Elements of Financial Risk Management Second Edition © 2012 by Peter Christoffersen 1 Non-Normal...

Elements of Financial Risk Management Second Edition © 2012 by Peter Christoffersen

1

Non-Normal Distributions

Elements of

Financial Risk Management

Chapter 6

Peter Christoffersen

Overview• Third part of the Stepwise Distribution Modeling (SDM)

approach: accounting for conditional nonnormality in portfolio returns.

• Returns are conditionally normal if the dynamically standardized returns are normally distributed.

• Fig.6.1 illustrates how histograms from standardized returns typically do not conform to normal density

• The top panel shows the histogram of the raw returns superimposed on the normal distribution and the bottom panel shows the histogram of the standardized returns superimposed on the normal distribution


2

Figure 6.1: Histogram of Daily S&P 500 Returns and Histogram of GARCH Shocks


3

Learning Objectives• We introduce the quantile-quantile (QQ) plot,

which is a graphical tool better at describing tails of distributions than the histogram.

• We define the Filtered Historical Simulation approach which combines GARCH with historical simulation.

• We introduce the simple Cornish-Fisher approximation to VaR in non-normal distributions.

• We consider the standardized Student’s t distribution and discuss the estimation of it.


4

Learning Objectives

• We extend the Student’s t distribution to a more flexible asymmetric version.

• We consider extreme value theory for modeling the tail of the conditional distribution

• For each of these methods we will consider the Value-at-Risk and the expected shortfall formulas


5

Visualising Non-normality Using QQ Plots

• Consider a portfolio of n assets with Ni,t units or shares of asset i then the value of the portfolio today is


6

• Yesterday’s portfolio value would be

• The log return can now be defined as


• Allowing for a dynamic variance model we can say


7

• where PF,t is the conditional volatility forecast• So far, we have relied on setting D(0,1) to N(0,1), but we

now want to assess the problems of the normality assumption


• QQ (Quantile-Quantile) plot: Plot the quantiles of the calculated returns against the quantiles of the normal distribution.

• Systematic deviations from the 45 degree angle signals that the returns are not well described by normal distribution.

• QQ Plots are particularly relevant for risk managers who care about VaR, which itself is essentially a quantile.


8


• 1) Sort all standardized returns in ascending order and call them zi

• 2) Calculate the empirical probability of getting a value below the value i as (i-.5)/T

• 3) Calculate the standard normal quantiles as • 4) Finally draw scatter plot


9

• If the data were normally distributed, then the scatterplot should conform to the 45-degree line.

Figure 6.2: QQ Plot of Daily S&P 500 Returns


10

Figure 6.2: QQ Plot of Daily S&P 500 GARCH Shocks


11

Filtered Historical Simulation Approach

• We have seen the pros and cons of both data-based and model-based approaches.

• The Filtered Historical Simulation (FHS) attempts to combine the best of the model-based with the best of the model-free approaches in a very intuitive fashion.

• FHS combines model-based methods of variance with model-free method of distribution in the following fashion.


12


• Assume we have estimated a GARCH-type model of our portfolio variance.

• Although we are comfortable with our variance model, we are not comfortable making a specific distributional assumption about the standardized returns, such as a Normal or a distribution.

• Instead we would like the past returns data to tell us about the distribution directly without making further assumptions.


13

dt~

Filtered Historical Simulation Approach• To fix ideas, consider again the simple example of

a GARCH(1,1) model


14

• where

• Given a sequence of past returns, we can estimate the GARCH model.

• Next we calculate past standardized returns from the observed returns and from the estimated standard deviations as

Filtered Historical Simulation Approach• We will refer to the set of standardized returns as

• To calculate the 1-day VaR using the percentile of the database of standardized residuals


15

• Expected shortfall (ES) for the 1-day horizon is

• The ES is calculated from the historical shocks via


• where the indicator function 1(*) returns a 1 if the argument is true and zero if not

• FHS can generate large losses in the forecast period even without having observed a large loss in the recorded past returns

• FHS deserves serious consideration by any risk management team


16

The Cornish-Fisher Approximation to VaR

• We consider a simple alternative way of calculating Value at Risk, which has certain advantages:

• First, it allows for skewness and excess kurtosis.• Second, it is easily calculated from the empirical

skewness and excess kurtosis estimates from the standardized returns.

