Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

Risk Management and Financial Institutions

By John C. Hull
Chapter 3 How Traders manage Their Exposures ........................................................................2
Chapter 4 Interest Rate Risk .........................................................................................................3
Chapter 5 Volatility.......................................................................................................................5
Chapter 6 Correlations and Copulas .............................................................................................7
Chapter 7 Bank Regulation and Basel II ......................................................................................9
Chapter 8 The VaR Measure .......................................................................................................11
Chapter 9 Market Risk VaR: Historical Simulation Approach ...................................................14
Chapter 10 Market Risk VaR: Model-Building Approach.........................................................16
Chapter 11 Credit Risk: Estimating Default Probabilities .........................................................17
Chapter 12 Credit Risk Losses and Credit VaR .........................................................................20
Chapter 13 Credit Derivatives ...................................................................................................22
Chapter 14 Operational Risk......................................................................................................24
Chapter 15 Model Risk and Liquidity Risk ...............................................................................25
Chapter 17 Weather, Energy, and Insurance Derivatives...........................................................27
Chapter 18 Big Losses and What We Can Learn From Them ...................................................28
T1 Bootstrap .............................................................................................................................30
T2 Principal Component Analysis............................................................................................30
T3 Monte Carlo Simulation Methods.......................................................................................31

-1-
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

Chapter 3 How Traders manage Their Exposures

1. Linear products: a product whose value is linearly dependent on the value of the
underlying asset price. Forward, futures, and swaps are linear products; options
are not.
E.g. Goldman Sachs have entered into a forward with a gold mining firm. Goldman
Sachs borrows gold from a central bank and sell it at the current market price. At the
end of the life of the forward, Goldman Sachs buys gold from the gold mining firm
and uses it to repay the central bank.
2. Delta neutrality is more feasible for a large portfolio of derivatives dependent on
a single asset. Only one trade in the underlying asset is necessary to zero out delta
for the whole portfolio.
3. Gamma: if it is small, delta changes slowly and adjustments to keep a portfolio
delta neutral only need to be made relatively infrequently.
∂ 2Π
Gamma =
∂S 2
- Gamma is positive for a long position in an option (call or put).
- A linear product has zero Gamma and cannot be used to change the gamma of a
portfolio.
4. Vega
- Spot positions, forwards, and swaps do not depend on the volatility of the
underlying market variable, but options and most exotics do.
∂Π
- ν=
∂σ
- Vega is positive for long call and put;
- The volatilities of short-dated options tend to be more variable than the volatilities
of long-dated options.
5. Theta: time decay of the portfolio.
- Theta is usually negative for an option. An exception could be an in-the-money
European put option on a non-dividend-paying stock or an in-the-money European
call option on a currency with a very high interest rate.
- It makes sense to hedge against changes in the price of the underlying asset, but it
does not make sense to hedge against the effect of the passage of time on an
option portfolio. In spite of this, many traders regard theta as a useful statistic. In a
delta neutral portfolio, when theta is large and positive, gamma tends to be large
and negative, and vice versa.
6. Rho: the rate of change of a portfolio with respect to the level of interest rates.
Currency options have two rhos, one for the domestic interest rate and one for
the foreign interest rate.
7. Taylor expansions:
- if ignoring terms of higher order dt and assuming delta neutral and volatility and
interest rates are constant, then

-2-
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

1
ΔΠ = ΘΔt + ΓΔS 2
2
- Options traders make themselves delta neutral – or close to delta neutral at
the end of each day. Gamma and vega are monitored, but are not usually
managed on a daily basis.
- There is one aspect of an options portfolio that mitigates problems of managing
gamma and vega somewhat. Options are often close to the money when they are
first sold so that they have relatively high gammas and vegas. However, as time
elapses, the underlying asset price has often changed sufficiently for them to
become deep out of the money or deep in the money. Their gammas and vegas are
then very small and of little consequence. The nightmare scenario for a trader is
where written options remain very close to the money as the maturity date is
approached.

Chapter 4 Interest Rate Risk

1. LIBID, the London Interbank Bid Rate. This is the rate at which a bank is
prepared to accept deposits from another bank. The LIBOR quote is slightly
higher than the LIBID quote.
- Large banks quote 1, 3, 6 and 12 month LIBOR in all major currencies. A bank
must have an AA rating to qualify for receiving LIBOR deposits.
- How LIBOR yield curve be extended beyond one year? Usually, create a yield
curve to represent the future short-term borrowing rates for AA-rated companies.
- The LIBOR yield curve is also called the swap yield curve or the LIBOR/swap
yield curve.
- Practitioners usually assume that the LIBOR/swap yield curve provides the
risk-free rate. T-rates are regarded as too low to be used as risk-free rates
because:
a. T-bills and T-bonds must be purchased by financial institutions to fulfill a variety
of regulatory requirements. This increase demand for these Ts driving their prices
up and yields down.
b. The amount of capital a bank is required to hold to support an investment in Ts is
substantially smaller than the capital required to support a similar investment in
other very low-risk instruments.
c. In USA, Ts are given a favorable tax treatment because they are not taxed at the
state level.
2. Duration
1 dB n ci e − yti
- Continuous compounding. D=− = ∑ ti ( ), ΔB = − BDΔy .
B dy i =1 B
This is the weighted average of the times when payments are made, with the

-3-
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

weight applied to time t being equal to the proportion of the bond’s total present
value provided by the cash flow at time t.
D
Compounding m times per year, then modified duration. D =
*
-
1+ y / m
3. Convexity.
n

1d B 2 ∑c t e i i
2 − yti

- C= = i =1
. This is the weighted average of the square of the time
B dy 2 B
to the receipt of cash flows.
1
ΔB = − BDΔy + BC (Δy ) 2
2
- The duration (convexity) of a portfolio is the weighted average of the durations
(convexity) of the individual assets comprising the portfolio with the weight
assigned to an asset being proportional to the value of the asset.
- The convexity of a bond portfolio tends to be greatest when the portfolio
provides payments evenly over a long period of time. It is least when the
payments are concentrated around one particular point in time.
- Duration zero Æ protect small parallel shit in the yield curve.
- Both duration and convexity zero Æ protect large parallel shifts in the yield
curve.
- Duration and convexity are analogous to the delta and gamma.
4. Nonparallel yield curve shifts
1 dPi
- Definition: D p = − , where dyi is the size of the small change made to the
P dyi
ith point on the yield curve and dpi is the resultant change in the portfolio value.
- The sum of all the partial duration measures equals the usual duration measure.
5. Interest rate deltas
- DV01: define the delta of a portfolio as the change in value for a one-basis-point
parallel shift in the zero curve. It is the same as the duration of the portfolio
multiplied by the value of the portfolio multiplied by 0.01%.
- In practice, traders like to calculate several deltas to reflect their exposures to all
the different ways in which the yield curve can move. One approach corresponds
to the partial duration approach. Æ the sum of the deltas for all the points in the
yield curve equals the DV01.
6. Principal components analysis
- The interest rate move for a particular factor is factor loading. The sum of the
squares of factor loadings is 1.
- Factors are chosen so that the factor scores are uncorrelated.
- 10 rates and 10 factors Æ solve the simultaneous equations: the quantity of a
particular factor in the interest rate changes on a particular day is known as the

-4-
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

factor score for that day.

- The importance of a factor is measured by the standard deviation of its factor
score.
- The sum of the variances of the factor scores equal the total variance of the
data. Æ the importance of 1st factor = the variance of 1st factor’s factor score/total
variance of the factor scores.
- Use principal component analysis to calculate delta exposures to factors: dP/dF =
dP/dy *dy/dF.

