4.

Volatility, Correlations, and Copulas

FIN 500Q: Quantitative Risk Management

Definition of Volatility

• The volatility σ of a financial variable is the standard

deviation of its return per unit of time.

• When volatility is used for risk management, the unit of

time is often one day.

• Suppose that Si is the value of a variable at the end of day i.

The volatility per day is the standard deviation of
ln(Si /Si−1 ).

• Note that we continue to work with continuously

compounded returns.

1
Definition of Volatility

• The variance rate is the square of volatility.

• Assuming uncorrelated returns, variances linearly increase

with the length of the time period.

• Research has shown that volatility is much higher on

business days than on non-business days. Therefore, days
when markets are closed are usually ignored in volatility
calculations.
√
• This implies that the volatility per year is 252 times the
daily volatility.

2
Implied Volatilities

• Besides estimating volatilities based on historical data, risk

managers also frequently keep track of implied volatilities.

• The idea is that if we use a no-arbitrage option pricing

formula such as the Black-Scholes-Merton option pricing
model, the only parameter that cannot be directly observed
is volatility.

• We can therefore compute implied volatilities from market

prices and vice versa.

3
VIX Index

• VIX index: a measure of the implied volatility of the S&P

500 (“fear index”)

4
Are Daily Changes in Financial Variables Normally Distributed?

Test of normality using daily movements in 10 different

exchange rates:

Real World (%) Normal Model (%)

> 1 SD 23.32 31.73
> 2 SD 4.67 4.55
> 3 SD 1.30 0.27
> 4 SD 0.49 0.01
> 5 SD 0.24 0.00
> 6 SD 0.13 0.00

5
Heavy Tails

• The previous table suggests that daily exchange rate

changes are not normally distributed.
• The distribution has heavier tails than the normal
distribution.
• It is more peaked than the normal distribution.

• This means that small changes and large changes are more
likely than the normal distribution would suggest.

• Many market variables have this property, known as excess

kurtosis.

6
Normal and Heavy-Tailed Distribution

7
Alternatives to Normal Distributions: The Power Law

• A random variable ν satisfies a Power Law if its cumulative

distribution function is given by:

P(ν > x) = Kx −α .

• This distribution seems to fit the behavior of the returns on

many market variables better than the normal distribution.

• Also works well for modeling variables as diverse as the

income of an individual, the size of a city, and the number
of visits to a website in a day.

8
Log-Log Test of Power Law for Exchange Rate Data

9
Standard Approach to Estimating Volatility

• Define σn as the daily volatility of a market variable on day

n, as estimated at end of day n − 1.

• Define Si as the value of market variable at end of day i,

and ui = ln(Si /Si−1 ).

• We can estimate the daily variance using the most recent m

observations as
m
1 X
σn2 = (un−i − ū)2
m−1
i=1
m
1 X
ū = un−i .
m
i=1

10
Simplifications Usually Made in Risk Management

Simplifications that are commonly made:

• At short horizons, continous compounding makes little

difference, so we can approximate ui by (Si − Si−1 )/Si−1 .

• At short horizons, means are small relative to volatilities, so

we can assume that the mean value of ui is approximately
zero.

• Replace m − 1 by m for simplicity.

This yields
m
1 X 2
σn2 = un−i .
m
i=1

11
Weighting Scheme

• The previous estimate gives equal weight to all past

observations in the sample.

• Instead of assigning equal weights to the observations, we

can set
m
X
σn2 = 2
αi un−i ,
i=1

where
m
X
αi = 1.
i=1

• If we choose αi < αj when i > j, less weight is given to

older observations.

12
ARCH(m) Model

• In an ARCH(m) model we also assign some weight to the

long-run variance rate, VL :

m
X
σn2 = γVL + 2
αi un−i ,
i=1

where

m
X
γ+ αi = 1.
i=1

13
EWMA Model

• In an exponentially weighted moving average (EWMA)

model, the weights assigned to the un−i
2 decline
exponentially as we move back through time:

αi+1 = λαi .

• This leads to

σn2 = λσn−1
2 2
+ (1 − λ)un−1 .

14
Attractions of EWMA

• Relatively little data needs to be stored.

