INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Lesson 2: Value-at-Risk (VaR) and

Conditional VaR
Module: Financial Risk Management
 Introduction

• Introduction and Background to Tail Risk Measures

• Value-at-Risk (VaR Models)
• VaR: Theoretical underpinnings and mathematical formulation
• VaR: A few numerical examples
• Conditional VaR (CVaR) or Expected Shortfall (ES) Models
• CVaR or ES: Theoretical underpinnings and mathematical formulation
• CVaR or ES: A few numerical examples

2
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Value-at-Risk (VaR) Models: Part 1

Value-at-Risk (VaR) Models
• This measure is widely employed by banks for their portfolio performance
and Financial Markets for Margin requirements.

• Consider the distribution of “T” period returns shown below

• “We are X percent certain that we will not lose more than V dollars in time T.”

4
Value-at-Risk (VaR) Models
• “We are X percent certain that we will not lose more than V dollars in time T.”

• Thus, variable V here is the T-day (100-X)% VaR of the portfolio.

• The variable V is the VaR of the portfolio, it is a function of time T, confidence X%,

• Often VaR is expressed in terms of loss, i.e., negative gain.

5
Value-at-Risk (VaR) Models
• For example, when T=5 days, X=97 percent confidence, then VaR is the loss
over the next 5 days at (a) the 3rd Percentile of gains distribution or (b) the
97th percentile of the distribution of losses.

• When the gain distribution is considered, VaR is the 100-X percentile, and
VaR is the X percentile when the loss distribution is considered.

6
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Value-at-Risk (VaR) Models: Part 2

Value-at-Risk (VaR) Models
• If you are asked on any given day, what is the probability that you can lose
more than Rs. 10 Mn? Then you may reply in the following ways: (1) any
given day, with 95% confidence, your loss can not be more than Rs. 10 Mn
(or X% negative return, i.e., losses); [this is same as saying that there is a
5% chance that loss can be more than Rs. 10 Mn or X%].

• Or you may say that 5% daily VaR is Rs. 10 Mn (or X% negative return, i.e.,
losses) for gain distribution or 95% daily VaR is Rs 10 Mn for loss
distribution.

8
Value-at-Risk (VaR) Models
𝑽𝒂𝑹𝜶 𝑿 = 𝐦𝐢𝐧 𝒛 𝑭𝒙 𝒛 ≥ 𝜶 𝒇𝒐𝒓 𝜶 𝝐 [𝟎, 𝟏]

Compare & Zio (2015). Genetic

algorithms for condition-based
maintenance optimization under
uncertainty. European Journal of
Operational Research

9
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Value-at-Risk (VaR) Models: Part 3

Value-at-Risk (VaR) Models
• Here three important inputs are (a) time period (e.g., daily, weekly), (b) Level
of confidence (95% or 99%), and (c) Estimate of loss (in absolute Rs or %
return terms)

• To estimate this probability (1) You can assume that returns follow a
distribution (e.g., standard normal distribution), or (2) use empirical data

11
Value-at-Risk (VaR) Models
• VaR estimation with discrete empirical observations.

• In the later case, assume that you have 1001 (0-1000) return observations
(daily). You can order them in a decreasing fashion. The observation at
position 990 will represent the cut-off 1 percentile. If we consider the data to
be segments of unit intervals 0-1, 1-2 so on. This 990th position divides data
into 990 segments below and 10 segments above. Thus, it represents the
cut-off level for the 99% confidence level of returns. This level of returns
(990th position) will become your 99% VaR (daily) for loss distribution (or 1%
VaR for gain distribution).

12
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Value-at-Risk (VaR) Models: Simple

Example 1
 Examples
A hypothetical gamble is designed such that from a loss of INR 50 Mn to a gain of
INR 50 Mn all outcomes are equally likely, over a period of year.

• What is 1% (Gain distribution) VaR (or 99% VaR on loss distribution)

• What is 5% (Gain distribution) VaR (or 95% VaR on loss distribution)

14
 Examples
A hypothetical gamble is designed such that from a loss of INR 50 Mn to a gain of
INR 50 Mn all outcomes are equally likely, over a period of year.