Chapter 5 Volatility

1. What causes volatility?

- It is natural to assume that the volatility of a stock price is caused by new
information reaching the market. Fama (1965), French (1980), and French and
Roll (1986) show that the variance of stock price returns between Friday and
Monday is only 22%, 19% and 10.7% higher than the variance of stock price
return between two consecutive trading days (not 3 times).
- Roll (1984) looked at the prices of orange juice futures. By far the most important
news for orange juice futures prices is news about the weather and news about the
weather is equally likely to arrive at any time. He found that the
Friday-to-Monday variance is only 1.54 times the first variance.
- The only reasonable conclusion from all this is that volatility is to a large extent
caused by trading itself.
2. Variance rate: risk managers often focus on the variance rate rather than the
volatility. It is defined as the square of the volatility.
3. Implied volatilities are used extensively by traders. However, risk management is
largely based on historical volatilities.
4. Suppose that most market investors think that exchange rates are log normally
distributed. They will be comfortable using the same volatility to value all options
on a particular exchange rate. But you know that the lognormal assumption is not
a good one for exchange rates. What should you do? – You should buy
deep-out-of-the-money call and put options on a variety of different
currencies – and wait. These options will be relatively inexpensive and more of
them will close in the money than the lognormal model predicts. The present
value of your payoffs will on average be much greater than the cost of the options.
In the mid-1980s, the few traders who were well informed followed the strategy
and made lots of money. By the late 1980s everyone realized that
out-of-the-money options should have a higher implied volatility than at the
money options and the trading opportunities disappeared.
5. An alternative to normal distributions: the power law- has been found to be a good
descriptions of the tails of many distributions in practice.
- The power law: for many variables, it is approximately true that the value of v of

-5-
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

−α
the variable has the property that, when x is large, Pr ob(v > x ) = Kx , where K
and alpha are constants.
6. Monitoring volatility
- The exponentially weighted moving average model (EWMA).
σ n2 = λσ n2−1 + (1 − λ )un−
2
1 , RiskMetrics use λ =0.94.
- The GARCH(1,1) MODEL
σ n2 = γ VL + α un2−1 + βσ n2−1 , Where γ +α + β =1, if γ =0, then GARCH
model is EWMA
- ML method to estimate GARCH (1,1)
7. How good is the model:
- The assumption underlying a GARCH model is that volatility changes with the
passage of time. If a GARCH model is working well, it should remove the

autocorrelation of ui2 . We can consider the autocorrelation of the variables ui2 / σ i2 .

If these show very little autocorrelation, the model for volatility has succeed in
explaining autocorrelations in the ui2 . We can use Ljung-Box statistic. If this statistic

is greater than 25, zero autocorrelation can be rejected with 95% confidence.
8. Using GARCH to forecast future volatility.
σ n2+t − VL = α (un2+t −1 − VL ) + β (σ n2+t −1 − VL ) , since the expected value of un2+t −1

is σ n2+t −1 , hence:

E[σ n2+ t − VL ] = (α + β ) E[σ n2+ t −1 − VL ] = (α + β )t (σ n2 −VL )

9. Volatility term structures – the relationship between the implied volatilities of
the options and their maturities.
1
Define V (t ) = E (σ n +t ) and a=ln
2
- Æ
α +β
1 T 1 − e − at
T ∫0
- V (t )dt = VL + [V (0) − VL ] , or in per yearÆ
aT
⎧ 1 − e − at ⎫
- σ (T ) = 252 ⎨VL +
2
[V (0) − VL ]⎬ Æ can be used to estimate a
⎩ aT ⎭
volatility term structure based on the GARCH (1,1) model.
10. Impact of volatility changes.
⎧ 1 − e − at σ (0) 2 ⎫
- σ (T ) = 252 ⎨VL + − VL ]⎬ .
2
[
⎩ aT 252 ⎭

-6-
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

1 − e − at σ (0)
When σ (0) change by d σ (0) , σ (T ) changes by Δσ (0)
aT σ (T )
- Many banks use analyses such as this when determining the exposure of their
books to volatility changes.

Chapter 6 Correlations and Copulas

1. Correlation measures linear dependence. There are many other ways in which
two variables can be related. E.g. for normal values, two variables may be
unrelated. However, their extreme values may be related. “During a crisis the
correlations all go to one.”
2. Monitoring correlation
- Using EWMA:
cov n = λ cov n −1 + (1 − λ ) xn −1 yn −1
- Using GARCH
cov n = ω + α cov n −1 + β xn −1 yn −1
3. Consistency condition for covariances

- Variance-covariance matrix should be positive-semidefinite. That is, ω T Ωω ≥ 0

for any vector omega.
- To ensure that a positive-semidefinite matrix is produced, variances and
convariances should be calculated consistently. For example, if variance rates are
updated using an EWMA model with λ=0.94, the same should be done for
covariance rates.
4. Multivariate normal distribution

- E[Y|x] = μY + ρ(σY/σX)(x - μX), and

- The conditional mean of Y is linearly dependent on X and the conditional
standard deviation of Y is constant.
5. Factor models

U i = ai1 F1 + ai 2 F2 + ... + aiM FM + 1 − ai21 − ai22 − ... − aiM

2
Zi . The

factors have uncorrelated standard normal distributions and the Zi are uncorrelated
both with each other and with the factors. In this case the correlation between Ui and
M

Uj is ∑a
m =1
im a jm

-7-
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

6. Copulas
- The marginal distribution of X (unconditional distribution) is its distribution
assuming we know nothing about Y. But often there is no natural way of defining
a correlation structure between two marginal distributions.
- Gaussian copula approach. Suppose that F1 and F2 are the cumulative marginal
probability distributions of V1 and V2. We map V1 = v1 to U1 = u1 and V2 = v2
to U2 = u2, where
F1 (v1 ) = N (u1 ) and F2 (v2 ) = N (u2 ) , and N is the cumulative normal
distribution function.
- The key property of a copula model is that it preserves the marginal distributions
of V1 and V2 while defining correlation structure between them.
- Student t-copula Æ U1 and U2 are assumed to have a bivariate Student
t-distribution. Tail correlation is higher in a bivariate t-distribution than in a
bivariate normal distribution.
- A factor copula model: analysts often assume a factor model for the correlation
structure between the Ui.
7. Application to loan portfolios
- Define Ti (1<= i <= N) as the time when company i default and cumulative
probability distribution of Ti by Qi. To define the correlation structure between the
Ti using the one-factor Gaussian copula model, we map, for each i, Ti to a
variable Ui that has a standard normal on a percentile-to-percentile basis.
U − ai F
We assume: U i = ai F + 1 − ai Z i Æ Pr ob(U i < U | F ) = N [
2
- ]
1− a 2
i

- The mappings between the Ui and Ti imply Prob(Ui<U) = Prob(Ti<T) when

N −1[Qi (T )] − ai F
U = N −1[Q(T )] Æ Pr ob(Ti < T | F ) = N [ ]
1 − ai2
- Assuming Qi of time to default is the same for all i and equal to Q and the copula

correlation between any two names is the same and equals rho Æ ai = ρ for
all i. Æ
N −1[Q(T )] − ρ F
- Pr ob(Ti < T | F ) = N [ ] . For a large portfolio of loans, this
1− ρ
equation provides a good estimate of the proportion of loans in the portfolio that
default by time T. Æ We refer to this as the default rate.
- As F decreases, the default rate increases. Since F is standard normal, the
probability that F will be less than N-1(Y) is Y. There is therefore a probability of
N −1[Q(T )] − ρ N −1 (Y )
Y that the default rate will be greater than N [ ]
1− ρ
- Define V(T,X) as the default rate that will not be exceeded with probability X, so

-8-
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

that we are X% certain that the default rate will not exceed V(T,X). The value of
V(T,X) is determined by substituting Y = 1 – X into the above expression:
N −1[Q(T )] + ρ N −1 ( X )
V(T, X) = N [ ]
1− ρ

Chapter 7 Bank Regulation and Basel II

1. The capital a financial institution requires should cover the difference

between expected losses over some time horizon and ‘worst-case losses’
over the same time horizon. The idea is that expected losses are usually
covered by the way a financial institution prices its products. For example, the
interest charged by a bank is designed to recover expected loan losses. Capital
is a cushion to protect the bank from an extremely unfavorable outcome.
2. The 1988 BIS Accord (Basel I)
- Assets-to-capital <+20
- The Cooke Ratio Æ calculate risk-weighted assets.
3. Netting: If a counterparty defaults on one contract if has with a financial
institution then it must default on all outstanding contracts with that financial
institution.
Without netting the financial institution’s exposure in the event of a default today
N

is ∑ max(V , 0) (N
i =1
i contracts with the defaulted party). With netting, it is

N
max(∑ Vi , 0)
i =1

4. Basel II is based on three pillars:

- Minimum capital requirements
- Supervisory review: allow regulators in different countries some discretion in
how rules are applied but seeks to achieve overall consistency in the application of
the rules. It places more emphasis on early intervention when problems arise. Part
of their role is to encourage banks to develop and use better risk management
techniques and to evaluate these techniques.
- Market discipline: require banks to disclose more information about the way they
allocate capital and the risks they take.
5. Minimum capital requirements:
Total capital = 0.08*(credit risk RWA + Market risk RWA + Operational risk
RWA).

