually
received by the other party. This risk, known as counterparty credit risk,
received a lot of attention during the financial crisis of 2007-2009.
Some of the institutions heavily involved in derivative transactions experi-
enced worsening credit quality or—in the case of Lehman Brothers—even
a default event.
Counterparty risk management is now a key issue for all financial insti-
tutions and the focus of many new regulatory developments.

© QRM Tutorial Section 10.1.3

Example of interest-rate swap
Two parties A and B agree to exchange a series of interest payments
on a given nominal amount of money for a given period.
A receives payments at a fixed interest rate and makes floating payments
at a rate equal to the three-month LIBOR rate.
Suppose that A defaults at time τA before the maturity of the contract.
If interest rates have risen relative to their value at inception of contract:
◮ The fixed interest payments have decreased in value and the value
of the contract has increased for B.
◮ The default of A constitutes a loss for B.
◮ The loss size depends on the term structure of interest rates at τA .

© QRM Tutorial Section 10.1.3

 If interest rates have fallen relative to their value at t = 0:
◮ The fixed payments have increased in value so that the swap has a
negative value for B.
◮ B will still has to pay the value of the contract into the bankruptcy
pool,
◮ There is no upside for B in A’s default.

If B defaults first the situation is reversed: falling rates lead to a

counterparty-risk-related loss for A.

© QRM Tutorial Section 10.1.3

Management of counterparty risk
Counterparty risk has to be taken into account in pricing and valuation.
This has led to the notion of credit value adjustments (CVA).
Counterparty risk needs to be controlled using risk-mitigation techniques
such as netting and collateralization.
Under a netting agreement the value of all derivatives transactions
between A and B is computed and only the aggregated value is subject
to counterparty risk; since offsetting transactions cancel each other out,
this has the potential to reduce counterparty risk substantially.
Under a collateralization agreement the parties exchange collateral (cash
and securities) that serves as a pledge for the receiver. The value of the
collateral is adjusted dynamically to reflect changes in the value of the
underlying transactions.

© QRM Tutorial Section 10.1.3

10.1.4 Credit default swaps and related credit derivatives
Credit derivatives are securities which are primarily used for the hedging
and trading of credit risk.
The promised pay-off of a credit derivative is related to credit events
affecting one or more firms.
Major participants in the market for credit derivatives are banks, insur-
ance companies and investment funds.
Retail banks are typically net buyers of protection against credit events;
other investors such as hedge funds and investment banks often act as
both sellers and buyers of credit protection.
Credit default swaps (CDSs) are the workhorses of the credit derivatives
market and the market for CDSs written on larger corporations is fairly
liquid.

© QRM Tutorial Section 10.1.4

Structure of CDS
Consider contract with maturity T and ignore counterparty credit risk.
Three parties are involved (only two directly):
C (reference entity); default at time τC < T triggers default payment.
A (protection buyer); pays premiums to B until min(τC , T ).
B (protection seller); makes default payment to A if τC < T .

C
✻
❄
✛premium payments at fixed times
yes: default payment ✲
B C defaults ? ✲ A
no: no payment ✲

© QRM Tutorial Section 10.1.4

CDS: Payment flows
If the reference entity experiences a default before the maturity date T
of the contract, the protection seller makes a default payment to the
protection buyer, which mimics the loss due to the default of a bond
issued by the reference entity (the reference asset); this part of a CDS
is called the default payment leg.
As compensation the protection buyer makes periodic premium payments
(typically quarterly or semiannually) to the protection seller (the premium
payment leg); after the default of the reference entity, premium payments
stop. There is no initial payment.
The premium payments are quoted in the form of an annualized per-
centage x∗ of the notional value of the reference asset; x∗ is termed the
(fair or market quoted) CDS spread.

© QRM Tutorial Section 10.1.4

Use of CDS
Investors enter into CDS contracts for various reasons.
Bond investors with a large credit exposure to the reference entity may
buy CDS protection to insure themselves against losses due to default of
a bond (easier than reducing the original bond position as CDS contracts
are more liquid).
CDS contracts are also held for speculative reasons: so-called naked
CDS positions, where the protection buyer does not own the bond are
often assumed by investors who are speculating on the widening of the
credit spread of the reference entity (similar to short-selling bonds issued
by the reference entity.)
Note that, in contrast to insurance, there is no requirement for the
protection buyer to have insurable interest, that is, to actually own a
bond issued by the reference entity.

© QRM Tutorial Section 10.1.4

10.1.5 PD, LGD and EAD
Exposure
If we make a loan or buy a bond, our exposure is relatively easy to
determine, since it is mainly the principal that is at stake. There is some
additional uncertainty about the value of lost interest payments.
A further source of exposure uncertainty is due to the widespread use
of credit lines, essentially a ceiling up to which a corporate client can
borrow money at given terms.
For OTC derivatives the counterparty risk exposure is even more difficult
to quantify, since it is a stochastic variable depending on the unknown
time at which a counterparty defaults and the evolution of the value of
the derivative up to that point.
In practice the concept used is exposure at default (EAD), which recog-
nises that exposure often depends on the exact default time.
© QRM Tutorial Section 10.1.5
Probability of default (PD)
When measuring the risk of losses over a fixed time horizon, for example
one year, we are particularly concerned with estimating the probability
that obligors default by the time horizon, a quantity known to practi-
tioners as probability of default or PD.
The probability of default is related to the credit quality of an obligor
and we discuss models of credit quality next.
For instruments where the loss is dependent on the exact timing of
default, for example OTC derivatives with counterparty risk, the risk of
default is described by the whole distribution of possible default times
and not just the probability of default by a fixed horizon.
In simple models of default time, the probability of default may be
expressed in terms of a hazard function which measures the risk of
default at any instant in time.

© QRM Tutorial Section 10.1.5

Loss given default (LGD)
In the event of default, it is unlikely that the entire exposure is lost.
When a mortgage holder defaults on a residential mortgage, and there
is no realistic possibility of restructuring the debt, the lender can sell
the property (the collateral asset) and the proceeds from the sale will
make good some of the lost principal.
When a bond issuer goes into administration, the bond holders join the
group of creditors who will be partly recompensed for their losses by the
sale of the firm’s assets.
Practitioners use the term loss given default or LGD to describe the
proportion of the exposure that is actually lost in the event of default,
or its converse, the recovery, to describe the amount of the exposure
that can be recovered through debt restructuring and asset sales.

© QRM Tutorial Section 10.1.5

Dependence of EAD, PD and LGD
EAD, PD and LGD are dependent quantities. For example, in a period of
financial distress, when PDs are high, asset values of firms are depressed
and firms are defaulting, recoveries are likely to be correspondingly low,
so that there is positive dependence between PDs and LGDs.

© QRM Tutorial Section 10.1.5

10.2 Measuring credit quality
Scores, ratings & measures inferred from prices
There are two philosophies for quantifying the credit quality or default
probability of an obligor.
1) Credit quality can be described by a credit rating or score that is based
on empirical data and expert judgement.
2) For obligors whose equity is traded on financial markets, prices can be
used to infer the market’s view of the credit quality of the obligor.
Credit ratings and scores fulfill a similar function—they allow us to order
obligors by their credit risk and map that risk to an estimate of the PD.

Credit ratings tend to be expressed on an ordered categorical scale whereas

credit scores are often expressed in points on a metric scale.

© QRM Tutorial Section 10.2

Rating and scoring
The task of rating obligors is often outsourced to a rating agency such
as Moody’s or Standard & Poor’s (S&P).
In the S&P rating system there are seven pre-default rating categories
labelled AAA, AA, A, BBB, BB, B, CCC, with AAA being the highest
and CCC the lowest rating.
Moody’s uses nine pre-default rating categories labelled Aaa, Aa, A,
Baa, Ba, B, Caa, Ca, C.
A finer alpha-numeric system is also used by both agencies.
Credit scores are traditionally used for retail customers and are based
on so-called scorecards. Historical data is used to model default risk as
a function of demographic, behavioural and financial covariates using
techniques like logistic regression. The covariates are weighted and
combined into a score.

