An Overview of Value at Risk

Abridged & Annotated

By Darrell Duffie & Jun Pan

1. Background & Definition

Value at Risk (VAR) – “for a given portfolio, time horizon, and probability p,
the p VaR is defined as a threshold loss value, such that the probability that
the loss on the portfolio over the given time horizon exceeds this value is p”
– Wikipedia definition

Market Risk – the risk of unexpected changes in prices or rates of market-

traded entities (i.e. stocks, bonds, derivative instruments)

Objectives in Managing Market Risk (Duffie & Pan):

1. Measure the extent of exposure by trade, profit center, & in various
aggregates
2. Charge each position a cost of capital appropriate to its market value &
risk
3. Allocate capital, risk limits, & other scarce resources such as
accounting capital to profit centers
4. Provide information on the firm’s financial integrity and risk-management
technology to contractual counterparties, regulators, auditors, rating
agencies, the financial press, & others whose knowledge might improve the
firm’s terms of trade, or regulatory treatment & compliance
5. Evaluate & improve the performance of profit centers, in light of the
risks taken to achieve profits
6. Protect the firm from financial distress costs
Ingredients for Computing VaR (Duffie & Pan):

1. a model of random changes in the prices of the underlying instruments

(equity indices, interest rates, foreign exchange rates, etc.)
2. a model for computing the sensitivity of the prices of derivatives to the
underlying prices

Recipe for Estimating VaR (Duffie & Pan):

1. Build a model for simulating changes in prices across all underlying

markets, and perhaps changes in volatilities as well, over the VaR time
horizon.
2. Build a database of portfolio positions, including the contractual
definitions of each derivative. Estimate the size of the current position
in each instrument.
3. Develop a model for the revaluation of each derivative for given changes in
the underlying market prices (and volatilites).
4. Simulate the change in market value of the portfolio, for each scenario of
the underlying market returns. Independently generate a sufficient number
of scenarios to estimate the desired critical values of the profit-and-loss
distribution with the desired level of accuracy

2. Price Risk

2.1. The Basic Model of Return Risk (Duffie & Pan)

A daily return model can be written as:

Rt+1 = μt + σt εt+1
Where
μt is the expectation of the return Rt+1 conditional on information
available at day t
σt is the standard deviation of Rt+1 conditional on information
available at day t
εt+1 is a “shock” with a conditional mean of zero and a conditional
standard deviation of one

The volatility of the asset is the annualized standard deviation of return

i.e. the volatility at day t is equal to √𝑛 ∗ 𝛔𝒕 where n is the number of
trading days per year

A plain-vanilla model of returns is one in which μ and σ are constant

parameters, and the “shocks” are i.i.d. and normally distributed
2.2. Risk-Neutral Versus Actual Value at Risk (Duffie & Pan)
Risk-neutral pricing – the idea that a security can be priced so that its
current value matches the expected discounted cash flow paid by the
security

Derivative prices are generally modeled assuming risk-neutral parameters

In short time horizons, the difference between risk-neutral and actual

price behavior is negligible

Generally speaking, actual price models are used for risk management

2.3. Fat Tails

Fatter Tail = larger overnight VaR at higher confidence intervals

Kurtosis - E((εt+1)^4), the expected fourth power of the “shock”
and a standard measure of “fat-tailedness" (Duffie & Pan)
- kurtosis estimates are very sensitive to extremely large
returns

An alternate measure of fat-tailedness:

-the number of standard deviations represented by the associated critical

values of the return distribution are an especially useful way to gauge
“fat-tailedness” when concerned with VaR
-The more greater the number of standard deviations for a particular
critical value, the “fatter” the tail
- When using kurtosis as a measure of “fat-tailedness,” S&P Returns from
1986-96 are very “fat-tailed”
- When considering standard deviations, the returns are in fact slightly
“thinner-tailed” than a normal distribution
- Negative skewness implies that “large negative returns are more common
than large positive returns” (Duffie & Pan)
 - Figures 4 & 5 show the propensity for “fat-tailedness” in
various real world markets when measured by the relation
between critical values and their associated standard
deviations
- Left-Tail Fatness is relevant when measuring VaR for long
positions, Right-Tail Fatness when measuring VaR for short
positions
Two Particular Sources for Fat Tails:
1. “Jumps” i.e. “significant unexpected discontinuous changes in
prices” (Duffie & Pan)
2. Stochastic Volatility – “volatility that changes at random
over time, usually with some persistence” (Duffie & Pan)