6. Market risk (1996 amendment to Basel I, continue to be used under Basel II):
- It requires financial institutions to hold capital to cover their exposures to market

-9-
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

risks as well as credit risks. It distinguishes between a bank’s trading book

(normally marked to market daily) and its banking book.
- The market risk capital requirement: k * VaR + SRC, where SRC is a specific
risk charge. The VaR is the greater of the pervious day’s VaR and the average VaR
over the last 60 days. The minimum value for k is 3.
- SRC is a capital charge for the idiosyncratic risks related to individual companies.
E.g. a corporate bond has two risks: interest rate risk and credit risk. The interest
rate risk is captured by the bank’s market VaR measure; the credit risk is specific
risk.

7. Credit risk capital (NEW FOR BASEL II)

- For an on-balance-sheet item a risk weight is applied to the principal to calculate
risk-weighted assets reflecting the creditworthiness of the counterparty. For
off-balance-sheet items the risk weight is applied to a credit equivalent amount.
This is calculated using either credit conversion factors or add-on amount.
- Standardized approach (for small banks. In USA, Basel II will apply only to the
largest banks and these banks must use the foundation internal ratings based (IRB)
approach). Æ risk weights for exposures to country, banks, and corporations as a
function of their ratings.
- IRB approach – one-factor Gaussian copula model of time to default.
WCDR: the worst-case default rate during the next year that we are 99.9% certain
will not be exceeded
PD: the probability of default for each loan in one year
EAD: The exposure at default on each loan (in dollars)
LGD: the loss given default. This is the proportion of the exposure that is lost in
the event of a default.
Suppose that the copula correlation between each pair of obligors is rho. Then
N −1[ PD] + ρ N −1 (0.999)
WCDR = N [ ]
1− ρ
It follows that there is a 99.9% chance that the loss on the portfolio will be less
than N times EAD*LGD*WCDR.
- For corporate, sovereign, and bank exposures, Basel II assumes the relationship

between rho and PD is: ρ ≈ 0.12(1 + e −50× PD ) . As PD increases, rho decreases.

Æ As a firm becomes less creditworthy, its PD increases and its probability of
default becomes more idiosyncratic and less affected by overall market
conditions.
- The formula for the capital required is: EAD*LGD*(WCDR – PD)*MA. We use
WCDR – PD instead of WCDR because we are interested in providing capital for
the excess of the 99% worst-case loss over the expected loss. The MA is the
maturity adjustment.
- The risk-weighted assets (RWA) are calculated as 12.5 times the capital required:
RWA = 12.5 * EAD*LGD*(WCDR – PD)*MA

- 10 -
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

8. Operational risk: Banks have to keep capital for operational risk. Three
approaches.
- Basic indicator approach: operational risk capital = the bank’s average annual
gross income (=net interest income + noninterest income) over the last three years
multiplied by 0.15.
- Standardized approach: similar to basic approach, except that a different factor is
applied to the gross income from different business lines.
- Advanced measurement approach: the bank uses its own internal models to
calculate the operational risk loss that it is 99.9% certain will not be exceeded in
one year.

9. The limitation of Basel II:

- The total required capital under Basel II is the sum of the capital for three different
risks (credit, market and operational). This implicitly assumes that the risks are
perfectly correlated. It will be desirable if banks can assume less than perfect
correlation between losses from different types of risk.
- Basel II does not allow a bank to use its own credit risk diversification
calculations when setting capital requirements for credit risk within the banking
book. The prescribed rho must be used. In theory a bank with $1 billion of lending
to BBB-rated firms in a single industry is liable to be asked to keep the same
capital as a bank that has $1 billion of lending to a much more diverse group of
BBB-rated corporations.

Chapter 8 The VaR Measure

1. Choice of parameters for VaR

- VaR = σ N ( X ) , Assuming normal distribution and the mean change in the

−1

portfolio value is zero, where X is the confidence level, sigma is the standard
deviation (in dollars) of the portfolio change over the time horizon.
- The time horizon: N-day VaR = 1-day VaR * square root N.
- autocorrelation Æ the above formula a rough one.
- The confidence level: if the daily portfolio changes are assumed to be normally
distributed with zero mean Æ convert a VaR to another VaR with different confidence
level.
N −1 ( X *)
VaR( X *) = VaR( X ) −1
N (X )

- 11 -
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

∂ (VaR)
2. Marginal VaR: , where xi is the ith component of a portfolio. For an
∂xi

investment portfolio, marginal VaR is closely related to the CAPM’s beta. If an

asset’s beta is high, its marginal VaR will tend to be high.
3. Incremental VaR is the incremental effect on VaR of a new trade or the
incremental effect of closing out an existing trade. It asks the question: “what is
the difference between VaR with and without the trade.” If a component is small
relative to the size of a portfolio, it may be reasonable to assume that the marginal
∂ (VaR )
VaR remains constant as xi is reduced to zero Æ xi
∂xi
4. Component VaR is the part of the VaR of the portfolio that can be attributed to this
component:
- The ith component VaR for a large portfolio should be approximately equal to the
incremental VaR.
- The sum of all the component VaRs should equal to the portfolio VaR.
N
∂ (VaR)
- VaR = ∑ xi
i =1 ∂xi
5. Back testing:
- The percentage of times the actual loss exceeds VaR.
Let p = 1-X, where x is confidence level. m = the number of times that the VaR
limits is exceeded, n the total number of days.
Two hypotheses:
a. The probability of an exception on any given day is p.
b. The probability of an exception on any given day is greater than p.
The probability (binominal distribution) of the VaR limit being exceeded on m or
n
n!
more days is: ∑ k !(n − k )! p
k =m
k
(1 − p ) n − k . We usually use confidence level as 5%.

If the probability of the VaR limit being exceeded on m or more days is less than 5%,
we reject the first hypothesis that the probability of an exception is p.
The above test is one-tailed test. Kupiec has proposed a two-tailed test
(Frequency-of-tail-losses or Kupiec test). If the probability of an exception under the
VaR model is p and m exceptions are observed in n trials, then
−2 ln[(1 − p) n − m p m ] + 2 ln[(1 − m / n)n − m (m / n) m ] should have a chi-square
distribution with one degree of freedom.
z The Kupiec test is a large sample test
z Kupiec test focuses solely on the frequency of tail losses. It throws away
potentially valuable information about the sizes of tail losses. This suggests that
the Kupiec test should be relatively inefficient, compared to a suitable test that
took account of the sizes as well as the frequency of tail losses.