© QRM Tutorial Section 10.2

10.2.1 Credit rating migration
In the credit-migration approach each firm is assigned to a credit-rating
category at any given time point.
We assume that the current credit rating completely determines the
default probability.
The probability of moving from one credit rating to another over a given
risk horizon (typically one year) is then specified.
These probabilities, known as transition probabilities, are typically pre-
sented in the form of a matrix. They are estimated from historical data
on empirical transition rates.
The following example is taken from Ou (2013), (Exhibit 26). It gives
average transition rates from one rating to another within one year. WR
stands for withdrawn rating.

© QRM Tutorial Section 10.2.1

 Rating at year-end (%)
Initial z }| {
rating Aaa Aa A Baa Ba B Caa Ca–C Default WR
Aaa 87.20 8.20 0.63 0.00 0.03 0.00 0.00 0.00 0.00 3.93
Aa 0.91 84.57 8.43 0.49 0.06 0.02 0.01 0.00 0.02 5.48
A 0.06 2.48 86.07 5.47 0.57 0.11 0.03 0.00 0.06 5.13
Baa 0.039 0.17 4.11 84.84 4.05 7.55 1.63 0.02 0.17 5.65
Ba 0.01 0.05 0.35 5.52 75.75 7.22 0.58 0.07 1.06 9.39
B 0.01 0.03 0.11 0.32 4.58 73.53 5.81 0.59 3.85 11.16
Caa 0.01 0.02 0.02 0.12 0.38 8.70 61.71 3.72 13.34 12.00
Ca-C 0.00 0.00 0.00 0.00 0.40 2.03 9.38 35.46 37.93 14.80

1-year default probability for an A-rated company is estimated to

be 0.06%, whereas for a Caa-rated company it is 13.3%.
In practice a correction to the figures would probably be undertaken to
account for rating withdrawals

© QRM Tutorial Section 10.2.1

 Rating agencies also publish cumulative default rates over longer time
horizons.
These provide estimates of cumulative default probabilities over several
years. Alternative estimates of multi-year default probabilities can be
inferred from one-year transition matrices as explained later.
The data are taken from Ou (2013), (Exhibit 33).

Term
Initial z }| {
rating 1 2 3 4 5 10 15
Aaa 0.00 0.01 0.01 0.04 0.11 0.50 0.93
Aa 0.02 0.07 0.14 0.26 0.38 0.92 1.75
A 0.06 0.20 0.41 0.63 0.87 2.48 4.26
Baa 0.18 0.50 0.89 1.37 1.88 4.70 8.62
Ba 1.11 3.08 5.42 7.93 10.18 19.70 29.17
B 4.05 9.60 15.22 20.13 24.61 41.94 52.22
Caa-C 16.45 27.87 36.91 44.13 50.37 69.48 79.18

© QRM Tutorial Section 10.2.1

TTC and PIT
Default rates tend to vary with the state of the economy, being high
during recessions and low during periods of economic expansion.
Transition rates as estimated by rating agencies are historical averages
over longer time horizons covering several business cycles.
For instance the transition rates we show have been estimated from
rating migration data over the period 1970–2012.
Rating agencies focus on the average credit quality through the business
cycle (TTC) when attributing a credit rating to a particular firm.
Hence the default probabilities used in the rating migration approach are
estimates of the average default probability, independent of the current
economic environment.
These can be contrasted with point-in-time (PIT) estimates of default
probabilities which reflect the current macroeconomic environment.
© QRM Tutorial Section 10.2.1
10.2.2 Rating transitions as a Markov chain
Let (Rt ) denote a discrete-time stochastic process taking values in
S = {0, 1, . . . , n} at times t = 0, 1, . . ..
The set S defines rating states of increasing creditworthiness with 0
representing default. (Rt ) models an obligor’s rating over time.
We will assume that (Rt ) is a Markov chain. This means that it has
the Markov property that
P(Rt = k | R0 = r0 , R1 = r1 , . . . , Rt−1 = j) = P(Rt = k | Rt−1 = j)
for all t ≥ 1 and all j, r0 , r1 , rt−2 , k ∈ S.
Conditional probabilities of rating transitions given an obligors’s rating
history depend only on the previous rating Rt−1 = j at the last time
point and not the more distant history.
There is evidence that rating histories show momentum and stickiness
which violates the Markov assumption (Lando and Skodeberg (2002)).
© QRM Tutorial Section 10.2.2
Properties of Markov chains
The Markov chain is stationary if, for all t ≥ 1 and rating states j, k,
P(Rt = k | Rt−1 = j) = P(R1 = k | R0 = j).

In this case we can define the transition matrix P = (pjk ) with elements
pjk = P(Rt = k | Rt−1 = j), for any t ≥ 1.
The Chapman-Kolmogorov equations say that
X
P(Rt = k | Rt−2 = j) = pjl plk .
l∈S

An implication of this is that the matrix of transition probabilities over

two time steps is given by P 2 = P × P .
It is not clear how a matrix of transition probabilities for a fraction of a
time period can be computed (one would need continuous-time chains).

© QRM Tutorial Section 10.2.2

Estimating default and transition probabilities
For t = 0, . . . , T − 1 and j ∈ S \ {0} let Ntj denote the number of
companies that are rated j at time t and for which a rating is available
at time t + 1; let Ntjk denote the subset of those companies that are
rated k at time t + 1.
Under the Markovian assumption the Ntj companies rated j can be
thought of as being randomly allocated to the ratings k ∈ S according
P
to probabilities pjk which satisfy nk=0 pjk = 1.
In this framework the likelihood is given by
  
TY
−1 n
Y n pNtjk
Y jk 
L((pjk ); (Ntj ), (Ntjk )) =  Ntj ! .
t=0 j=1 k=0
Ntjk !

© QRM Tutorial Section 10.2.2

 P
If this is maximized subject to the constraints that nk=0 pjk = 1 for
j = 1, . . . , n we obtain the maximum likelihood estimator
PT −1
Ntjk
p̂jk = Pt=0
T −1
. (70)
t=0 Ntj

Continuous-time Markov transition models

The main drawback of modelling rating transitions as a discrete-time
Markov chain is that we ignore any information about intermediate
transitions taking place between two times t and t + 1.
For this reason, better to consider transitions in continuous time. Prob-
abilities cannot be modelled directly but are instead modelled in terms
of transition rates.
Over any small time step of duration δt we assume that the probability
of a transition from rating j to k is given approximately by λjk δt. The
P
probability of staying at rating j is given by 1 − k6=j λjk δt.
© QRM Tutorial Section 10.2.2
 If we now define a matrix Λ to have off-diagonal entries λjk and diagonal
P
entries − k6=j λjk , we can summarise these transition probabilities for
a small time step in the matrix
In+1 + Λδt.

Λ is the so-called generator matrix .

The generator matrix

Let P (t) be the matrix of transition probabilities for the period [0, t].
Divide [0, t]into N small time steps of size δt = t/N for N large.
The matrix of transition probabilities can be approximated by
 N
Λt
P (t) ≈ In+1 +
N
This converges, as N → ∞,to the so-called matrix exponential of Λt.
P (t) = exp (Λt)
© QRM Tutorial Section 10.2.2
 We can compute transition probabilities for any time horizon.
A Markov chain with generator Λ can be constructed in the following
way. An obligor remains in rating state j for an exponentially distributed
P
amount of time with parameter λ = k6=j λjk . When a transition takes
place the probability that it is from j to state k is given by λjk /λ.