• Third, it can be viewed as an approximation to the VaR from a wide range of conditionally nonnormal distributions.


17



18

• Standardized portfolio returns is defined by

• where D(0,1) denotes a distribution with a mean equal to 0 and a variance equal to 1

• i.i.d. denotes independently and identically distributed

• The Cornish-Fisher VaR with coverage rate, p, can be calculated as



19

• Where

• Where is the skewness and is the excess kurtosis of the standardized returns

• If we have neither skewness nor excess kurtosis so that . , then we get the quantile of the normal distribution



20

• Consider now for example the one percent VaR, where

• Allowing for skewness and kurtosis we can calculate the Cornish-Fisher 1% quantile as

• and the portfolio VaR can be calculated as


• Thus, for example, if skewness equals –1 and excess kurtosis equals 4, then we get


21

• which is much higher than the VaR number from a normal distribution, which equals 2.33PF,t+1



22

• The expected shortfall can be derived as

• Where



23

• Recall that the ES for the normal case is

• Which can be derived by setting in the equation for .

• The CF approach is easy to implement and we avoid having to make an assumption about exactly which distribution fits the data best

The Standardized t distribution


24

• The Student’s t distribution is defined by

• (*) notation refers to the gamma function• the distribution has only one parameter d• In the Student’s t distribution we have the following

first two moments



25

• Define Z by standardizing x so that,

• The Standardized t distribution, , is then defined as

• where



26

• In standardized t distribution random variable z has mean equal to zero and a variance equal to 1

• Note also that the parameter d must be larger than two for standardized distribution to be well defined

• In distribution, the random variable, z, is taken to a power, rather than an exponential, which is the case in the standard normal distribution where

• The power function driven by d will allow for distribution to have fatter tails than the normal

The Standardized t distribution• The distribution is symmetric around zero, and the

mean , variance 2, skewness 1, and excess kurtosis 2 of the distribution are


27

The Standardized t distribution• Note that d must be higher than 4 for the kurtosis to be

well defined.

• Note also that for large values of d the distribution will have an excess kurtosis of zero, and we can show that it converges to the standard normal distribution as d goes to infinity.

• For values of d above 50, the distribution is difficult to distinguish from the standard normal distribution.


28

Maximum Likelihood Estimation

• We can combine dynamic volatility model such as GARCH with the standardized t distribution to specify our model portfolio returns as


29

• If we ignore the fact that variance is estimated with error, we can treat standardized return as a regular random variable, calculated as

Maximum Likelihood Estimation• The d parameter can then be chosen to maximize

the log likelihood function


30

Maximum Likelihood Estimation• If we want to jointly maximize over the parameter

d and we should adjust the distribution to take into account the variance


31

• We can then maximize

Maximum Likelihood Estimation• As a simple univariate example of the difference between

QMLE and MLE consider the GARCH(1,1)- model with leverage:


32

• We can estimate all the parameters in one step using lnL2 from before, which would correspond to exact MLE.

• Exact MLE is clearly feasible as the total number of parameters is only five

An Easy Estimate of d • There is a simple alternative estimation procedure

to the QMLE estimation procedure above. • If the conditional variance model has already been

estimated, then we are only estimating one parameter, namely d.