- 12 -
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

- Sizes-of-tail-losses test: compare the distribution of empirical tail losses against

the tail-loss distribution predicted by model – Kolmogorov-Smirnov test (it is the
maximum value of the absolute difference between the two distribution
functions).
Another backtest is Crnkovic and Drachman (CD) test. The test is to evaluate a
market model by testing the difference between the empirical P/L distribution and
the predicted P/L distribution, across their whole range of values.
- The extent to which exceptions are bunched: in practice, exceptions are often
bunched together, suggesting that losses on successive days are not independent.
6. Stress testing: involves estimating how the portfolio would have performed under
extreme market moves.
Two Main approaches:
a. Scenario analyses, in which we evaluate the impact of specified scenarios
(e.g., such as a particular fall in the stock market) on our portfolio. The
emphasis is on specifying the scenario and working out its ramifications.
b. Mechanical stress tests, in which we evaluate a number (often a large number)
of mathematically or statistically defined possibilities (e.g., such as increases
or decreases of market risk factors by a certain number of standard deviations)
to determine the most damaging combination of events and the loss it would
produce.
7. stress testing – scenario analyses:
- Stylized scenarios: Some stress tests focus on particular market variables.
a. Shifting a yield curve by 100 basis points
b. Changing implied volatilities for an asset by 20% of current values.
c. Changing an equity index by 10%.
d. Changing an exchange rate for a major currency by 6% or changing the
exchange rate for a minor currency by 20%.
- Actual historical events: stress tests more often involve making changes to several
market variables – use historical scenarios. E.g. set the percentage changes in all
market variables equal to those on October 19, 1987.
- If movements in only a few variables are specified in a stress test, one approach is
to set changes in all other variables to zero. Another approach is to regress the
nonstressed variables on the variables that are being stressed to obtain forecasts
for them, conditional on the changes being made to the stressed variables
(conditional stress testing)
8. Stress testing – mechanical stress testing
- Factor push analysis: we push the price of each individual security or (preferably)
the relevant underlying risk factor in the most disadvantageous direction and work
out the combined effect of all such changes on the value of the portfolio.
a. We start by specify a level of confidence, which gives us a confidence level
parameter alpha.
b. We then consider each risk factor on its own, ‘push’ it by alpha times its
standard deviation, and revalue the portfolio at the new risk factor value;
c. We do the same for all risk factors, and select that set of risk factor

- 13 -
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

movements that has the worst effect on the portfolio value.

d. Collecting these worst price movements for each instrument in out portfolio
gives us our worst-case scenario, and the maximum loss (ML) is equal to the
current value of our portfolio minus the portfolio value under this worst-case
scenario.
e. Factor push test is only appropriate for certain relatively simple types of
portfolio in which the position value is a monotonic function of a risk factor.
- Maximum loss optimization: search over the losses that occur for intermediate as
well as extreme values of the risk variables

Chapter 9 Market Risk VaR: Historical Simulation

Approach

1. Methodology
Suppose that we calculate VaR for a portfolio using a one-day time horizon, a 99%
confidence level, and 500 days of data.
- Identify the market variables affecting the portfolio. They typically are exchange
rates, equity prices, interest rates, and so on.
- Collect data on the movements in these variables over the most recent 500 days.
vi
- The value of the market variable tomorrow under ith scenario is vn (n –
vi −1
today, say n=500, i = 0, 1, … 499)
- Get change in portfolio value under each scenario according to the calculated
market variables.
- VaR = 5th worst number of the change in portfolio value.
2. The confidence interval:
- Kendall and Stuart calculate a confidence interval for the quantile of a distribution
when it is estimated from sample data.
1 q (1 − q )
- The standard errors of the estimate is: , where f(x) is the
f ( x) n
probability density function of the loss evaluated at x.
3. Weighting of observations: the basic historical simulation approach assumes that
each day in the past is given equal weight. Boudoukh et alpha. Suggest that more
recent observations should be given more weight – the weight given to the change
λ i −1 (1 − λ )
between day n-i and n-i+1 is . VaR is calculated by ranking the
1− λ n
observations from the worst outcome to the best. Starting at the worst outcome,

- 14 -
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

weights are summed until the required quantile of the distribution is reached.
4. Incorporating volatility updating (Hull and White (1998)).
- Get daily volatility for each market variable.
- Suppose σ n +1 (current estimate of the volatility of the market variable between

today and tomorrow) is twice σ i . Æ we expect to see changes between today and

tomorrow that are twice as big as changes between day i-1 and day i. Æ the value
vi −1 + (vi − vi −1 )σ n +1 / σ i
of a market variable under the ith scenario becomes: vn
vi −1
5. Extreme value theory is a way of smoothing the tails of the probability
distribution of portfolio daily changes calculated using historical simulation. It
leads to estimates of VaR that reflect the whole shape of the tail of the distribution,
not just the positions of a few losses in the tails. It can also be used to estimate
VaR when the VaR confidence level is very high. E.g., even if we have only 500
days of data, it could be used to come up an estimate of VaR for a VaR confidence
level of 99.9%:
- F(x) is CDF, Fu(y) as the probability that x lies between u and u+y conditional on
F (u + y ) − F (u )
x>u. Æ Fu ( y ) =
1 − F (u )
- Gnedenko states that for a wide class of distributions F(x), the distribution of Fu(y)
converges to a generalized Pareto distribution as the threshold u is increased.
y
Gξ , β ( y ) = 1 − (1 + ξ )−1/ ξ , where ξ is the shape parameter and determines
β
the heaviness of the tail of the distribution. Beta is a scale parameter.
- Estimating ξ and beta – MLE. The PDF of G(y) is

1 y
gξ , β ( y ) = (1 + ξ ) −1/ ξ −1
β β
- First, choose a value of u, say the 95 percentile. Then focus attention on those
observations for which x > u. Suppose there are nu such observations and they are
xi (i<=nu). Æ
nu
1 xi − u
lng(x) = ∑ ln[ β (1 + ξ
i =1 β
)−1/ ξ −1 ]

- Prob(x>u+y| x>u) = 1 –G(y), prob(x>u) = 1 – F(u) ~ nu/n Æ prob(x>u+y) =

nu x − u −1/ ξ
[1-F(u)][1-G(y)] Æ F ( x) = 1 − (1 + ξ ) , this is the estimator of the
n β
tail of the CDF of x when x is large. This is reduced to the Power Law if we set

- 15 -
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

u = β /ξ .
- Calculation of VaR with a confidence level of q Æ F(VaR) = q Æ
−ξ
β ⎡⎛ n ⎞ ⎤
VaR = u + ⎢⎜ (1 − q) ⎟ − 1⎥
ξ ⎢⎣⎝ nu ⎠ ⎥⎦

Chapter 10 Market Risk VaR: Model-Building Approach

1. Basic methodology – two-asset case Æ diversification. Note: VaR does not always
reflect the benefits of diversification.
n

- The linear model: ΔP = ∑ α i Δxi , where dx is the return on asset i in one day.
i =1
Alphai is the dollar value being invested in asset i. Æ get mean and standard deviation
of dp then we are done under multivariate normal distribution.
2. Handling interest rates- cash flow mapping. The cash flows from instruments in
the portfolio are mapped into cash flows occurring on the standard maturity dates.
Since it is easy to calculate zero T-bills or T-bonds’ volatilities and correlations,
after mapping, it is easy to calculate the portfolio’s VaR in terms of cash flows of
zero T-bills.
3. Principal components analysis: A PCA can be used (in conjunction with cash flow
mapping) to handle interest rates when VaR is calculated using the model-building
approach.
4. Applications of Linear model:
- A portfolio with no derivatives.
- Forward contracts on foreign exchange (to be treated as a long position in the
foreign bond combined with a short position in the domestic bond).
- Interest rate swaps
5. The linear model and options:
n n
ΔP = δΔS = SδΔx = ∑ Siδ i Δxi = ∑ α i Δxi , where dx = ds/s is the return on a
i =1 i =1

stock in one day. Ds is the dollar change in the stock price in one day.
- The weakness of the model: when gamma is positive (negative), the pdf of the value
of the portfolio tends to be positively (negatively) skewed Æ have a less heavy
(heavier) left tail than the normal distribution. Æ if we assume the distribution is
normal, we will tend to calculate a VaR that is too high (low).
6. The Quadratic model:

- 16 -
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

1 1
- ΔP = δΔS + γ (ΔS ) 2 = SδΔx + S 2γ (Δx) 2 (ignoring theta term).
2 2
- For a portfolio with n underlying market variables, with each instrument in the
portfolio being dependent on only one of the market variables, Æ
n
1 n 2
ΔP = ∑ Siδ i Δxi + ∑ Si γ i (Δxi ) 2 , where Si is the value of the ith market
i =1 2 i =1
variable.
- When some of the individual instruments are dependent on more than one market
variable, this equation takes the more general form:
n
1 n n
ΔP = ∑ Siδ i Δxi + ∑∑ Si S j γ ij Δxi Δx j , where gammaij is a ‘cross gamma’.
i =1 2 i =1 j =1
- Cornish – Fisher expansion:Æ estimate percentiles of a probability distribution
from its moments that can take account of the skewness of the probability
distribution.
Using the first three moments of dP, the Cornish-Fisher expansion estimates the

q-percentile of the distribution of dP as μ P + ωqσ P , where

1
ωq = zq + ( zq2 − 1)ξ P and Z is q-percentile of the standard normal distribution
6
and ξ P is the skewness of dp.