Estimating generator in continuous time

This construction leads to natural estimators for the matrix Λ.
Since λjk is the rate of migrating from j to k we can estimate it by
Njk (T )
λ̂jk = R T , (71)
0 Yj (t) dt
where Njk (T ) is the total number of observed transitions from j to k
in [0, T ] and Yj (t) is the number of obligors with rating j at time t.
The denominator represents the total time spent in state j by all the
companies in the dataset.
© QRM Tutorial Section 10.2.2
 Note that this is the continuous-time analogue of the maximum likelihood
estimator in (70).
It can be shown to be the maximum likelihood estimator for the transition
rates in a homogenous continuous-time Markov chain.

© QRM Tutorial Section 10.2.2

10.3 Structural models of default
10.3.1 The Merton model
Merton’s model (1974) is the prototype of all firm value models.
Consider firm with stochastic asset-value (Vt ), financing itself by equity
(i.e. by issuing shares) and debt.
Assume that debt consists of single zero coupon bond with face or
nominal value B and maturity T .
Denote by St and Bt the value at time t ≤ T of equity and debt so that
Vt = St + Bt , 0 ≤ t ≤ T.

Assume that default occurs if the firm misses a payment to its debt
holders and hence only at T .

© QRM Tutorial Section 10.3

Equity and debt as contingent claims on assets
At T we have two possible cases:
1) VT > B. In that case the debtholders receive B; shareholders receive
residual value ST = VT − B, and there is no default.
2) VT ≤ B. In that case the firm cannot meet its financial obligations,
and shareholders hand over control to the bondholders, who liquidate
the firm; hence we have BT = VT , ST = 0.
In summary we obtain
ST = (VT − B)+
BT = min(VT , B) = B − (B − VT )+ .
The value of equity at T equals the pay-off of a European call option
on VT with exercise price equal to B.
The value of the debt at T equals the nominal value of debt minus the
pay-off of a European put option on VT .
© QRM Tutorial Section 10.3.1
 The option interpretation explains certain conflicts of interest between
shareholders and bondholders.
For example, shareholders have more interest in the firm taking on risky
projects/investments since the value of an option increases with the
volatility of the underlying security.
Bondholders have a short position on the firm’s assets and would like to
see the volatility reduced.

© QRM Tutorial Section 10.3.1

The asset value process
It is assumed that asset value (Vt ) follows a diffusion of the form
dVt = µV Vt dt + σV Vt dWt
for constants µV ∈ R, σV > 0, and a Brownian motion (Wt )t≥0 , so that
 1 
VT = V0 exp (µV − σV2 )T + σV WT ;
2
in particular ln VT ∼ N(ln V0 +(µV − 12 σV2 )T, σV2 T ). The default probability
is thus
!
ln VB0 − (µV − 21 σV2 )T
P(VT ≤ B) = P(ln VT ≤ ln B) = Φ √ ; (72)
σV T
it is increasing in B and σV (for V0 > B) and decreasing in V0 and µV .

© QRM Tutorial Section 10.3.1

A default path

1.0
0.9
Vt

0.8
0.7
0.6

0.0 0.2 0.4 0.6 0.8 1.0

© QRM Tutorial Section 10.3.1

A non-default path

1.1
1.0
0.9
Vt

0.8
0.7
0.6

0.0 0.2 0.4 0.6 0.8 1.0

© QRM Tutorial Section 10.3.1

10.3.2 Pricing in Merton’s model
Under some technical assumptions we can price equity and debt using
the Black–Scholes formula.
The assumptions are that:
1) The risk-free rate is deterministic and equal to r ≥ 0.
2) The asset-value process (Vt ) is independent of the debt level B.
3) The asset value (Vt ) can be traded on a frictionless market.
Recall that equity is a call option on the asset value (Vt ). Hence
Black–Scholes formula yields
St = C BS (t, Vt ; σV , r, T, B) := Vt Φ(dt,1 ) − Be−r(T −t) Φ(dt,2 ),
where the arguments are given by
ln VBt + (r + 21 σV2 )(T − t) √
dt,1 = √ , dt,2 = dt,1 − σV T − t.
σV T − t
© QRM Tutorial Section 10.3.2
Pricing of debt
The price at t ≤ T of a default-free zero-coupon bond with maturity T
and a face value of one equals
p0 (t, T ) = exp(−r(T − t)).

The value of the firm’s debt equals the difference between the value of
default-free debt and a put option on (Vt ) with strike B, i.e.
Bt = Bp0 (t, T ) − P BS (t, Vt ; r, σV , B, T ).

The Black–Scholes formula for European puts now yields

Bt = p0 (t, T )BΦ(dt,2 ) + Vt Φ(−dt,1 ). (73)

The path of (Bt ) is shown on the previous plots. The value of default-free
debt Bp0 (t, T ) is shown as a green curve.

© QRM Tutorial Section 10.3.2

Risk-neutral and physical default probabilities
Under the risk-neutral measure Q the process (Vt ) satisfies the SDE
dVt = rVt dt + σV Vt dW̃t for a standard Q-Brownian motion W̃ .
The drift µV is replaced by the risk-free interest rate r.
Hence the risk-neutral default probability is given by
ln B − ln V0 − (r − 12 σV2 )T
 
q = Q(VT ≤ B) = Φ √ .
σV T
Comparison with physical default probability p = P(VT ≤ B) yields
µV − r √
 
q = Φ Φ−1 (p) + T . (74)
σV
The correction term (µV − r)/σV equals the Sharpe ratio of the firm’s
assets (a popular measure of the risk premium earned by the firm).
The formula is sometimes applied in practice to go from physical to
risk-neutral default probabilities.
© QRM Tutorial Section 10.3.2
Credit spreads in Merton’s model
The credit spread measures the difference between the (continuously
compounded) yield of a default-free zero coupon bond p0 (t, T ) and a
defaultable zero coupon bond p1 (t, T ), i.e.
−1
c(t, T ) = (ln p1 (t, T ) − ln p0 (T − t))
T −t
−1 p1 (t, T )
= ln .
T − t p0 (t, T )
1
In Merton’s model we have p1 (t, T ) = B Bt and hence
 
−1 Vt
c(t, T ) = ln Φ(dt,2 ) + Φ(−dt,1 ) . (75)
(T − t) Bp0 (t, T )
For a fixed time to maturity c(t, T ) depends only on σV and on the
ratio Bp0 (t, T )/Vt (a measure of indebtedness of the firm).
In line with economic intuition it is increasing in both quantities.
© QRM Tutorial Section 10.3.2
Illustration of credit spreads in Merton’s model

6
5
credit spread (%)

4
3
2
1
0

0.1 0.2 0.3 0.4 0.5

volatility
1.2
credit spread (%)

0.8
0.4
0.0

0 1 2 3 4 5

time to maturity

Credit spread c(t, T ) (%) as function of σV (top) and time to maturity T − t

(bottom) for fixed debt to firm value ratio 0.6. In upper picture T − t = 2; in
lower picture σV = 0.25.
© QRM Tutorial Section 10.3.2
10.3.3 Structural models in practice: EDF and DD
A number of industry models descend from the Merton model.
An important example is the so-called public-firm EDF model that is
maintained by Moody’s Analytics.
The methodology builds on earlier work by KMV (a private company
named after its founders Kealhofer, McQuown and Vasicek) in the 1990s.
Literature: Crosbie and Bohn (2002) and Sun et al. (2012).
Expected Default Frequency. The EDF is an estimate of the default
probability of a given firm over a one-year horizon.
Suppose we use Merton’s model for a company issuing debt with face
value B maturing at time T = 1. The analogous quantity would be
!
ln V0 − ln B + (µV − 21 σV2 )
EDFM erton =1−Φ . (76)
σV