• What is 1% (Gain distribution) VaR (or 99% VaR on loss distribution)

• What is 5% (Gain distribution) VaR (or 95% VaR on loss distribution)

Ans: Since it is a uniform distribution, the 1% loss is INR 49 Mn and the 5% loss is
INR 45 Mn

15
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Value-at-Risk (VaR) Models: Simple

Example 2
 Examples
VaR estimation with distributional assumptions

• Z-value corresponding to 1% probability (99% confidence) and 5% probability for

(95% confidence)
𝒓−𝒖
𝒛=
𝝈

17
 Examples
Question: An investor purchased a share of Rs 1000, whose continuously
compounded daily returns are distributed normally with a mean of 12.5% pa. and a
standard deviation of 50% pa.. How much VaR margin would he have to deposit, if it
is calculated at 99% level of confidence? We can assume 250 days of trading in a
year and that returns are serially uncorrelated. This is to be calculated for 1-day and
3-day.

Assume standard normal distribution and Z= -2.326 for a 1% significance level.

18
 Examples
Ans. A) For one day, 𝑢𝑑 =? ? ; 𝜎𝑑 =? ?
𝑟−𝑢
𝑧= =Z= -2.326
𝜎

With 99% confidence the maximum one-day loss will be =𝑢𝑑 + z*𝜎𝑑 = ??

Ans B) For 3 days, VaR=> 𝑢3𝑑 = ??; 𝜎3𝑑 =? ?

99% confidence the maximum three-day loss will be =𝑢3𝑑 + z*𝜎3𝑑 = ??

19
 Examples
50%
Ans. A) For one day, 𝑢𝑑 = 12.5/250 = 0.05%; 𝜎𝑑 = = 3.16%.
250

𝑟−𝑢
𝑧= =Z= -2.326
𝜎

With 99% confidence the maximum one-day loss will be =𝑢𝑑 + z*𝜎𝑑 = = 0.05%-
2.326*3.16%= -7.30%

Ans B) For 3 days, VaR=> 𝑢3𝑑 = 0.05%∗3; 𝜎3𝑑 = 3 ∗ 3.16%

99% confidence the maximum three-day loss will be =𝑢3𝑑 + z*𝜎3𝑑 = 0.05%*3-
2.326* 3 ∗ 3.16%= -12.55%
20
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Value-at-Risk (VaR) Models: Last few

words
 Value-at-Risk (VaR) Models: Pitfalls
• A portfolio manager can set up his risk Portfolio A

position as B, which may appear to be of

the same risk as A.

• Both positions will appear as of similar

risk on VaR. Portfolio B

22
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Conditional VaR (CVaR) or

Expected Shortfall (ES): Part 1
 CVaR or ES
• The VaR method covers all the possibilities within a certain confidence interval.
However, the position is exposed to those losses that are beyond the confidence
interval.
• To cover this exposure/risk, a more advanced version of the risk measure that is
CVaR is proposed.
• The measure computes expected losses given that (conditional upon) the
confidence level is breached. That is, what were to happen if the scenarios
beyond that 99% (or 95%) were to occur?

24
 CVaR or ES
For example, assuming X=99%, T=10 days, VaR is INR 10 Mn.

Then, ES is the average (or expected loss) over a 10-day period assuming that the
loss is greater than INR 10 Mn.

25
 CVaR or ES
• In that case, what is the expected value of loss? The objective here is to compute
the expected (mean or average) losses for that extreme 1% (or 5%) scenario.

5%

1.645 Std Dev

-10MM
Possible Profit/Loss

26
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Conditional VaR (CVaR) or

Expected Shortfall (ES): Part 2
 CVaR or ES
• In case, empirical data is not available sufficiently, then one has to use some
distributional assumption and integrate over the tail region.

𝑪𝑽𝒂𝑹𝜶 = 𝑬[𝑿|𝑿 > 𝑽𝒂𝑹𝜶 𝑿 ]

𝟏
𝑪𝑽𝒂𝑹𝜶 (𝑿) = ‫𝟏׬‬−𝜶 𝑿 ∗ 𝒅𝑭𝜶 (𝑿)

• Setting limits over ‘1 − 𝛼’ to ‘1’ same as setting limits over VaR to ∞, both
represent the same region.