7. The model-building approach is frequently used for investment portfolios. It is

less popular for the trading portfolios of financial institutions because it does not
work well when deltas are low and portfolios are nonlinear.

Chapter 11 Credit Risk: Estimating Default Probabilities

1. Credit-risk arises from the possibility that borrowers, bond issuers, and
counter-parties in derivatives transactions may default. In theory, a credit rating is
an attribute of a bond issue, not a company. However, in most cases all bonds
issued by a company have the same rating. A rating is therefore often referred to
as an attribute of a company.
- Ratings changes relatively infrequently. One of rating agencies’ objectives is
ratings stability. They want to avoid ratings reversals where a firm is downgraded
and then upgraded a few weeks later. Ratings therefore change only when there is
reason to believe that a long-term change in the firm’s creditworthiness has taken
place. The reason for this is that bond traders are major users of ratings. Often
they are subject to rules governing what the credit ratings of the bonds they hold

- 17 -
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

must be. If these ratings changed frequently they might have to do a large amount
of trading just to satisfy the rules.
- Rating agencies try to ‘rate through the cycle’. Suppose that an economic
downturn increases the probability of a firm defaulting in the next 6 months, but
makes very little difference to the firm’s cumulative probability of defaulting over
the next three to five years. A rating agency would not change the firm’s rating.
2. Internal credit ratings: most banks have procedures for rating the creditworthiness
of their corporate and retail clients. The internal ratings based (IRB) approach in
Basel II allows bank to use their internal ratings in determining the probability of
default, PD. Under the advanced IRB approach, they are also allowed to estimate
the loss given default, LGD, the exposure at default, EAD, and the maturity, M.
3. Altman’s Z-score
Using discriminant analysis, Altman attempted to predict defaults from five
accounting ratios.
Z = 1.2*working capital + 1.4*Retained earnings + 3.3* EBIT + 0.6*Market value of
equity + 0.999*sales, all variables are scaled by assets, except for market equity,
which is scaled by book value of total liabilities.
If Z-score > 3, the firm is unlikely to default. If it is between 2.7 and 3.0, we should
be ‘on alert’. If it is between 1.8 and 2.7, there is a good chance of default. If it is less
than 1.8, the probability of a financial embarrassment is very high.
4. Historical default probabilities
- For investment-grade bonds, the probability of default in a year tends to be an
increasing function of time.
- For bonds with a poor credit rating, the probability of default is often a decreasing
function of time. The reason is that, for a bond like this, the next year or two may
be critical. If the issuer survives this period, its financial health is likely to have
improved.
- Default intensities (hazard rate): λ(t) at t is defined so thatλ(t)Δt is the
probability of default between time t and t +Δt conditional on no default between
time 0 and time t.
Let V(t) is the cumulative probability of the firm surviving to time t (no default by
time t), then V(t +Δt) – V(t) = -λ(t) V(t)Δt Æ
dV (t )
t

= −λ (t )V (t ) → V(t)=e ∫0
- λ (τ ) dτ
= e − λt , where λ is the
dt
average default intensity between time zero and t.
Define Q(t) as the probability of default by time t Æ
t

e ∫0
- λ (τ ) dτ 1
Q(t) = 1 – V(t) = 1 - = 1- e − λt Æ λ (t ) = − ln[1 − Q (t )] , where Q(t)
t
comes from historical data.
5. Recovery rates: the bond’s market value immediately after a default as a percent
of its face value = 1 – LGD.
- Recovery rates are significantly negatively correlated with default rates.

- 18 -
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

- Average recovery rate = 0.52 – 6.9*average default rate. Æ a bad year for the
default rate is usually doubly bad because it is accompanied by a low recovery
rate.
6. Estimating default probabilities from bond prices
− RfT
- An approximate calculation: [1 − PD ]100e + PD100 Re − RfT = 100e − yT

1 − e −[ y − Rf ]T
Æ PD(T) = ~ spread/(1-R)
1− R
- Risk-free rate: Traders usually use LIBOR/swap rates as proxies for risk-free
rates when calculating default probabilities.
- Credit default swaps can be used to imply the risk-free rate assumed by
traders. The rate used appears to be approximately equal to the LIBOR/swap rate
minus 10 basis points on average. (Credit risk in a swap rate is the credit risk from
making a series of 6-month loans to AA-rated counterparties and 10 basis points is
a reasonable default risk premium for an AA-rated 6-month instrument.
7. Asset swap: traders often use asset swap spreads as a way of extracting default
probabilities from bond prices. This is because asset swap spreads provide a direct
estimate of the spread of bond yields over the LIBOR/swap curve. Æ The present
value of the asset swap spread is the amount by which the price of the corporate
bond is exceeded by the price of a similar risk-free bond.
8. Real-world vs. Risk-neutral probabilities
- The default probabilities implied from bond yields are risk-neutral default
probabilities.
- The default probabilities implied from historical data are real-world default
probabilities. If there was no expected excess return, the real-world and
risk-neutral default probabilities would be the same, and vice versa.
- Reasons for the difference
a. Corporate bonds are relatively illiquid and bond traders demand an extra
return to compensate for this.
b. The subjective default probabilities of traders may be much higher than those
given in table 11.1. Traders may be allowing for depression scenarios much
worse than anything seen during the 1970 to 2003 period.
c. Bonds do not default independently of each other. This gives rise to
systematic risk and traders should require an expected excess return for
bearing the risk. The variation in default rates from year to year may be
because of overall economic conditions or because a default by one company
has a ripple effect resulting in defaults by other companies (the latter is
referred as credit contagion).
d. Bond returns are highly skewed with limited upside. As a result it is much
more difficult to diversify risks in a bond portfolio than in an equity portfolio.
- When valuing credit derivatives or estimating the impact of default risk on the
pricing of instruments, we should use risk-neutral default probabilities.
- When carrying out scenario analyses to calculate potential future losses from

- 19 -
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

defaults, we should use real-world default probabilities. The PD used to calculate

regulatory capital is a real-world default probabilities.
9. Estimating default probabilities from equity prices:
- Equity is an option on the assets of the company. Æ the value of equity today is:

- E0 = V0 N ( d1 ) − De − rT N ( d 2 ) , where
ln(V0 / D) + ( r + σ V2 / 2)T
- d1 = and d 2 = d1 − σ V T
σV T
- The risk-neutral probability that a firm will default on the debt is N(-d2). We need

V0 and
σ V , which are not directly observable.

∂E
- From Ito’s lemma Æ σ E E0 = σ V V0 = N (d1 )σ V V0
∂V

- The above two equations Æ the values of V0 and

σ V Æ N(-d ).
2

10. The default probability can be estimated from historical data (real-world
probabilities), bond prices (risk-neutral probabilities), or equity prices (in
theory risk-neutral probabilities. But, the output from the model can be calibrated
so that either risk-neutral or real-world default probabilities are produced).