© QRM Tutorial Section 10.3.3

How Moody’s adapt the Merton formula
The decreasing function 1 − Φ is replaced by an empirically estimated
function.
B is replaced by a new default threshold B̃ representing the structure
of the firm’s liabilities more closely.
The term (µV − 12 σV2 ) in the numerator is usually omitted.
The current asset value V0 and the asset volatility σV are inferred or
‘backed out’ from information about the firm’s equity value.
Why?
◮ In contrast to the assumptions underlying Merton’s model, in most
cases there is no market for the assets of a firm, so that the asset
value is not directly observable.
◮ The market value can differ widely from the value of a company as
measured by accountancy rules (the so-called book value).
© QRM Tutorial Section 10.3.3
Inferring asset values in Merton’s model
Recall that in Merton’s model we have that
St = C BS (t, Vt ; r, σV , B, T ). (77)

We consider the debt structure (B and T ) as well as the interest rate r

to be known. Equity values (St ) are observable.
For fixed t, (77) is an equation with two unknowns, Vt and σV .
To overcome this difficulty an iterative procedure is used.
(0)
In step (1), an initial estimate σV is used to infer a time series of asset
(0)
values (Vt ) from equity values (St ).
(1)
Then a new volatility estimate σV is estimated from this time series.
(1) (1)
A new time series (Vt ) is then constructed using (77) with σV .
This procedure is iterated n-times, until the volatility estimates converge.

© QRM Tutorial Section 10.3.3

 The procedure in the public-firm EDF model is similar but a more
sophisticated capital structure is assumed and the BS formula in (77) is
replaced by a more complex formula.

EDF and DD
In the public-firm EDF model a new state variable is introduced. This
is the so-called distance-to-default (DD), given by
DD := (log V0 − log B̃)/σV . (78)
Here B̃ represents the default threshold; in some versions of the model
B̃ is modelled as the sum of the liabilities payable within one year and
half of the longer term debt.
Note that (78) is in fact an approximation of the argument of (76),
since µV and σV2 are usually small.
It is assumed that the distance-to-default ranks firms in the sense that
firms with a higher DD exhibit a higher default probability.
© QRM Tutorial Section 10.3.3
 The functional relationship between DD and EDF is determined em-
pirically; using a database of historical default events, the proportion
of firms with DD in a given small range that default within a year is
estimated. This proportion is the empirically estimated EDF.

Variable J&J RadioShack Notes

Market value of assets V0 $236 bn $1834 mo Determined from time series
Asset volatility σV 11% 24% of equity prices.
Default threshold B̃ $39 bn $1042 m Short-term liabilities and half
of long-term liabilities.
DD 16.4 2.3 Given by (log V0 − log B̃)/σV .
EDF (one year) 0.01% 3.58% Using empirical mapping be-
tween DD and EDF.

The example is taken from Sun et al. (2012); it is concerned with the situation of Johnson
and Johnson (J&J) and RadioShack as of April 2012.

© QRM Tutorial Section 10.3.3

10.3.4 Credit migration models revisited
In a credit migration model, consider a firm rated j at t = 0 with
transition probabilities pjk , 0 ≤ k ≤ n for the period [0, T ].
Suppose that the asset-value process (Vt ) of the firm follows the Merton
diffusion model so that
VT = V0 exp((µV − 21 σV2 )T + σV WT ) (79)
is lognormally distributed.
We can choose thresholds 0 = d˜0 < d˜1 < · · · < d˜n < d˜n+1 = ∞ such
that P(d˜k < VT ≤ d˜k+1 ) = pjk for k ∈ {0, . . . , n}.
Thus we have translated the transition probabilities into a series of
thresholds for an assumed asset-value process.
The threshold d˜1 is the default threshold, often interpreted as the value
of the firm’s liabilities.

© QRM Tutorial Section 10.3.4

 The higher thresholds are the asset-value levels that mark the boundaries
of higher rating categories.
The firm-value model can be summarized by saying that the firm belongs
to rating class k at the time horizon T if and only if d˜k < VT ≤ d˜k+1 .
The migration probabilities remain invariant under simultaneous strictly
increasing transformations of VT and the thresholds d˜j .
If we define
ln VT − ln V0 − (µV − 21 σV2 )T
XT := √ , (80)
σV T
ln d˜k − ln V0 − (µV − 21 σV2 )T
dk := √ , (81)
σV T
then we can also say that the firm belongs to rating class k at the time
horizon T if and only if dk < XT ≤ dk+1 .
Observe that XT is a standardized version of the asset-value log-return
√
ln VT − ln V0 and we can easily verify that XT = WT / T ∼ N (0, 1).
© QRM Tutorial Section 10.3.4
 In this case the formulas for the thresholds are easily obtained and are
P
dk = Φ−1 ( k−1l=0 pjl ) for k = 1, . . . , n.

Credit migrations and public-firm EDFs compared

Advantages of EDFs.
1) The EDF reacts quickly to changes in the economic prospects of a firm,
whereas agencies are often slow to adjust ratings.
2) EDFs tend to reflect the current macroeconomic environment and tend
to be better predictors of default over short time horizons.
Advantages of credit migration approach.
1) The EDF approach is sensitive to over- and under-reactions in equity
markets. If widely followed this might have destabilizing effects.
2) As rating agencies focus on average credit quality “through the business
cycle”, risk capital requirements based on rating transitions fluctuate
less, helping to provide liquidity in credit markets.
© QRM Tutorial Section 10.3.4
10.4 Bond and CDS pricing in hazard rate models
10.4.1 Hazard rate models
These are the simplest reduced-form credit risk models.
A hazard rate model is a model in which the distribution of the default
time of an obligor is directly specified by a hazard function without
modelling the mechanism by which default occurs.
To set up a hazard rate model we consider a probability space (Ω, F, P)
and a random default time τ defined on this space, i.e. an F-measurable
rv taking values in [0, ∞].
We denote the df of τ by F (t) = P(τ ≤ t) and the tail or survival
function by F̄ (t) = 1 − F (t); we assume that P(τ = 0) = F (0) = 0,
and that F̄ (t) > 0 for all t < ∞.

© QRM Tutorial Section 10.4

 The jump or default indicator process (Yt ) associated with τ is
Yt = I{τ ≤t} , t ≥ 0. (82)

(Yt ) is a right-continuous process which jumps from 0 to 1 at the default

time τ .
1 − Yt = I{τ >t} is the survival indicator of the firm at time t.

Definition 10.1 (cumulative hazard and hazard function)

The function Γ(t) = − ln(F̄ (t)) is called the cumulative hazard function
of the random time τ . If F is absolutely continuous with density f , the
function
f (t) f (t) d
γ(t) = = = − ln(F̄ (t))
1 − F (t) F̄ (t) dt
is called the hazard function of τ .

The hazard function γ(t) gives the hazard rate at t, which is a measure
of the instantaneous risk of default at t, given survival up to time t.
© QRM Tutorial Section 10.4.1
 We can represent the survival function of τ by
 Z t 
F̄ (t) = exp − γ(s) ds . (83)
0

We may show that

1 1 F (t + h) − F (t)
lim P(τ ≤ t + h | τ > t) = lim = γ(t).
h→0 h F̄ (t) h→0 h
Example 10.2 (Weibull distribution)
For illustrative purposes we determine the hazard function for the Weibull
distribution with df F (t) = 1 − exp(−λtα ) for parameters λ, α > 0.
Differentiation yields
f (t) = λαtα−1 exp(−λtα ) and γ(t) = λαtα−1 .
In particular, γ is decreasing in t if α < 1 and increasing if α > 1. For
α = 1 (exponential distribution) the hazard rate equals the constant λ.