• The simple closed-form relationship between d and the excess kurtosis 2 suggests calculating 2, from the zt variable and calculating d from


33

Calculating VaR and ES• Having estimated d we can calculate the VaR of

the portfolio return


34

• as

• Where is the pth quantile of distribution. Therefore,

Calculating VaR and ES

• The formula for ES is,


35

QQ Plots• We can relate the t(d) to the standardized t(d) by


36

• where is the pth quantile of the distribution

QQ Plots

• We are now ready to construct the QQ plots,


37

• where zi denotes the ith sorted standardized return

Figure 6.3: QQ Plot of S&P500 GARCH Shocks Against the Standardized t Distribution


38

The Asymmetric t distribution• If one would like to have skewness in t distribution, the

asymmetric t distribution can be used,


39

The Asymmetric t distribution

• Here d1 > 2, and -1 < d2 < 1.

• Note that C(d1) = C(d) from the symmetric Student’s t distribution


40

Figure 6.4: The Asymmetric t Distribution


41

The Asymmetric t distribution

• In order to derive the moments of the distribution we first define,


42

The Asymmetric t distribution• The moments of the asymmetric t distribution can

be derived as


43

Figure 6.5: Skewness and Kurtosis in the Asymmetric t Distribution


44

The Asymmetric t distribution• Notice that the symmetric t distribution is nested in the

asymmetric t distribution and can be derived by setting d1 = d, d2 = 0 which implies A = 0 and B = 1

• Therefore,


45

• which yields

Estimation of d1 and d2

• MLE can be used to estimate the parameters of the asymmetric distribution, d1 and d2

• The only complication is that the shape of the distribution on each day depends on zt

• As before the likelihood function for the shock can be defined as,


46


• Where


47

• This assumes the estimation of is done without estimation error.


• Alternatively the joint estimation of volatility and distribution parameters can be done via,


48

• We can also estimate d1 and d2 using the sample moments.

• Equations below can be solved numerically

Calculating VaR and ES• Having estimated d1 and d2 we can calculate the

VaR of the portfolio return


49

• as

• Where is the pth percentile of the asymmetric t distribution.

Calculating VaR and ES

• is given by


50

• The ES is given by

QQ Plots• Knowing the CDF we can construct the QQ plot as,


51

• where zi denotes the ith sorted standardized return

• The asymmetric t distribution is cumbersome to estimate and implement but it is capable of fitting GARCH shocks from daily asset returns quite well

• The t distributions attempt to fit the entire range of outcomes using all the data available

• Consequently, the estimated parameters in the distribution (for example d1 and d2 ) may be influenced excessively by data values close to zero

Figure 6.6: QQ Plot of S&P 500 GARCH Shocks against the Asymmetric t Distribution


52

Extreme Value Theory (EVT)

• Typically, the biggest risks to a portfolio is the sudden occurrence of a single large negative return.

• Having an as precise as possible knowledge of the probabilities of such extremes is therefore at the essence of financial risk management.

• Consequently, risk managers should focus attention explicitly on modeling the tails of the returns distribution.

• Fortunately, a branch of statistics is devoted exactly to the modeling of these extreme values.


53

Extreme Value Theory (EVT)• The central result in EVTstates that the extreme tail of a

wide range of distributions can approximately be described by a relatively simple distribution, the so-called Generalized Pareto distribution.

• Virtually all results in Extreme Value Theory assumes that returns are i.i.d. and are therefore not very useful unless modified to the asset return environment.

• Asset returns appear to approach normality at long horizons, thus EVT is more important at short horizons, such as daily.


54

Extreme Value Theory (EVT)

• Unfortunately, the i.i.d assumption is the least appropriate at short horizons due to the time-varying variance patterns.

• We therefore need to get rid of the variance dynamics before applying EVT.