Chapter 12 Credit Risk Losses and Credit VaR

1. Estimating credit losses:

- Credit losses on a loan depend primarily on the probability of default and the
recovery rate.
- The credit risk in a derivative transaction is more complicated because the
exposure at the time of default is uncertain. Some derivatives transactions (e.g.,
written options) are always liabilities and give rise to no credit risk. Some (e.g.,
long positions in options) are always assets and entail significant credit risks.
Some may become either assets or liabilities during their life (e.g., forward and
swaps).
2. Adjusting derivatives valuations for counterparty default risk
- Assume the value of the derivative to the financial institution at time ti is fi,
i=0,1,…n, and qi is the risk-neutral probability of default at time ti, the expected
recovery rate is R. Æ the risk-neutral expected loss from default at time ti is:
l [ max( f , 0) ] , taking
qi (1 − R ) E present values and sum them up over iÆ
i

costs of defaults. ∑ q (1 − R)v

i =1
i i , where v is the value today of an instrument that

pays off the exposure on the derivative under consideration at time ti.

- 20 -
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

- The impact of default risk on interest rate swaps is considerably less than that on
currency swaps, largely because principals are exchanged at the end of the life of a
currency swap and there is uncertainty about the exchange rate at that time.
- Two sided default risk (when contracts can become either assets or liabilities).
Human nature being what it is, most firms consider that there is no chance that
they themselves will default but want to make an adjustment to contract terms for
a possible default by their counterparty. This can make it very difficult for the two
firms to agree on terms and explains why it is difficult for financial institutions
that are not highly creditworthy to be active in the derivatives market.
3. Credit risk mitigation
- Netting. Because of netting, the incremental effect of a new contract on expected
default losses can be negative. This tends to happen when the value of the new
contract is highly negatively correlated with the value of existing contracts. A firm
may get different quotes in a well-functioning capital market. The company is
likely to get the most favorable quote from a financial institution it has done
business with in the past – particularly if that business gives rise to exposures for
the financial institution that are opposite to the exposure generated by the new
transaction.
- Collateralization
- Downgrade triggers. This is a clause stating that if the credit rating of the
counterparty falls below a certain level, say Baa, then the financial institution has
the option to close out a derivatives contract at its market value.
4. Credit VaR. Whereas the time horizon for market risk is usually between one day
and one month that for credit risk is usually much longer – often one year.
5. Vasicek’s modela:
- We are X% certain that the default rate will not exceed V(T,X).
N −1[Q(T )] + ρ N −1 ( X )
V(T, X) = N [ ] , where Q(T) is the cumulative
1− ρ
probability of each loan defaulting by time T.
6. Credit risk plus: the probability of n defaults follows the Poisson distribution and
this is combined with a probability distribution for the losses experienced on a
single counterparty default to obtain a probability distribution for the total default
losses from the counterparties.
7. CreditMetrics. It is based on an analysis of credit migration. This is the
probability of a firm moving from one rating category to another during a certain
period of time. Calculating a one-year VaR for the portfolio using CreditMetrics
involves carrying out Monte Carlo simulation of ratings transitions for bonds in
the portfolio over a one-year period. On each simulation trial the final credit rating
of all bonds is calculated and the bonds are revalued to determine total credit
losses for the year. The 99% worst result is the one-year 99% VaR for the
portfolio.
- The credit rating changes for different counterparties should not be assumed to be
independent. Æ use Gaussian copula model Æ The copula correlation between the

- 21 -
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

rating transitions for two companies is typically set equal to the correlation
between their equity returns using a factor model.

Chapter 13 Credit Derivatives

1. Credit derivatives are contracts where the payoff depends on the

creditworthiness of one or more companies or countries. Banks have been the
largest buyers of CDS credit protection and insurance companies have been
the largest sellers.
- The n-year CDS spread should be approximately equal to the excess of the par
yield on an n-year corporate bond over the par yield on an n-year risk-free bond. If
it is markedly less than this, an investor can earn more than the risk-free rate by
buying the corporate bond and buying protection. If it is markedly great than this,
an investor can borrow at less than the risk-free rate by shorting the corporate
bond and selling CDS protection.
- The payoff from a CDS in a credit event is notional principal*(1-recovery rate).
Usually a CDS specifies that a number of different bonds can be delivered in the
credit event. This gives the holder of CDS a cheapest-to-deliver bond option.
Therefore, recovery rate should be the lowest recovery rate applicable to a
deliverable bond.
2. Valuation of credit default swaps
- CDS can be analyzed by calculating the present value of the expected payments
(including accrued interests) and the present value of the expected payoff.
- The default probabilities used to value a CDS should be risk-neutral default
probabilities, which can be estimated from bond prices or asset swaps. An
alternative is to imply them from CDS quotes.
- Binary CDS: it is similar to a regular CDS except that the payoff is a fixed dollar
amount.
- Is the recovery rate important? – whether we use CDS spreads or bond prices to
estimate default probabilities, we need an estimate of the recovery rate. However,
provided that we use the same recovery rate for (a) estimating risk-neutral default
probabilities and (b) valuing a CDS, the value of the CDS is not very sensitive to
the recovery rate. This is because the implied probabilities of default are
approximately proportional to 1/(1-R) and the payoffs from a CDS are
proportional to 1-R.
3. A total return swaps is an agreement to exchange the total return on a bond (or
any portfolio of assets) for LIBOR plus a spread. The total return includes
coupons, interest, and the gain or loss on the asset over the life of the swap.
The spread over LIBOR received by the payer is compensation for bearing
the risk that the receiver will default. The payer will lose money if the
receiver defaults at a time when the reference bond’s price has decline. The
spread therefore depends on the credit quality of the receiver and of the bond

- 22 -
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

issuer, and on the default correlation between the two.

4. Basket CDS
- Add-up CDS provides a payoff when any of the reference entities default.
- An nth-to-default CDS provides a payoff only when the nth default occurs.
5. CDOs
- Cash CDO (based on bonds)
- Synthetic CDO: the creator sells a portfolio of CDSs to third parties. It then passes
the default risk on to the synthetic CDO’s tranche holders. The first tranche may
be responsible for the payoffs on the CDS until they have reached 5% of the total
notional principal;…; The income from the CDS is distributed to the tranches in a
way that reflects the risk they are bearing.
6. Valuation of a basket CDS and CDO
- The spread for an nth-to-default CDS and the tranche of a CDO is critically
dependent on default correlation. As the default correlation increases the
probability of one or more defaults declines and the probability of ten or more
defaults increases.
- The one-factor Gaussian copula model of time to default has become the standard
market model for valuing an nth-to-default CDS or a tranche of a CDO.
- Consider a portfolio of N firms, each having a probability Q(T) of defaulting by
time T. Æ the probability of default, conditional on the level of the factor F, is

N −1[Q(T )] − ρ F
Q(T | F ) = N [ ]
1− ρ
- The trick here is to calculate expected cash flows conditional on F and then
integrate over F. The advantage of this is that, conditional on F, defaults are
independent. The probability of exactly k defaults by time T, conditional on F, is
n!
Q(T | F ) k (1 − Q(T | F )) n − k
k !( n − k )!
- Derivates dealers calculate the implied copula correlation rho from the spreads
quoted in the market for tranches of CDOs and tend to quote these rather than
the spreads themselves. This is similar to the practice in options markets of
quoting B-S implied volatilities rather than dollar prices.
- Correlation smiles: the compound correlation is the correlation that prices a
particular tranche correctly. The base correlation is the correlation that prices all
tranches up to a certain level of seniority correctly. In practice, we find that
compound correlations exhibit a ‘smile’ with the correlations for the most junior
(equity) and senior tranches higher than those for intermediate tranches. The base
correlations exhibit a ‘skew’ where the correlation increases with the level of
seniority considered.