© QRM Tutorial Section 10.4.1

Introducing filtrations
Filtrations model information available to investors over time.
A filtration (Ft ) on (Ω, F) is an increasing family {Ft : t ≥ 0} of
sub-σ-algebras of F : Ft ⊂ Fs ⊂ F for 0 ≤ t ≤ s < ∞.
Ft represents the state of knowledge of an observer at time t. A ∈ Ft
means that at time t we can determine if A has occurred.
In this section we assume that only observable quantity is the default
indicator (Yt ) associated with τ . The appropriate filtration is (Ht ) with
Ht = σ({Yu : u ≤ t}), (84)
the default history up to and including time t.
τ is a (Ht )-stopping time, since {τ ≤ t} = {Yt = 1} ∈ Ht for all t ≥ 0.
In order to study bond and CDS pricing in hazard rate models we need
to compute conditional expectations with respect to the σ-algebra Ht .

© QRM Tutorial Section 10.4.1

A useful result
Lemma 10.3
Let τ be a default time with jump indicator process Yt = I{τ ≤t} and
natural filtration (Ht ). Then, for any integrable rv X and any t ≥ 0,
we have
E(I{τ >t} X)
E(I{τ >t} X | Ht ) = I{τ >t} . (85)
P(τ > t)

This result can be used to determine conditional survival probabilities. For

t < T , applying (85) with X := I{τ >T } we get
 Z T 
P(τ > T | Ht ) = I{τ >t} exp − γ(s) ds , t<T. (86)
t

© QRM Tutorial Section 10.4.1

Martingale property of jump indicator process

Proposition 10.4
The process (Mt ) defined as
Z t
Mt = Yt − I{τ >u} γ(u) du, t≥0
0
is an (Ht )-martingale, that is E(Ms | Ht ) = Mt for all 0 ≤ t ≤ s < ∞.

10.4.2 Risk-neutral pricing revisited

According to the first fundamental theorem of asset pricing, a model for
security prices is arbitrage free if and (essentially) only if it admits at
least one equivalent martingale measure Q.
When building a model for pricing derivatives it is a natural shortcut
to model the objects of interest—such as interest rates and default
times—directly, under a martingale measure Q.
© QRM Tutorial Section 10.4.2
Martingale modelling
So-called martingale modelling is particularly convenient if the value H
of the underlying assets at some maturity date T is exogenously given,
as in the case of zero-coupon bonds.
The underlying asset at time t < T can be computed as the conditional
expectation under Q of the discounted value at maturity via the risk-
neutral pricing rule
RT

Q − rs ds
Vt = E e t H | Ft . (87)

Model parameters are determined using the requirement that at time

t = 0 the model price should coincide with the market price of the
security; this is known as calibration to market data.

© QRM Tutorial Section 10.4.2

Pros and cons of Martingale modelling
Martingale modelling ensures that the resulting model is arbitrage free,
which is important for pricing many different securities simultaneously.
The approach is frequently adopted in default-free term structure models
and in reduced-form models for credit-risky securities.
Martingale modelling is however problematic if the underlying market
is incomplete (meaning that not all risk cannot be hedged away). In
practice martingale modelling is best applied in situations where many
liquidly traded derivatives are available.

10.4.3 Bond pricing

It suffices to consider zero-coupon bonds.
We use martingale modelling and work directly under some martingale
measure Q.

© QRM Tutorial Section 10.4.3

 We assume that under Q the default time τ is a random time with
deterministic risk-neutral hazard function γ Q (t).
The information available to investors at time t is given by the sigma
algebra Ht = σ({Yu : u ≤ t}).
We take interest rates and recovery rates to be deterministic.
The percentage loss given default is denoted by δ ∈ (0, 1).
The continuously compounded interest rate is denoted by r(t) ≥ 0.
The price of the default-free zero-coupon bond with maturity T ≥ t is
R
p0 (t, T ) = exp(− tT r(s) ds).

© QRM Tutorial Section 10.4.3

Analysing the payments
The payments of a defaultable zero-coupon bond can be represented
as a combination of a survival claim that pays one unit at the maturity
date T and a recovery payment in case of default.
The survival claim has pay-off I{τ >T } .
Recall from (86) that
Z T


Q(τ > T | Ht ) = I{τ >t} exp − γ Q (s) ds
t

and define R(t) = r(t) + γ Q (t).

Then the price of a survival claim at time t equals
 Z T 
EQ (p0 (t, T )I{τ >T } | Ht ) = exp − r(s) ds Q(τ > T | Ht )
t
 Z T 
= I{τ >t} exp − R(s) ds . (88)
t

© QRM Tutorial Section 10.4.3

 Note that for τ > t, this can be viewed as the price of a default-free
zero-coupon bond with adjusted interest rate R(t) > r(t).
A similar relationship between defaultable and default-free bond prices
can be established in many reduced-form credit risk models.

Recovery models
1) Recovery of Treasury (RT).
The RT model was proposed by Jarrow and Turnbull (1995).
If default occurs at some point in time τ ≤ T , the owner of the
defaulted bond receives (1 − δτ ) units of the default-free zero-coupon
bond p0 (· , T ) at time τ , where δτ ∈ [0, 1] models the percentage
loss given default.
At maturity T the holder of the defaultable bond therefore receives
the payment I{τ >T } + (1 − δτ )I{τ ≤T } .

© QRM Tutorial Section 10.4.3

2) Recovery of Face Value (RF).
Under RF, if default occurs at τ ≤ T , the holder of the bond receives
a recovery payment of size (1 − δτ ) immediately at the default time τ .
Note that even with deterministic loss given default and deterministic
interest rates, the value at maturity of the recovery payment is
random as it depends on the exact timing of default.
RF is slightly more realistic; RT is slightly easier to analyse.

Pricing recovery payment under RT

The value of the recovery payment at the maturity date T is
(1 − δ)I{τ ≤T } = (1 − δ) − (1 − δ)I{τ >T } .

Using (88), the value of the recovery payment at time t < T is hence
 Z T 
(1 − δ)p0 (t, T ) − (1 − δ)I{τ >t} exp − R(s) ds .
t
© QRM Tutorial Section 10.4.3
 Hence the value of the bond is
 Z T 
p1 (t, T ) = (1 − δ)p0 (t, T ) + δI{τ >t} exp − R(s) ds .
t

Pricing recovery payment under RF

Under the RF-hypothesis the recovery payment takes the form (1 −
δ)I{τ ≤T } where the payment occurs directly at time τ .
A payments of this form is a payment-at-default claim.
The value of the recovery payment at time t ≤ T equals
  Z τ  
Q
E (1 − δ)I{t<τ ≤T } exp − r(s) ds Ht .
t

Using (85) we may show that

  Z τ  
EQ (1 − δ)I{t<τ ≤T } exp − r(s) ds Ht
t
Z T  Z s 
Q
= (1 − δ)I{τ >t} γ (s) exp − R(u) du ds.
t t
© QRM Tutorial Section 10.4.3
10.4.4 CDS pricing
Recap: Structure of CDS
C (reference entity); default at time τC < T triggers default payment.
A (protection buyer); pays premiums to B until min(τC , T ).
B (protection seller); makes default payment to A if τC < T .

C
✻
❄
✛premium payments at fixed times
yes: default payment ✲
B C defaults ? ✲ A
no: no payment ✲

© QRM Tutorial Section 10.4.4

Payment flows
For simplicity write τ = τC and consider the following contract:
Premium payments.
◮ These are due at times 0 < t1 < · · · < tN measured in years.
◮ If τ > tk , A pays a premium of size x∗ (tk − tk−1 ) at tk , where x∗
denotes the fair swap spread.
◮ After τ premium payments stop.
◮ No initial payment.