2.4. Jump-Diffusions

Jump-Diffusion Model: the same as the plain-vanilla model, except that an

a predetermined yearly frequency λ the daily return “shock” is “jumped”
by adding an independent normal random variable, with mean 0 and standard
deviation ν
- jump arrivals are Poisson and thus independent of past
“shocks”
- the jump-diffusion model captures the possibility of extreme
loss with more accuracy than the plain-vanilla model
Things to note:
1.) The two models pictured are calibrated to have the same annual
volatility. The jump diffusion model is expected to have one “jump”
per year with a standard deviation of 10 %
2.) While the 2-Week 99% VaR are fairly similar between the two models,
the greater distinction between the two is seen further away from
the mean
3.) The jump-diffusion model predicts that once every 140 years, one
will experience a shock equating to an overnight loss of 25% of
share value. According to the plain-vanilla model, one might wait
billions of years before such an occurrence. Daily returns of 5
standard deviations of share value occurred numerous times from
1986-96. Even returns of 10 standard deviations have been noted in
rare occasions. However, according to the plain-vanilla model, a
return of 5 standard deviations is expected less than once per
million days, while a return of 10 standard deviations is not even
expected quite once every 10^23 days.
Note:
1.) Once again, the two models pictured are calibrated to have the same
annual volatility
2.) This time, in the jump diffusion model, there are two expected
“jumps” per year, each with a standard deviation of 5 %
3.) We can see that this model accounts even more generously for extreme
loss. This is implied in part by the greater difference at the 99%
level than the previous model. Thus, the jump-diffusion of Figure 7
will allow for extreme loss at an even greater rate than that of
Figure 6

2.5. Stochastic Volatility

Stochastic volatility: volatility that changes at random over time,

usually with some persistence – i.e. a high current volatility implies a
high volatility in the near future and a low current volatility implies a
low volatility in the near future (Duffie & Pan)

Our Model:

We will consider only Markovian stochastic volatility models, in

which the current volatility is a function of the current level of
volatility, among other variables, but not the path which the volatility
has taken up until the current time. Thus, we get the form
 Where F is some function of the 3 enclosed variables, and the z are white
noise just like the ε of the plain-vanilla model of daily returns.

Note:
1.) This formulation for volatility rules out any sort of
structured dependence of the distribution of changes in
volatility on other variables such as the volatility in
related markets and overarching macro-economic factors, which
in real world application may be very relevant
2.) This model does allow the possibility of correlation between
the volatility shock z and the return shock ε. Negative correlation
implied a negatively skewed distribution of daily returns, and thus the VaR estimate
for a short position would be less than that of a long position of equal value

Classes of the Markovian Stochastic Volatility Model:

2.5.1 Regime-Switching Volatility

regime-switching model – a model in which volatility behaves according to a

finite-state Markov chain (Duffie & Pan)

Two-State Example:

Consider a situation where there are two possible volatilities, νa and νb.
Let the following matrix represent the transitional probabilities of σt between
νa and νb :

The following figure illustrates an example using estimates inferred from oil
pricing (Duffie & Gray 1995):

Note that this

particular example shows a high rate of volatility persistence. In practice,
the number of possible states is not limited to two. Rather, it can reach any
finite number, although as one would expect as one adds more states the related
mathematical models grow in complexity.

2.5.2 Auto-Regressive Volatility

The log-auto-regressive model:

Where alpha, gamma & kappa are constants, and gamma indicates the level of
volatility persistence. The closer gamma is to 1, the higher the level of
persistence. Except in cases of “explosive volatility,” we assume that
-1<gamma<1. If the shocks are independent to returns and volatility, the term
structure of conditional volatility is the following:

If the shocks are not independent to returns and volatility, the term structure
of conditional volatility can be calculated but that is beyond the scope of
this overview. Nonetheless, allowing this correlation is often necessary in
practical application.