28
 CVaR or ES
∞
𝑪𝑽𝒂𝑹α = 𝑬[𝑿|𝑿 > 𝑽𝒂𝑹α 𝑿 ] = ‫𝑿 𝑹𝒂𝑽׬‬ ∗ 𝒅𝑭α (𝑿)

Compare & Zio (2015). Genetic

algorithms for condition-based
maintenance optimization under
uncertainty. European Journal of
Operational Research

29
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Conditional VaR (CVaR) or

Expected Shortfall (ES): Part 3
 CVaR or ES
• All observations are equally likely: In the later case with 1001 (0-1000) return observations,

the observation at position 990 will represent the cut-off 1 percentile. This level of returns

(990th position) will become your 99% VaR (daily). The observations beyond this (991-1000)

are our tail losses given that VaR is breached. Thus, we will take an average of observations

from 991 to 1000 position (these are our extreme scenarios available from empirical data)

and take the average to compute CVaR or expected shortfall.

31
 CVaR or ES
Given probability case: If discrete probabilities and corresponding observations are available

for such tail events, for example, if in the tail region, ‘n’ possible scenarios X= (x1, x2, x3, x4,

x5…xn) are given with probabilities P= (p1, p2, p3, p4, p5….pn). Then CVaR (or ES)=

q1x1+q2x2+q3x3+q4x4+q5x5+…..+qnxn, with

𝑝𝑖
𝑞𝑖 = σ𝑛 ; this is to ensure that σ𝑛𝑖=1 𝑞𝑖 =1 (Given that VaR is breached these ‘n’ tail events
𝑖=1 𝑝𝑖

are the only possibilities left

32
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Conditional VaR (CVaR) or

Expected Shortfall (ES):
Simple example 1
 Examples
A given portfolio has 97.5% VaR of INR 1 Mn, 98% VaR of INR 1 Mn,

and a 2% chance that the loss will be INR 10 Mn. What is CVaR (or ES)

at 97.5% confidence

We need to compute the expected loss for the extreme 2.5% given that

97.5% VaR has been breached.

34
 Examples
A given portfolio has 97.5% VaR of INR 1 Mn, 98% VaR of INR 1 Mn, and a 2%

chance that the loss will be INR 10 Mn. What is CVaR (or ES) at 97.5% confidence

We need to compute the expected loss for the extreme 2.5% given that 97.5% VaR

has been breached.

0.5 2
σ 𝑃𝑖 𝑋𝑖 = ∗1+ ∗ 10 = INR 8.2 Mn
2.5 2.5

35
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Conditional VaR (CVaR) or

Expected Shortfall (ES):
Simple example 2
 Examples
A portfolio has 95% VaR of INR 1 Mn, there is a 3% probability that a loss of

INR 2 Mn, 1% Probability that a loss of INR 5 Mn, 0.75% probability that a

loss of INR 10 Mn, and 0.25% probability that a loss of INR 20 Mn may occur.

What is 95% ES

37
 Examples
A portfolio has 95% VaR of INR 1 Mn, there is a 3% probability that a loss of INR 2

Mn, 1% Probability that a loss of INR 5 Mn, 0.75% probability that a loss of INR 10

Mn, and 0.25% probability that a loss of INR 20 Mn may occur. What is 95% ES

3 1 0.75 0.25
σ 𝑃𝑖 𝑋𝑖 = ∗ 2 + ∗5+ ∗ 10 + ∗ 20 = INR 4.7 Mn
5 5 5 5

38
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Summary and Concluding Remarks

 Summary and Concluding Remarks
• Conventional risk models (ARCH, GARCH, etc.) do not provide sufficient
emphasis on the negative tail of the return distribution
• This inefficiency with the conventional risk models is overcome with tail risk
measures such as Value-at-Risk (VaR) and Conditional Value at Risk (CVaR)
measures
• To begin with Value at Risk models, these models estimate the maximum loss
value expected over a given horizon (T) with a certain confidence (X%)
• While this method tells you how bad things can get, but in case things do get
bad, what may happen to our investment portfolio, is not known

40
 Summary and Concluding Remarks
• Fund managers may set up positions with a lower maximum loss value with a
certain confidence level; however, these positions may carry extreme tail
losses outside the confidence band; such inefficiency in the portfolio position
design will not be identified by the VaR models
• To account for this inefficiency of the VaR model, we go to conditional VaR or
expected shortfall models
• These models examine those losses that exceed VaR; or to put it another
way, what is the expected loss, given that losses exceed VaR

41
Thanks!