Default payment.
◮ If τ < tN = T , B makes a default payment δ at τ .
◮ Sometimes B receives an accrued premium payment of size x∗ (τ − tk )
for τ ∈ (tk , tk−1 ). We ignore this feature for simplicity.

© QRM Tutorial Section 10.4.4

Valuing the premium leg
The premium leg consists of a set of survival claims.
Introduce a function of x given by
Vtprem (x; γ Q )
 X  Z tk  
= EQ exp − r(u) du x(tk − tk−1 )I{τ >tk } | Ht
k : tk >t t
X
=x p0 (t, tk )(tk − tk−1 )Q(τ > tk | Ht ),
k : tk >t
R tk Q
which is easily computed using Q(τ > tk | Ht ) = exp(− t γ (s) ds).
We obtain
X  Z tk 
Vtprem (x; γ Q ) = I{τ >t} x (tk − tk−1 ) exp − R(u) du .
k : tk >t t

© QRM Tutorial Section 10.4.4

Valuing the default leg
The default payment leg is a typical payment-at-default claim.
We obtain
Vtdef (γ Q )
  Z τ  
Q
=E δI{t<τ ≤tN } exp − r(s) ds Ht
t
Z tN  Z s 
Q
= I{τ >t} δ γ (s) exp − R(u) du ds.
t t

The fair CDS spread

The fair CDS spread x∗t quoted for the contract at time t is chosen such
that the value of the contract is equal to zero.

© QRM Tutorial Section 10.4.4

 The equation Vtprem (x∗t ; γ Q ) = Vtdef (γ Q ) yields
R tN Q  R 
tδ γ (s) exp − ts R(u) du ds
x∗t = I{τ >t} P  R . (89)
tk
k : tk >t (tk − tk−1 ) exp − t R(s) ds

Model calibration
We have to calibrate our model to the available market information.
Hence we have to determine the implied risk-neutral hazard function
γ Q (t), which ensures that the fair CDS spreads implied by the model
equal the spreads quoted in the market.
Suppose that the market information at time t = 0 consists of the fair
spread x∗ of one CDS with maturity tN .
In that case γ Q (s) is taken constant: for all s ≥ 0, γ Q (s) = γ̄ Q for some
γ̄ Q > 0.

© QRM Tutorial Section 10.4.4

 γ̄ Q has to solve the equation
N
X Z tN
∗ −γ̄ Q tk Qt
x p0 (0, tk )(tk − tk−1 )e = δγ̄ Q
p0 (0, t)e−γ̄ dt.
k=1 0

There is a unique solution.

If we observe spreads for several CDSs on the same reference entity but
with different maturities, a constant function is not sufficient. Instead
one typically uses piecewise constant or linear hazard functions.
An exception occurs in the special case where: (1) the spread curve
is flat (i.e. all CDSs on the reference entity have the same spread x∗ ,
independent of the maturity); (2) the risk-free interest rate is constant;
(3) the time points tk are equally spaced (tk − tk−1 = ∆t for all k).
In that case the implied risk-neutral hazard rate γ̄ Q is the solution of
Z ∆t
Q ∆t Qt
x∗ ∆t p0 (0, ∆t)e−γ̄ = δγ̄ Q e−rt e−γ̄ dt. (90)
0
© QRM Tutorial Section 10.4.4
 For ∆t relatively small (quarterly or semi-annual spread payments) a
good approximation to the solution of (90) is given by γ̄ Q ≈ x∗ /δ.
This approximation is frequently used in practice and implies that the
Q
one-year default probability satisfies Q(τ ≤ 1) = 1 − e−γ̄ ≈ γ̄ Q ≈ x∗ /δ.

Shortcomings of simple hazard rate models

In the models of this section the only risk factor affecting a defaultable
bond or CDS is default risk.
In these models credit spreads evolve deterministically prior to default,
which is unrealistic.
The models are not sophisticated enough to price options on defaultable
bonds or CDSs.
To obtain more realistic models we can replace the deterministic hazard
functions by stochastic hazard processes.

© QRM Tutorial Section 10.4.4

 This means that default times are modelled as so-called doubly-stochastic
random times.
We might also consider adding stochastic interest rate models; and more
complex assumptions on recoveries in the event of default.

10.4.5 P versus Q: Empirical results

There are some empirical studies of the relationship between physical
and risk-neutral default probabilities.
Risk-neutral default probabilities are generally estimated from CDS
spreads. These can be compared, for example, with EDFs.
Berndt et al. (2008) compare five-year CDSs against five-year EDFs for
a large pool of firms. The five-year EDF is an annualized estimate of
the physical five-year default probability.

© QRM Tutorial Section 10.4.5

 Let x∗t,i and EDFt,i denote the CDS spread and five-year EDF of firm i
at date t. Their (most basic) model took the form
x∗t,i = α + β EDFt,i +εt,i ,
with estimates α = 33bp and β = 1.6; the R2 was 0.73.
More crudely x∗t,i / EDFt,i ≈ 1.6.
Using qt,i = x∗t,i /δ as a proxy for the risk-neutral default probability
yields
qt,i x∗t,i
≈ ≈ 1.6 δ −1 .
pt,i δ EDFt,i

© QRM Tutorial Section 10.4.5

10.5 Pricing with stochastic hazard rates
Why stochastic hazard rates?
In hazard rate models the only risk factor is default risk ⇒ Credit spreads
evolve deterministically prior to default, which is clearly unrealistic.
Moreover, it is not possible to price options on bonds or CDSs or to do
risk management for bond portfolios in such models.
Hence it is of interest to consider models where hazard rate is a stochastic
process (γt )t≥0 ; typically hazard rate is driven by a second stochastic
process Ψ, that is γt = γ(Ψt ).
Simplest such model class are doubly-stochastic random times.

© QRM Tutorial Section 10.5

10.5.1 Doubly stochastic random times
Setup. We work on (Ω, F, P) with background filtration (Ft ) containing
information about all other economic events except the default event.

Consider a random time τ , that is a measurable rv with values in (0, ∞).

Yt = I{τ ≤t} is the associated default indicator and (Ht ) = σ{Ys , s ≤ t}
is the default history up to t.
Define new filtration Gt = Ft ∨ Ht , t ≥ 0, i.e. Gt contains background
info Ft and default history up to t (this is the information available to
investors).

© QRM Tutorial Section 10.5.1

Definition 10.5
τ is called doubly stochastic if there is a positive (Ft )-adapted process
(γt ) (the hazard rate process) such that for all t ≥ 0
 Z t 
P(τ > t | F ∞ ) = exp − γs ds . (91)
0

Comments.
S
Here F∞ = σ( t≥0 Ft ). Conditioning on F∞ thus means that we know
the past and future economic environment and in particular the entire
trajectory (γs (ω))s≥0 of the hazard rate.
Relation (91) implies that, given the economic environment F ∞ , τ is a
random time with deterministic hazard function s 7→ γs (ω).
In the literature doubly stochastic random times are also known as
conditional Poisson or Cox random times.

© QRM Tutorial Section 10.5.1

Sampling doubly stochastic random times
A simple algorithm is based on the following result

Lemma 10.6
Let E be a standard exponentially distributed rv independent of F ∞ ,
that is P(τ > t | F ∞ ) = e−t . Let (γt ) be a positive Ft -adapted process
R
with 0t γs ds) < ∞ for all t. Define τ by
n Z t o
τ := inf t ≥ 0 : γs ds ≥ E . (92)
0
Then τ is doubly stochastic with hazard-rate process (γt ).