2.5.3 Garch

Generalized Autoregressive Conditional Heteroscedasticity (GARCH) model for

volatility:

Where alpha, beta, & gamma are positive constants. The gamma here is a stronger
persistence indicator than in the autoregressive model of 2.5.2 due to the
quadratic nature of the volatility term. Returning to crude oil, Duffie & Gray
(1995) estimated the following parameters (T-statistics in parenthesis):

Assuming the “non-explosivity condition,” i.e. delta = Beta + gamma < 1, the
steady-state volatility is the following:

It can be shown that the term structure of volatility for the GARCH model is
One major drawback of this model is related to the impact of persistence.
Because the term involving returns Rt is quadratic any sudden large loss or gain
coupled with a reasonably high persistence factor gamma will cause a
persistently high volatility forecast.

2.5.4 EGARCH
exponential GARCH:

Corresponding term structure of volatility:

Where Ck is a fairly complicated constant beyond the purview of these notes. It

has been shown that under certain circumstances, EGARCH & the log-auto-
regressive models converge to the same model.

2.5.5 Cross-Market GARCH

The volatilities across markets are often related, and this relationship can be
captured by the multivariate GARCH model.
2-market example (Duffie & Gray, 1995):

With Beta & Gamma assumed to be diagonal, Duffy & Gray (1995) get the following
estimates for (a) heating oil & (b) crude oil using maximum likelihood
estimation over a bivariate GARCH model (T-statistics in parenthesis):
Note that the t-statistics in this model show great improvement upon the
estimates in the univariate GARCH model of section 2.5.3

2.6 Term Structures of Tail-Fatness & Volatility

For the plain-vanilla model, tail-fatness & volatility exhibit a flat term
structure. The term structure generally exhibited by tail-fatness & volatility
are dependent on the source of tail fatness.

1. Jumps

We can see that the term structure of kurtosis for these 3 jump-diffusion
models is continuously declining to that of a Normal Distribution, i.e. K
= 3. This follows from the central limit theorem.
Unlike Kurtosis, whose term structure is steadily decreasing, the standard
deviations to 0.01 critical values of the 3 jump-diffusion models first
increase and then slowly decrease to the standard 2.326 of a Normal
variable. This is due to the fact that since the “jump” is expected to occur
no more than 3 times a year in any of the models, the standard deviations
are relatively low for a time horizon of one day. However, as the time
horizon increases, the effect of the “jumps” become more poignant and reach
their respective peak for each model somewhere between 1 day & 100 days.
However, over larger time horizons, the effect of the “jumps” become more
and more negligible and the 3 jump-diffusion models tend toward normal.

2. Stochastic Volatility

For this section we will let the mean returns be constant, the “shocks” be
i.i.d., and the volatility stochastic. Due to the stochastic nature of the
volatility, it is theoretically possible for the term structure to take any
form.
The above figure represents the term structure of volatility for an
autoregressive model of stochastic volatility for the Hang Seng Index fitted
by Heynen & Kat (1993). The maximum-likelihood estimates in this model are
alpha = -5.4, gamma = 0.38, & kappa = 1.82. Line B represents the term
structure of stochastic volatility for an initial volatility at the mean of
the steady-state distribution. Line A & Line C represent the term structure
for an initial volatility of one standard deviation above & below the mean,
respectively. One can easily see how in this model, the term structure will
approach a common asymptote despite the different starting values.

Setting the initial volatility equal to the steady-state mean level

volatility from Figure 12, we can observe in figure 13 that under such
conditions, the term structure of Normalized Kurtosis first increases, then
decreases over time toward the flat level of K = 3 rendered by a normal
model. This hump is a result of the mixture of normal models with various
variances. However, in accordance with the Central Limit Theorem, over
longer Time Horizons it will tend towards normal.

Line A shows the complications that can arise when the initial volatility is
allowed a level of randomness. This, combined with the allowance for “jumps”
in the model, can cause problems related to bias & misspecification.
Nonetheless, it is a complication that can be necessary considering that
Jorion[1989] finds evidence for both jumps and stochastic volatility modeled
in GARCH form.