- 23 -
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

Chapter 14 Operational Risk

1. Operational risk: the risk of loss resulting from inadequate or failed internal
processes, people and systems or from external events.
- It includes legal risk, but does not include reputation risk or the risk resulting from
strategic decisions.
2. 7 Categories of operational risk
- Internal fraud (Barings),
- External fraud (Republic NY corp. Lost $611 million because of fraud committed
by a custodial client,
- Employment practices and workplace safety (Merrill Lynch lost $250 million in a
gender discrimination lawsuit),
- Clients, products, & business practices (Household International lost $484 million
from improper lending practices),
- Damage to physical assets (911 attacks),
- Business disruption and system failures (Solomon Brothers lost $303 million from
a change in computing technology).
- Execution, delivery, and process management: failed transaction processing or
process management, and relations with trade counter-parties and vendors. E.g.,
Bank of America and Wells Fargo Bank lost $225 and $150 million, respectively,
from systems integration failures and transactions processing failures.
3. Loss severity and loss frequency
- Loss frequency distribution – the distribution of the number of losses observed
during the time horizon (a Poisson distribution is usually used).
- Loss severity distribution – the distribution of the size of a loss, given that a loss
occurs. (usually assume the two are independent). (a lognormal probability
distribution is often used)
- The two distributions must be used for each loss type and business line to
determine a total loss distribution. Monte Carlo simulation can be used here:
a. Sample from the frequency distribution to determine the number of loss
events (=n).
b. Sample n times from the loss severity distribution to determine the loss
experienced for each loss event (L1, L2, … Ln)
c. Determine the total loss experienced ( = L1 + L2 +…+ Ln).
d. Repeat this many times.
- Data: the frequency distribution should be estimated from the bank’s own data as
far as possible. For the loss severity distribution, regulators encourage banks to
use their own data in conjunction with external data (through sharing
arrangements between banks or from data vendors. Both internal and external data
must be adjusted for inflation. A scale adjustment should be made to external
data).
4. Operational risk capital should be allocated to business units in a way that
encourages them to improve their operational risk management. The overall

- 24 -
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

result of operational risk assessment and operational risk capital allocation

should be that business units become more sensitive to the need for managing
operational risk.
5. The power law holds well for the large losses experienced by banks. This
makes the calculation of VaR with high degree of confidence such as 99.9%
possible. When loss distributions are aggregated, the distribution with the
heaviest tails tends to dominate Æ it may only be necessary to consider one
or two business-line/loss-type combinations.
6. Insurance – two problems (moral hazard [deductibles, coinsurance provisions,
and policy limits can be employed] and adverse selection). Some insurance
companies do offer rogue trader insurance policies. These companies tend to
specify carefully how trading limits are implemented. They may also require
that the existence of the insurance policy not be revealed to anyone on the
trading floor. They are likely to want to retain the right to investigate the
circumstances underlying any loss.
7. Sarbanes-Oxley Æ It requires boards of directors to become much more
involved with day-to-day operations. They must monitor internal controls to
ensure risks are being assessed and handled well.
8. Two sorts of operational risk: high-frequency low-severity risks and
low-frequency high-severity risks. The former are relatively easy to quantify,
but operational risk VaR is largely driven by the latter.

Chapter 15 Model Risk and Liquidity Risk

1. Model risk is the risk related to the models a financial institution uses to value
derivatives. Liquidity risk is the risk that there may not be enough buyers/sellers
in the market for a financial institution to execute the trades it desires. The two
risks are related. Sophisticated models are only necessary to price products that
are relatively illiquid. When there is an active market for a product, prices can be
observed in the market and models play a less important role.
2. Exploiting the weakness of a competitor’s model: a LIBOR-in-arrears swap is an
interest rate swap where the floating interest rate is paid on the day it is observed,
not one accrual period later. Whereas a plain vanilla swap is correctly valued by
assuming that future rates will be today’s forward rates, a LIBOR-in-arrears swap
should be valued on the assumption that the future rate is today’s forward interest
rate plus a ‘convexity adjustment’.
3. Models for actively traded products.
- Traders frequently quote implied volatilities rather than the dollar prices of
options, because implied volatility is more stable than the price.
- Volatility smiles: the volatility implied by B-S as a function of the strike price for
a particular option maturity is known as a volatility smile. The relationship
between strike price and implied volatility should be exactly the same for calls

- 25 -
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

and puts in the case of European options and approximately the same in the case
of American options.
- Volatility smile for foreign currency options. The reason for the volatility smile is
that B-S assumes: a. the volatility of the asset is constant; b. the price of the asset
changes smoothly with no jumps. Neither of these conditions is satisfied for an
exchange rate.
- Volatility skew – equity options. The volatility decreases as the strike price
increases. The volatility used to price a low-strike-price option (i.e., a
deep-out-of-the-money put or a deep in the money call) is significantly higher
than that used to price a high strike-price option (i.e., a deep in the money put or a
deep out of the money call). One explanation for the skew in equity options
concerns leverage. As a firm’s equity declines in value, Æ leverage increase Æ
equity becomes more risky and its volatility increases.
Crashophobia: the volatility smile for equities has existed only since the stock
market crash of Oct 1987. Prior to Oct 1987 implied volatilities were much less
dependent on strike price. Æ Mark Rubinstein suggested that one reason for the
equity volatility smile may be ‘crashophobia’. Traders are concerned about the
possibility of another crash similar to Oct 1987 and assign relatively high prices (and
therefore relatively high implied volatilities) for deep out of the money puts.
- Volatility surfaces Æ implied volatility as a function of both strike price and time
to maturity Æ volatility smile becomes less pronounced as the time to maturity
increases. To value a new option, traders look up the appropriate volatility in the
volatility surface table using interpolation.
- Models play a relatively minor role in the pricing of actively traded products.
Models are used in a more significant way when it comes to hedging.
a. Within-model hedging is designed to deal with the risk of changes in variables
that are assumed to be uncertain by the model.
b. Outside-model hedging deals with the risk of changes in variables that are
assumed to be constant/deterministic by the model. When B-S is used, hedging
against movements in the underlying stock price (delta and gamma hedging) is
with-model hedging because the model assumes that stock price changes are
uncertain. However, hedging against volatility (vega hedging) is outside-model
hedging.
4. Models for structured products: exotic options and other nonstandard products that
are tailored to the needs of clients are referred to as structured products. Usually
they do not traded actively and a financial institution must rely on a model to
determine the price it charges the client. A financial institution should, whenever
possible, use several different models. Æ get a price range and a better
understanding of the model risks being taken.
5. Detecting model problems:
- The risk management function should keep track of the following:
a. The type of trading the financial institution is doing with other financial
institutions.
b. How competitive it is in bidding for different types of structured transactions.

- 26 -
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

c. The profits being recorded from the trading of different products.

- Getting too much of a certain type of business or making huge profits from
relatively simple trading strategies, or the financial institution is unable to unwind
trades at close to the prices given by its models, can be a warning sign.
6. Liquidity risk: as the quantity increases, the price paid by the buyer increases and
the price received by the sell decreases.
- The percentage bid-offer spread: s = (offer price – bid price)/mid-price
n

- Liquidity-adjusted VaR = VaR + ∑sα

i =1
i i / 2 , where αi is the amount of money

invested in the ith position. As the number of position, n, grows, VaR benefits from
diversification but the liquidity adjustment does not. Consequently, the percentage difference
between VaR and liquidity-adjusted VaR grows as n grows.
7. Liquidity black holes is created when a price decline causes more market
participants to want to sell, driving prices well below where they will eventually
settle. During the sell-off, liquidity dries up and the asset can be sold only at a
fire-sale price. Reasons for herd behavior and the creation of liquidity black holes
are:
- The computer models used by different traders are similar.
- All financial institutions are regulated in the same way and respond in the same
way to changes in volatilities and correlations. E.g., when volatilities and
correlations increase, market VaR and the capital required for market risks
increase Æ banks tend to reduce their exposures. Since banks often have similar
positions to each other, they try to do similar trades. Æ liquidity black hole. E.g.,
consider credit risk. During economic recessions, default probabilities are
relatively high and capital requirements for loans under Basel II internal ratings
based models tend to be high Æ banks may be less willing to make loans, creating
a liquidity black hole for small and medium-sized businesses. Æ to solve this,
Basel Committee requires that the probability of default should be an average of
the probability of default through the economic or credit cycle, rather than an
estimate applicable to one particular point in time.
- There is a natural tendency to feel that if other people are doing a certain type of
trade then they must know something that you do not.

Chapter 17 Weather, Energy, and Insurance Derivatives

1. Weather derivatives (US department of Energy estimated that 1/7 of the US

economy is subject to weather risk): we will see contracts on rainfall, snow and
similar variables become more commonplace.
- HDD: heating degree days – a measure of the volume of energy required for
heating during the day.