Algorithm. (threshold simulation)

1) Generate E ∼ Exp(1).
2) Generate a trajectory (γs )∞
s=0 of hazard rate process.
Rt
3) Return τ := inf{t ≥ 0 : 0 γs ds ≥ E}
© QRM Tutorial Section 10.5.1
Graphical illustration.

0.8

0.6

E
Γ 0.4

0.2

0 τ
0 2 4 6 8 10
Time

A graphical illustration of threshold simulation; E ≈ 0.44, τ ≈ 6.59.

© QRM Tutorial Section 10.5.1

Intensities
Definition 10.7
Consider a filtration (Gt ) and a random time τ with (Gt )-adapted jump
indicator process (Yt ). A non-negative (Gt )-adapted process (λt ) is called
R
(Gt )-(default) intensity of the random time τ if Mt := Yt − 0t∧τ λs ds
is a (Gt )-martingale.

The next result extends Proposition 10.4 to doubly stochastic τ .

Proposition 10.8
Let τ be a doubly stochastic random time with (Ft )-conditional hazard
R
rate process (γt ). Then Mt := Yt − 0t∧τ γs ds is a (Gt )-martingale, that
is the hazard rate γt is the (Gt ) default intensity.

Conditional expectations
Conditional expectations wrt Gt are crucial for pricing formulas.
© QRM Tutorial Section 10.5.1
Proposition 10.9 (Dellacherie formula)
Let τ be an arbitrary random time (not necessarily doubly stochastic)
such that P(τ > t | Ft ) > 0 for all t ≥ 0. Then we have for every
integrable rv X that
E(I{τ >t} X | Ft )
E(I{τ >t} X | Gt ) = I{τ >t} .
P(τ > t | Ft )

Corollary 10.10
Let T > t and assume that τ is doubly-stochastic with hazard
rate process (γt ). If X̃ is integrable
R
and FT -measurable, we have
T
E(I{τ >T } X̃ | Gt ) = I{τ >t} E(e− t
γs ds
X̃ | Ft ).

© QRM Tutorial Section 10.5.1

Application: 1-year default probabilities
γt gives good approximation to the one-year default probability: Let T =
t + 1 and X̃ = 1 to obtain
  Z t+1  
P(τ > t + 1 | Gt ) = I{τ >t} E exp − γs ds Ft . (93)
t

For τ > t and a fairly stable hazard rate over the time interval [t, t + 1]
the right-hand side of (93) is ≈ exp(−γt ) and for γt small,
P(τ ≤ t + 1 | Gt ) ≈ 1 − exp(−γt ) ≈ γt . (94)

© QRM Tutorial Section 10.5.1

10.5.2 Pricing formulas
Setup.
Consider arbitrage-free security market model on (Ω, F, (Ft ), Q) where
Q is equivalent martingale measure. Prices of default-free securities
R
(Ft )-adapted; Bt = exp( 0t rs ds) models default-free savings account.
Let τ be the default time of some company. As before we set Ht =
σ({Ys : s ≤ t}) and Gt = Ft ∨ Ht ; this is the information available to
investors at time t.
We use martingale-modelling. Hence price at t of an GT -measurable
contingent claim H is given by
  Z T  
Ht = EQ exp − rs ds H Gt . (95)
t

Under Q, τ is a doubly stochastic random time with background filtra-

tion (Ft ) and hazard rate process (γt ).
© QRM Tutorial Section 10.5.2
Key building blocks
The pricing of bonds and CDSs can be reduced to the pricing of the
following building blocks:
A survival claim, i.e. a promised FT -measurable payment X which is
made at time T if there is no default; the actual payment of the survival
claim equals XI{τ >T } .
A payment-at-default claim of the form Zτ I{τ ≤T } , where Z = (Zt )t≥0 is
an (Ft ) adapted stochastic process and where Zτ is short for Zτ (ω) (ω).
Note that the payment is made directly at τ , provided that τ ≤ T
where T is the maturity date of the claim.

Example. Defaultable bond is a combination of a survival claim and a

payment at default claim (the recovery payment).

© QRM Tutorial Section 10.5.2

Pricing the building blocks
Next result shows that pricing of building blocks can be reduced to pricing
problem for default-free claims with adjusted interest rate.
Theorem 10.11
Define adjusted interest rate Rt = rt + γt . Under the above assumptions
(in particular for τ doubly stochastic) it holds that
  Z T     Z T  
EQ exp − rs ds I{τ >T } X Gt = I{τ >t} EQ exp − Rs ds X Ft ,
t t
  Z τ  
EQ I{t<τ ≤T } exp − rs ds Zτ Gt
t
Z T  Z s  
= I{τ >t} EQ Zs γs exp − Ru du ds Ft .
t t

© QRM Tutorial Section 10.5.2

10.5.3 Applications
Corporate bonds and RF. The price at t of a defaultable zero-coupon
bond with maturity T ≥ t is p1 (t, T ); price of corresponding default-free
bond is p0 (t, T ).

Recall the following recovery models

1) Recovery of Treasury (RT). Under RT, if default occurs at τ ≤ T ,
the bond holder receives (1 − δτ ) units of p0 (· , T ) at time τ , where
δ ∈ [0, 1] models the percentage loss given default (LGD). Under RT
the holder of the defaultable bond therefore receives the payment
I{τ >T } + (1 − δ)I{τ ≤T } = 1 − δ + δI{τ >T } .

2) Recovery of Face Value (RF). Under RF, if default occurs at τ ≤ T ,

the bondholder receives (1 − δτ ) immediately at τ . ⇒ Value of the
recovery payment depends on the timing of default.

© QRM Tutorial Section 10.5.3

3) Recovery of market value (RM). Duffie and Singleton (1999) Under RM
recovery payment equals (1 − δτ )Vτ I{τ ≤T } , where (δt ) ∈ (0, 1) gives
the percentage LGDl and where the (Ft )-adapted process (Vt ) gives
the pre-default value of the claim. This is a recursive definition (but
explicit solution exists)

Application to corporate bonds.

1) Under RT the bond price in t < T is
  Z T  
p1 (t, T ) = (1 − δ)p0 (t, T ) + I{τ >t} δEQ exp − Rs ds Ft
t

2) Under RF the bond is the sum of a survival claim and a payment at

© QRM Tutorial Section 10.5.3

 default claim. One has
   Z T  
Q
p1 (t, T ) = I{τ >t} E exp − Rs ds | Ft
t
Z T  Z s  
+ (1 − δ)EQ γs exp − Ru du ds | Ft )
t t

3) Under RM assumption one has

Proposition 10.12
Suppose that, under Q, τ is doubly stochastic with hazard rate pro-
cess (γt ). Then under RM the pre-default value (Vt ) of a corporate
bond is uniquely determined and given by
 Z T  
Vt = EQ exp − (rs + δs γs ) ds Ft , 0 ≤ t ≤ T. (96)
t

Special cases: δ = 1 ⇒ standard survival claim; δ = 0 ⇒ default-free.

© QRM Tutorial Section 10.5.3

Credit spreads
With doubly stochastic default times hazard rate process (γt ) and credit
spread c(t, T ) = − T 1−t (ln p1 (t, T ) − ln p0 (t, T )) of defaultable bonds are
closely related. Analytic results for the instantaneous credit spread
∂
c(t, t) = lim c(t, T ) = − (ln p1 (t, T ) − ln p0 (t, T )). (97)
T →t ∂T T =t

Proposition 10.13
In all recovery models c(t, t) = δγtQ .

Instantaneous credit spreads are product of LGD and instantaneous

default probability.
In hazard-rate models short term spreads strictly positive.
For T > t spreads in the three models differ.