• Consider therefore again the standardized portfolio returns


55

The Distribution of Extremes• Define a threshold value u on the horizontal axis of the

histogram in Figure 6.1• As you let the threshold u go to infinity, the distribution

of observations beyond the threshold (y) converge to the Generalized Pareto Distribution, where


56

• With >0 and y≥u. Tail-index parameter controls the shape of the distribution tail

Estimating Tail Index Parameter, • If we assume that the tail parameter, , is strictly positive,

then we can use the Hill estimator to approximate the GPD distribution


57

• for y > u and > 0. Recall now the definition of a conditional distribution

• Note that from the definition of F(y) we have


58

Estimating Tail Index Parameter, • We can get the density function of y from F(y):

• The likelihood function for all observations yi larger than the threshold, u,

• where Tu is the number of observations y larger than u

• The log-likelihood function is therefore


59

Estimating Tail Index Parameter,

• Taking the derivative with respect to and setting it to zero yields the Hill estimator of tail index parameter

• We can estimate the c parameter by ensuring that the fraction of observations beyond the threshold is accurately captured by the density as in


60

Estimating Tail Index Parameter,

• Cumulative density function for observations beyond u is

• Solving this equation for c yields the estimate

Estimating Tail Index Parameter, • Notice that our estimates are available in closed form

• So far, we have implicitly referred to extreme returns as being large gains. As risk managers, we are more interested in extreme negative returns corresponding to large losses

• We can simply do the EVT analysis on the negative of returns (i.e. the losses) instead of returns themselves.


61

Choosing the Threshold, u

• When choosing u we must balance two evils: bias and variance.

• If u is set too large, then only very few observations are left in the tail and the estimate of the tail parameter, , will be very noisy.

• If on the other hand u is set too small, then the data to the right of the threshold does not conform sufficiently well to the Generalized Pareto Distribution to generate unbiased estimates of .


62

Choosing the Threshold, u • Simulation studies have shown that in typical data sets

with daily asset returns, a good rule of thumb is to set the threshold so as to keep the largest 50 observations for estimating

• We set Tu = 50.

• Visually gauging the QQ plot can provide useful guidance as well.

• Only those observations in the tail that are clearly deviating from the 45-degree line indicating the normal distribution should be used in the estimation of the tail index parameter,


63

Constructing the QQ Plot from EVT

• Define y to be a standardized loss


64

• The first step is to estimate and c from the losses, yi, using the Hill estimator

• Next, we need to compute the inverse cumulative distribution function, which gives us the quantiles


• We now set the estimated cumulative probability function equal to 1-p so that there is only a p probability of getting a standardized loss worse than the quantile, F-1

1-p


65

• From the definition of F(*), we can solve for the quantile to get


• We are now ready to construct the QQ plot from EVT using the relationship


66

• where yi is the ith sorted standardized loss

Figure 6.7: QQ Plot of Daily S&P 500 Tail Shocks against the EVT Distribution


67

Calculating VaR and ES from EVT

• We are ultimately interested not in QQ plots but rather in portfolio risk measures such as VaR.

• Using the loss quantile F-11-p defined above by


68

• The VaR from the EVT combined with the variance model is now easily calculated as

Calculating VaR and ES from EVT• We usually calculate the VaR taking -1

p to be the pth quantile from the standardized return so that


69

• But we now take F-11-p to be the (1-p)th quantile of the

standardized loss so that

• The expected shortfall can be computed using

• Where when <1

Calculating VaR and ES from EVT

• In general, the ratio of ES to VaR for fat-tailed distribution will be higher than that of the normal.

• When using the Hill approximation of the EVT tail the previous formulas for VaR and ES show that we have a particularly simple relationship, namely


70

• so that for fat-tailed distributions where > 0, the fatter the tail, the larger the ratio of ES to VaR:

Calculating VaR and ES from EVT• The preceding formula shows that when = 0.5 then the

ES to VaR ratio is 2

• Thus even though the 1% VaR is the same in the two distributions by construction, the ES measure reveals the differences in the risk profiles of the two distributions, which arises from one being fat-tailed

• The VaR does not reveal this difference unless the VaR is reported for several extreme coverage probabilities, p.


71

Figure 6.8: Tail Shapes of the Normal Distribution (blue) and EVT (red)


72

Elements of Financial Risk Management Second Edition © 2012 by Peter Christoffersen 1 Non-Normal...

Documents

Transcript of Elements of Financial Risk Management Second Edition © 2012 by Peter Christoffersen 1 Non-Normal...