• We need only remember the current estimate of the

variance rate and the most recent observation on the
market variable.

• The EWMA approach is designed to track changes in

volatilitiy. If λ is low, the estimate is very responsive to the
most recent daily percentage change.

• λ = 0.94 has been found to be a good choice across a wide

range of market variables (RiskMetrics model).

15
GARCH(1, 1)

• In GARCH(1, 1) we also assign some weight to a long-run

average variance rate VL :

σn2 = γVL + αun−1

2 2
+ βσn−1 ,

where the weights sum to 1:

γ + α + β = 1.

• The EWMA model is a special case of GARCH(1,1) where

γ = 0, α = 1 − λ, and β = λ.

16
GARCH(1, 1)

• Settings ω = γVL , the GARCH(1, 1) model can also be

written as
σn2 = ω + αun−1
2 2
+ βσn−1

with long-run variance

ω
VL = .
1−α−β

• For a stable GARCH(1,1) model, we require α + β < 1.

17
Example

• Suppose
σn2 = 0.000002 + 0.13un−1
2 2
+ 0.86σn−1 .

• The long-run variance rate is 0.0002 so that the long-run

volatility per day is 1.4%.

• Suppose the previous estimate of the volatility was 1.6%

per day and the most recent percentage change in the
market variable is 1%.

• The new variance rate is

0.000002+0.13×0.0001+0.86×0.000256 = 0.00023336.
The new volatility is 1.53% per day.
18
Other Models

• The GARCH(p, q) model is given by

p
X q
X
σn2 =ω+ 2
αi un−i + 2
βj σn−j .
i=1 j=1

• Many other GARCH models have been proposed.

• For example, we can design a GARCH model so that the

weight given to ui2 depends on whether ui is positive or
negative.

19
Maximum Likelihood Estimation

• How can we estimate the parameters of the models

discussed so far based on historical data?

• Approach: maximum likelihood estimation.

• In maximum likelihood methods, we choose parameters

that maximize the likelihood of the observations occurring.

20
Example 1

• We observe that a certain event happens one time in ten

trials. What is our estimate of the proportion of the time, p,
that it happens?

• The probability of the outcome is:

 
10
· p(1 − p)9 .
1

• We maximize this probability to obtain a maximum

likelihood estimate: p̂ = 0.1.

21
Example 2

• Estimate the variance ν of observations from a normal

distribution with mean zero:

n
" !#
Y 1 −ui2
Maximize : √ exp
2πν 2ν
i=1
n
" #
X ui2
or equivalently : − ln(ν) −
ν
i=1
n
1X 2
This gives : ν̂ = ui .
n
i=1

22
Application to GARCH(1, 1)

• To estimate the parameters of a GARCH model, we choose

the parameters that maximize
n
" #
X ui2
− ln(νi ) − ,
νi
i=1

where νi = σi2 is the variance estimated for day i.

• We estimate the parameters of a GARCH(1, 1) model for

the S&P 500 between July 2005 and August 2010.

• See Excel file “GARCH [Link]” for all calculations.

23
Time Series on S&P 500

24
Calculations on S&P 500 Data

Day Si ui νi = σi2 − ln νi − ui2 /νi

1 1221.13
2 1229.35 0.006731
3 1235.20 0.004759 0.00004531 9.5022
4 1227.04 −0.006606 0.00004447 9.0393
...
1279 1079.25 −0.004024 0.00016327 8.6209
10 228.2349

25
Estimated Daily Volatility on the S&P 500

26
Variance Targeting

• One way of implementing GARCH(1, 1) that increases

stability is by using variance targeting.

• We set the long-run average volatility equal to the sample

volatility.

• Only two other parameters then have to be estimated.

27
Forecasting Future Volatility

• For the GARCH(1, 1) model, a few lines of algebra show

that:
2
En−1 [σn+t ] = VL + (α + β)t (σn2 − VL ).

• To estimate the volatility for an option lasting T days, we

must integrate this expected variance from day n to n + T .

• The average variance rate per day between today and T

days from now is
" #
1 T 2 1 − e−aT 2
Z
En−1 σn+t dt = VL + [σn − VL ],
T 0 aT
 
where a = ln α+β1
.