© QRM Tutorial Section 10.5.3

CDS contracts
Here premium payments constitute a sequence of survival claims; default
payment is a payment-at-default claim. This gives the following formula
for the fair CDS spread x∗ at t:

RT Rs

∗ δEQ t γs e− t
Ru du
ds | Ft
x =P R tk

. (98)
Tk >t (tk − tk−1 )EQ exp − t Rs ds | Ft

For T → t we get that x∗ converges to δγt .

The formula is a generalization of (89).

© QRM Tutorial Section 10.5.3

10.6 Affine models
In most models with doubly stochastic default time used in practice it is
assumed that (rt ) and (γt ) are functions of some Markov process (Ψt ) on
D ⊂ Rp .
Natural background filtration is (Ft ) = σ({Ψs : s ≤ t}).
Rt := rt + γt is of the form Rt = R(Ψt ) for some R : D ⊆ Rp → R+ .
To evaluate general pricing formulas we hence have to compute conditional
expectations of the form

 RT Z T Rs 
− R(Ψs ) ds
E e t g(ΨT ) + h(Ψs )e− t
R(Ψu ) du
ds | Ft (99)
t
for generic g, h : D → R+ . Since (Ψt ) is Markov, (99) is a function f (t, Ψt )
of time and of Ψt . f can sometimes be computed by solving a PDE; this
is the well-known Feynman-Kac formula.
© QRM Tutorial Section 10.6
Theorem 10.14 (Feynman–Kac)
Consider generic R, g : D → R+ . Suppose that f : [0, T ] × D → R is
bounded, continuous and solves the terminal-value problem
ft + µ(ψ)fψ + 12 σ 2 (ψ)fψψ = R(ψ)f, (t, ψ) ∈ [0, T ) × D, (100)
with f (T, ψ) = g(ψ), ψ ∈ D. Suppose that (Ψt ) is the unique solution
of the SDE
dΨt = µ(Ψt ) dt + σ(Ψt ) dWt , Ψ0 = ψ ∈ D, (101)
with state space D ⊆ R, (Wt ) a standard, Brownian motion and µ and
σ continuous functions from D to R resp. R+ . Then
 RT 
− R(Ψs )ds
E e t g(ΨT ) | Ft = f (t, Ψt ) . (102)

© QRM Tutorial Section 10.6

Comments
The Feynman Kac formula can be used in two ways:
We can use probabilistic techniques or Monte-Carlo simulation to com-
pute the (conditional) expectation (102) in order to solve numerically
the PDE (100).
We can solve the PDE (100) perhaps numerically, in order to compute
the expectation (102).
For an extension to the d-dimensional case (and weaker regularity conditions
on f ) we refer to the literature such as Karatzas and Shreve (1988)

© QRM Tutorial Section 10.6

Affine term structure
Consider a model where r and γ are functions of a diffusion Ψ. Define a
function f by
 RT 
f (t, Ψt ) = E e− t
R(Ψs )ds uΨT
e Ft
where u, D are such uψ ≤ 0 for all ψ ∈ D. Note that for u = 0 we have a
bond price with zero recovery.

Definition 10.15
The model has an affine (defaultable) term structure if
f (t, ψ) = exp(α(t, T ) + β(t, T )ψ) (103)
for deterministic functions α(·, T ) and β(·, T ).

© QRM Tutorial Section 10.6

The following assumption guarantees an affine term structure.

Assumption 10.16 (affine term structure)

R, µ and σ 2 are affine functions of ψ, i.e. there are constants ρ0 , ρ1 ,
k 0 , k 1 , h0 and h1 such that
R(ψ) = ρ0 + ρ1 ψ, µ(ψ) = k 0 + k 1 ψ, σ 2 (ψ) = h0 + h1 ψ.
Moreover, for all ψ ∈ D we have h0 + h1 ψ ≥ 0 and ρ0 + ρ1 ψ ≥ 0.

An ODE system for α and β

The educated guess f (t, ψ) = exp(α(t, T ) + β(t, T )ψ) gives
ft = (α′ + β ′ ψ)f, fψ = βf and fψψ = β 2 f.

© QRM Tutorial Section 10.6

Substituting this in the PDE ft + µ(ψ)fψ + 12 σ 2 (ψ)fψψ ) = R(ψ)f , using
the special form of µ, σ 2 and rearranging terms gives the following ODE
system
β ′ (t, T ) = ρ1 − k 1 β(t, T ) − 12 h1 β 2 (t, T ), β(T, T ) = u, (104)
α′ (t, T ) = ρ0 − k 0 β(t, T ) − 12 h0 β 2 (t, T ), α(T, T ) = 0. (105)

Comments.
The ODE (104) for β(· , T ) is a so-called Ricatti equation.
The ODE (105) for α(· , T ) can be solved by (numerical) integration
once β has been determined.

© QRM Tutorial Section 10.6

Summary. Suppose that the affine-term-structure assumption holds,
that the ODE system (104), (105) has a unique solution (α, β) on [0, T ]
and that there is some C such that β(t, T )ψ ≤ C for all t ∈ [0, T ],
ψ ∈ D. Then the model has an affine term structure.

10.6.1 The CIR square-root diffusion

The CIR or square-root diffusion model due to Cox et al. (1985) is a popular
affine model.

CIR dynamics.
p
dΨt = κ(θ̄ − Ψt ) dt + σ Ψt dWt , Ψ0 = ψ > 0, (106)
for parameters κ, θ̄, σ > 0 and state space D = [0, ∞).

© QRM Tutorial Section 10.6.1

Properties.
(106) is an affine model; the parameters are given by k 0 = κθ̄, k 1 = −κ,
h0 = 0 and h1 = σ 2 .
The SDE (106) admits a global solution (non-trivial)
(106) implies that (Ψt ) is mean reverting.
Mean reversion sufficiently strong ⇒ trajectories never reach zero: Let
τ0 (Ψ) := inf{t ≥ 0 : Ψt = 0}. For κθ̄ ≥ 21 σ 2 , P(τ0 (Ψ) < ∞) = 0; for
κθ̄ < 12 σ 2 , P(τ0 (Ψ) < ∞) = 1.

© QRM Tutorial Section 10.6.1

CIR term structure
Theorem 10.17
Suppose that the factor Ψ follows the CIR model and that adjusted
interest rate is an affine function of the state, R(ψ) = ρ0 + ρ1 ψ. Then
it holds that
  Z T  
E exp − (ρ0 + ρ1 Ψs ) ds Ψt = exp(α(T − t) + β(T − t)Ψt ),
t
where
−2ρ1 (eγτ − 1)
β(τ ) = ,
γ − κ + eγτ (γ + κ)
 
κθ̄
0 2γeτ (γ+κ)/2
α(τ ) = −ρ τ + 2 2 ln ,
σ γ − κ + eγτ (γ + κ)
p
and τ := T − t, γ := κ2 + 2σ 2 ρ1

© QRM Tutorial Section 10.6.1

10.6.2 Extensions
It is possible to extend the above to CIR models with jumps,
p
dΨt = κ(θ̄ − Ψt ) dt + σ Ψt dWt + dZt
P
where Zt = Tn ≤t Zn is a compound Poisson process with jump intensity
λt = λ0 + λ1 Ψt , λ0 and λ1 > 0 and the Zn are iid positive rvs, for
instance exponentially distributed.
The computation of payment-at-default claims is also possible with
“affine model technology”.

© QRM Tutorial Section 10.6.2

Numerical example
7
6
Credit spread (%) 5 Zero recovery

4
3 RM
2
1
0
0 5 10 15 20
Time to maturity

Spreads of defaultable zero-coupon bonds in an affine model for various recovery

assumptions. It holds Ψ0 ≈ 0.0533, r = 6% and δ = 0.5. Note that under the
RF recovery model (dashed line) the spread becomes negative for large times to
maturity; this is not true under other recovery assumptions.
© QRM Tutorial Section 10.6.2