However, time horizons in years are rarely relevant to practical

applications in risk management.
 The above figure shows the term structure of kurtosis for stochastic
volatility models of 3 different indices. As one can see, when the model is
restricted in time horizons more applicable to risk management, the term
structure gives the impression of monotone increasing behavior. The British
Pound (A) exhibits a very high level of both mean reversion and volatility
of volatility. The S&P 500 (C) is far more moderate, and the Hang Seng Index
is in between. These results are based on the maximum likelihood estimates
by Heynen & Kat [1993]

3. Mean Reversion

Let the volatility in our daily Returns model be constant to go along with
i.i.d. shocks and introduce mean reversion into the model. One possible way
is to let μt = α(R*-R-t-1) with alpha > 0 serves as a dampener in order to
center the return model around an asymptotic mean R*. In this model the term
structure of volatility is decreasing asymptotically to a positive value,
and the term structure of volatility is constant.

2.7 Estimating Current Volatility

A major component to estimating VaR is coming up with a good estimate of the

current volatility of the market one is concerned with. We will now consider a
handful of estimation methods.

2.7.1 Historical Volatility

The historical volatility is a standard uncouth estimator for the current

volatility and is given by:

Where

This constitutes a maximum likelihood estimator which is optimal in a plain-

vanilla context. However, historical data tends to strongly imply, even
aver, that volatilities are not constant.

2.7.2 Black-Scholes Implied Volatility

The standard Black-Scholes equation gives the value of a European option at

time t, given the underlying price Pt, the strike K, the time τ to expiry, the
constant interest rate r, and the volatility σ. It is given by

Due to the fact that this formula is monotone increasing on σ, with the option
price Ct one can obtain a value for the volatility σ using the Black-Scholes
Implied Volatility Function:

Although the stochastic complexity of the Black Scholes formula prevents the
existence of an explicit formula for the above, an implied volatility can be
calculated to any desired accuracy using numerical techniques. In most markets,
option-implied volatility is more reliably accurate than any method utilizing
historic data.
2.7.3 Option-Implied Stochastic Volatility

Assume our model’s shocks are independent with respect to returns and
volatility. One can use the following slightly modified version of Black
Scholes to obtain the price of an option:

Is the root-mean-squared term volatility, CBS is the Black-Scholes formula, &

E* is the risk-neutral expectation. With this altered version of the option
price, one can then follow the same method in the previous sub-section to
obtain an estimate for the implied volatility.

2.7.4 Day-of-the-week and Other Seasonal Volatility Effects

In certain cases, the day of the week, the season, or other seasonal
effects can influence the volatility. In the realm of energy products, the
demand for heating oil and gasoline depend upon winter and summer weather
patterns, respectively. Thus, their pricing models exhibit greater volatility
in the summer and winter months, respectively.

2.8 Skewness

skewness – “a measure of the degree to which positive deviations from the mean
are larger than negative deviations from the mean, as measured by the expected
third power of these deviations”

Skewness that is caused only by skewness in the shocks will be “diversified

away” over time in accordance with the central limit theorem, and its term
structure will tend to 0 over long time horizons.

Skewness related to correlation between shocks and changes in volatility would

remain present over long time horizons.

2.9 Correlations

One component not discussed in any further detail in this article but critical
to the real-world accurate estimation of price risk is the modeling of cross-
market correlations between the shocks of multiple markets.
Works Cited:

Duffie, D., & Pan, J. (1997). An Overview of Value at Risk. The Journal of

Derivatives, 4(3), 7-49. doi:10.3905/jod.1997.407971

GIOVANNINI, A., & JORION, P. (1989). The Time Variation of Risk and Return in the

Foreign Exchange and Stock Markets. The Journal of Finance, 44(2), 307-325.

doi:10.1111/j.1540-6261.1989.tb05059.x

Heynen, R. C., & Kat, H. M. (1994). Volatility Prediction. The Journal of

Derivatives, 2(2), 50-65. doi:10.3905/jod.1994.407912

Value at risk - Wikipedia. (n.d.). Retrieved November 8, 2016, from

https://en.wikipedia.org/wiki/Value_at_risk

“Volatility in Energy Prices,” with S. Gray and P. Hoang, in Managing

Energy Price Risk, edited by Lou Pai and Peter Field, Risk Publications,

1995, revised for second edition, 1999, pp. 273-290.

Note: the compiler of these notes claims no intellectual ownership of the ideas
presented here. All intellectual rights are relinquished to the original
authors. All images belong to the original authors Duffie & Pan.