28
Volatility Term Structures

• Thus, the volatility per year for an option lasting T days is

s
1 − e−aT 2
 
σ̄n (T ) = 252 VL + [σn − VL ] .
aT

• The GARCH (1,1) model allows us to predict volatility term

structure changes.

• When σn changes by ∆σn , GARCH(1, 1) predicts that σ̄n (T )

changes by

1 − e−aT σn
252 · · ∆σn .
aT σ̄n (T )

29
Correlation and Covariance

• As we have discussed before, the correlation between

returns on different investments determines to what extent
risks can be diversified.

• The coefficient of correlation between two variables V1 and

V2 is defined as

E[V1 V2 ] − E[V1 ]E[V2 ]

ρ1,2 =
sd[V1 ]sd[V2 ]

• The covariance is

Cov(V1 , V2 ) = E[V1 V2 ] − E[V1 ]E[V2 ].

30
Independence

• V1 and V2 are independent if the knowledge of one does

not affect the probability distribution for the other:

f (V2 |V1 = x) = f (V2 ),

where f (·) denotes the probability density function.

• Independence is stronger than having zero correlation.

Correlation is a measure of linear dependence.

31
Independence is Not the Same as Zero Correlation

• Suppose V1 = −1, 0, +1 with equal probabilities

• Distribution of V2 :

1 if V = −1 or V = 1
1 1
V2 =
0 if V1 = 0.

• V2 is clearly dependent on V1 (and vice versa) but the

coefficient of correlation is zero.

32
Types of Dependence

33
Monitoring Correlation Between Two Variables X and Y

• Define xi = (Xi − Xi−1 )/Xi−1 and yi = (Yi − Yi−1 )/Yi−1 .

• varx,n : daily variance of x on day n, calculated at n − 1

• vary ,n : daily variance of y on day n, calculated at n − 1

• covn : covariance on day n, calculated at n − 1

• The correlation is
covn
√ .
varx,n vary ,n

34
Covariance

• The covariance on day n is

covn = En−1 [xn yn ] − En−1 [xn ]En−1 [yn ].

• It is usually approximated as En−1 [xn yn ].

• EWMA:

covn = λcovn−1 + (1 − λ)xn−1 yn−1 .

• GARCH(1, 1):

covn = ω + αxn−1 yn−1 + βcovn−1 .

35
Positive Semi-Definite Condition

• A variance-covariance matrix Ω is internally consistent if

the positive semi-definite condition

w ′ Ωw ≥ 0

holds for all vectors w.

• Example: the variance-covariance matrix

 
1 0 0.9
 0 1 0.9
 

0.9 0.9 1

is not consistent.
• Try w = (−0.5, −0.5, 1) → w ′ Ωw = −0.3.

36
Bivariate Normal Distributions

• Suppose that V1 and V2 follow a bivariate normal

distribution:
• V1 ∼ N(µ1 , σ12 )
• V2 ∼ N(µ2 , σ22 )
• corr[V1 , V2 ] = ρ

• Conditional on the value of V1 , V2 is normally distributed

with mean
V1 − µ 1
µ2 + ρσ2
σ1
and standard deviation
p
σ2 1 − ρ2 .

37
Multivariate Normal Distribution

• Multivariate normal distributions are well understood and

relatively easy to deal with.

• A variance-covariance matrix defines the variances of and

correlations between variables.

• To be internally consistent, a variance-covariance matrix

must be positive semi-definite.

38
Generating Random Samples for Monte Carlo Simulation

• =[Link](RAND()) gives a random sample from a

normal distribution in Excel.

• For a multivariate normal distribution, a method known as

Cholesky’s decomposition can be used to generate random
samples.

• The idea is to first simulate independent draws of a normal

distribution, and then take linear combinations of those to
generate a sample with the appropriate variance-covariance
structure.

39
Factor Models

• When there are N variables, Vi (i = 1, 2, . . . , N), there are

N(N − 1)/2 correlations.

• We can reduce the number of correlation parameters that

have to be estimated with a factor model.

• A one-factor model has a single common component and N

independent idiosyncratic components.

40
One-Factor Model

• If the Ui have standard normal distributions, we can set

q
Ui = ai F + 1 − ai2 Zi

where the common factor F and the idiosyncratic

component Zi have independent standard normal
distributions.