- 27 -
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

HDD = max (0, 65 –A), where A is the average of the highest and lowest
temperature during the day at a specified weather station, measured in degrees
Fahrenheit.
- CDD: cooling degree days – a measure of the volume of energy required for
cooling during the day.
CDD = max (0, A – 65)
- A typical over-the-counter product is a forward or option contract providing a
payoff dependent on the cumulative HDD or CDD during a month.
2. Energy Derivatives
- Oil: virtually any derivative that is available on common stocks is now available
with oil as the underlying asset.
- Natural gas and electricity.
- An energy producer faces two risks: price risk and volume risk. Although prices
do adjust to reflect volumes, there is a less-than-perfect relationship between the
two. The price risk can be hedged using the energy derivative contracts. The
volume risks can be hedged using the weather derivatives.
3. Insurance Derivatives – are beginning to become an alternative to traditional
reinsurance as a way for insurance companies to manage the risks of
catastrophic events: CAT (catastrophic) bond: a bond issued by a subsidiary of
an insurance firm that pays a higher-than-normal interest rate. In exchange for
the extra interest, the holder of the bond agrees to provide an excess-of-cost
reinsurance contract. Depending on the terms of the CAT bond, the interest or
principal (or both) can be used to meet claims. There is no statistically
significant correlation between CAT risks and market returns. CAT bonds are
therefore an attractive addition to an investor’s portfolio.

Chapter 18 Big Losses and What We Can Learn From Them

1. Risk Limits: it is essential that all companies define in a clear and unambiguous
way limits to the financial risks that can be taken. Ideally, overall risk limits
should be set at board level. These should then be converted to limits applicable to
the individuals responsible for managing particular risks.
- It is tempting to ignore violations of risk limits when profits result. The penalties
for exceeding risk limits should be just as great when profits result as when losses
result. Otherwise, traders that make losses are liable to keep increasing their bets
in the hope that eventually a profit will result and all will be forgiven.
- Do not underestimate the benefits of diversification
- Carry out scenario analyses and stress tests: The calculation of risk measures such
as VaR should always be accompanied by scenario analyses and stress testing to
obtain an understanding of what can go wrong.
2. Managing the trading room
- There is a tendency to regard high-performing traders as ‘untouchable’ and to not
subject their activities to the same scrutiny as other traders. It is important that all

- 28 -
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

traders – particularly those making high profits – be fully accountable.

- Separate the front (consists of the traders who are executing trades, taking
positions, etc.), middle (consists of risk managers who are monitoring the risks
being taken) and back office (record-keeping and accounting)
- Do not blindly trust models: if large profits are reported when relatively simple
trading strategies are followed, there is a good chance that the models are wrong.
Similarly, if a financial institution appears to be particularly competitive on its
quotes for a particular type of deal, there is a good chance that it is using a
different model from other market participants. Getting too much business of a
certain type can be just as worrisome as getting too little business of that type.
- Be conservative in recognizing inception profits: it is much better to recognize
inception profits slowly so that traders are motivated to investigate the impact of
several different models and several different sets of assumptions before
committing themselves to a deal.
3. Liquidity risk
- Financial engineers usually base the pricing of exotic instruments and other
instruments that trade relatively infrequently on the prices of actively traded
instruments.
a. Often calculate a zero curve from actively traded government bonds and uses
it to price bonds that trade less frequently (off-the-run bonds).
b. Often implies the volatility of an asset from actively traded options and uses it
to price less actively traded options.
c. Often implies information about the behavior of interest rates from actively
traded interest rate caps and swap options and uses it to price products that are
highly structured.
These practices are not unreasonable. However, it is dangerous to assume that
less actively traded instruments can always be traded at close to their theoretical
price.
- Beware when everyone is following the same trading strategy: In the months
leading up to the crash of October 1987, many portfolio managers were
attempting to insure their portfolios by creating synthetic put options. They bought
stocks or stock index futures after a rise in the market and sold them after a fall.
This created an unstable market.
4. lessons for nonfinancial corporation:
- Make sure you fully understand the trades you are doing. If a trade and the
rationale for entering into it are so complicated that they cannot be understood by
the manager, it is almost certainly inappropriate for the corporation.
- Make sure a hedger does not become a speculator.
- Be cautious about making the treasury department a profit center
5. INTERNAL CONTROLS!!!

- 29 -
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

T1 Bootstrap

1. The procedure:
- We start with a sample of n.
- Draw a new sample of the same size from the original sample.
- Obtain a new parameter estimate from the new sample.
- Repeat the process B times and obtain a bootstrapped sample of B parameter
estimates;
- We use this sample to estimate a confidence interval.
2. Dealing with data dependency: the main limitation of the bootstrap is that standard
boostrap procedures presuppose that observations are independent, and they can
be unreliable if this assumption does not hold.
- If we are prepared to make parametric assumptions, we can model the dependence
parametrically (e.g., using a GARCH procedure). We can then bootstrap from the
residuals. However, the drawback of this solution is that is requires us to make
parametric assumptions and of course presupposes that those assumptions are
valid.
- An alternative is to use a block approach: we divide sample data into
non-overlapping blocks of equal length, and select a block at random. However,
this approach can be tricky to implement and can lead to problems because it
tends to ‘whiten’ the data.
- A third solution is to modify the probabilities with which individual observations
are chosen. Instead of assuming that each observation is chosen with the same
probability, we can make the probabilities of selection dependent on the time
indices of recently selected observations.

T2 Principal Component Analysis

1. The theory:
- Assume x is an mx1 random vector, with covariance matrix Σ, and let Λ be a
diagonal matrix whose diagonal elements are eigenvalues of Σ, let A is the

matrix of eigenvectors of Σ. Then Σ = AT ΛA

- The principal component of x are the linear combinations of the individual
x-variables produced by pre-multiplying x by A: p= Ax Æ the
variance-covariance matrix of p, VC(p), is then:

VC ( p ) = VC ( Ax) = AΣAT = Λ
- since Λ is a diagonal matrix. Æ the different principal components are
uncorrelated with each other. And the variances of principal components are given

- 30 -
Study Notes: Risk Management and Financial Institutions By Zhipeng Yan

by the diagonal elements ofΛ, the eigenvalues.

- In addition, we can choose the order of our principal components so that the
eigenvalues are in declining order. The first principal component therefore
‘explains’ more of the variability of our original data than the second principal
component, and so on.
- In short, the principal components of our m original variables are m artificial
variables constructed so that the first principal component ‘explains’ as much as it
can of the variance of these variables; the second principal component ‘explains’
as much as it can of the remaining variance, but it uncorrelated with the first
component; and so forth.
2. Estimates of principal components based on historical data can be quite unstable.
Some simple rules of thumb – such as taking moving averages of our principal
components – can help to mitigate this instability. We also should be careful about
using too many principal components. We only want to add principal components
that represent stable relationships that are good for forecasting, and there will
often come a point where additional principal components merely lead to the
model tracking noise – and so undermine the forecasting ability of our model.

T3 Monte Carlo Simulation Methods

1. The steps
- Select a model for the stochastic variables of interest.
- Estimate its parameters –volatilities, correlations, and so on – on the basis of
whatever historical or market data are available.
- Construct simulated paths for the stochastic variables using ‘random’ numbers.
Each set of ‘random’ numbers then produces a set of hypothetical terminal price(s)
for the instrument(s) in our portfolio.
- Repeat these simulations enough times.
2. Monte Carlo Simulation with single risk factors
- if stick price S follows a geometric Brownian motion process then, lnS:

d ln S = ( μ − σ 2 / 2)dt + σ dz Æ

S (Δt ) = S (0) exp ⎡⎣ ( μ − σ 2 / 2)Δt + σφ (Δt ) Δt ⎤⎦

S (T ) = S (0) exp ⎡⎣( μ − σ 2 / 2)T + σφ (T ) T ⎤⎦ Æ provided we are only

interested in the terminal stock value, this approach is both more accurate and less
time-consuming than the Euler method (get stock price for each incremental dt).

- 31 -