• The correlation between Ui and Uj is ai aj .

• The advantage is that the resulting covariance matrix is

always positive-semidefinite, and we only need to estimate
N parameters ai instead of N(N − 1)/2 correlations.

41
Gaussian Copula Models

• Suppose we wish to define a correlation structure between

two variable V1 and V2 that do not have normal
distributions.

• We transform the variable V1 to a new variable U1 that has

a standard normal distribution on a
“percentile-to-percentile” basis.

• We transform the variable V2 to a new variable U2 that has

a standard normal distribution on a
“percentile-to-percentile” basis.

• U1 and U2 are assumed to have a bivariate normal

distribution.
42
Correlation Structure Between the Vi ’s is Defined by that Be-
tween the Ui ’s

43
Example: V1 Mapping to U1

Distribution of V1 :

V1 Percentile U1
0.2 20 −0.84
0.4 55 0.13
0.6 80 0.84
0.8 95 1.64

44
Example: V2 Mapping to U2

Distribution of V2 :

V2 Percentile U2
0.2 8 −1.41
0.4 32 −0.47
0.6 68 0.47
0.8 92 1.41

45
Example: Joint Distribution

• Use of copula: correlation structure for modeling joint

distribution, while taking individual distributions as given.

• What is the probability that V1 and V2 are both less than

0.2?

• One-to-one mapping: same as probability that U1 < −0.84

and U2 < −1.41.

• When copula correlation is 0.5, this probability is

M(−0.84, −1.41; 0.5) = 0.043,

where M is the cumulative distribution function (CDF) for

the bivariate normal distribution.
46
Other Copulas

• Instead of a bivariate normal distribution for U1 and U2 we

can assume any other joint distribution.

• One possibility is the bivariate Student’s t-distribution.

• t-distribution has fatter tails than normal distribution.

47
5000 Random Samples from the Bivariate Normal

• Correlation is 0.5.
48
5000 Random Samples from the Bivariate t-Distribution

• 4 degrees of freedom, correlation is 0.5.

49
Multivariate Gaussian Copula

• We can similarly define a correlation structure between

V1 , V2 , . . . Vn .

• We transform each variable Vi to a new variable Ui that has

a standard normal distribution on a
“percentile-to-percentile” basis.

• The Ui ’s are assumed to have a multivariate normal

distribution.

• In a factor copula model, the correlation structure between

the Ui ’s is generated by assuming one or more factors.

50
Credit Default Correlation

• The credit default correlation between two companies is a

measure of their tendency to default at about the same
time.

• Default correlation is important in risk management when

analyzing the benefits of credit risk diversification.

• It is also important in the valuation of some credit

derivatives.

51
Model for Loan Portfolio

• The Vasicek Model is an application of the one-factor

Gaussian copula model to loan portfolios.

• We map the time to default for company i, Ti , to a new

variable Ui and assume
q
Ui = ai F + 1 − ai2 Zi ,
where F and the Zi have independent standard normal
distributions with CDF Φ.

• Define Qi as the cumulative probability distribution of Ti .

• P(Ti < T ) = P(Ui < U) if we set Qi (T ) = Φ(U) ⇒

T = Qi−1 (Φ(U)), U = Φ−1 (Qi (T )).

52
Model for Loan Portfolio

• Conditional distribution of Ui :
 
U − ai F 
P(Ui < U|F ) = Φ  q .
1 − ai2

• Hence, conditional probability of individual default is

P(Ti < T |F ) = P(Ui < Φ−1 (Qi (T ))|F )

 
−1
Φ [Qi (T )] − ai F 
= Φ q .
1 − ai2

53
Model for Loan Portfolio

• Assuming the Qi ’s and ai ’s are the same for all companies,

√ !
Φ−1 [Q(T )] − ρF
P(Ti < T |F ) = Φ √ ,
1−ρ

where ρ is the copula correlation.

• The worst-case default rate (WCDR) for the portfolio on a

time horizon of T and with a confidence of X (e.g., 95%) is
√ !
Φ−1 [Q(T )] + ρΦ−1 (X )
WCDR(T , X ) = Φ √ .
1−ρ

54