7.4 Value at Risk and Stress Testing
One number for a bad day: Value at Risk by historical simulation, by the normal (variance–covariance) formula and by Monte Carlo; portfolio VaR with correlations; expected shortfall for the losses beyond VaR; checking VaR against reality; stress scenarios; and everything VaR cannot see.
📍 The situation
Priya, a 34-year-old software engineer in Bengaluru, holds ₹50 lakh in two places: ₹30 lakh in a Nifty 50 index fund and ₹20 lakh in shares of a hypothetical private bank, Alpha Bank Ltd. Her broker's portfolio screen shows a line she has never understood: "1-day VaR (99%): ₹1.44 lakh". Her colleague says it means "the most you can lose tomorrow". It does not.
Value at Risk (VaR) is the risk number used by banks, brokers, fund houses and clearing corporations around the world — for internal limits, for reporting to boards and regulators, and (in India) as one ingredient of the margins charged on equity trades. This section answers Priya's questions. What exactly does a VaR statement say, and what does it leave out? How is it computed — from history, from a formula, or by simulation? Why does diversification lower it, and why can a rise in correlation wipe that benefit out? How bad are the days beyond VaR (expected shortfall)? And how do risk managers check a VaR model and stress-test a portfolio for events that the model has never seen? The mathematics is the volatility, correlation and Monte Carlo simulation of Chapter 3 and the portfolio variance $\mathbf w^{\mathsf T}\Sigma\mathbf w$ of §4.1, turned into rupees.
All volatilities, correlations, return histories and portfolios on this page are simulated or hypothetical, chosen to be realistic in size; they are not estimates for any real index, fund or company, and nothing here is investment advice. Regulatory uses of VaR — bank capital rules, exchange margin methods, mutual-fund risk disclosures — are described only in general terms; they are revised from time to time, so check the current RBI, SEBI and exchange documents before relying on any detail.
What VaR Measures, and Historical Simulation
Every risk measure starts from the distribution of tomorrow's profit or loss. VaR picks one point of it: a loss that is exceeded only rarely. "Rarely" is fixed by a confidence level (95%, 99%) and "tomorrow" by a horizon (one day, ten days).
Let $V$ be today's portfolio value and $\Delta V$ its change over a horizon of $h$ days; $\Delta V$ is the profit and loss (P&L) and $L=-\Delta V$ the loss (negative when there is a profit). For a confidence level $\alpha$ (such as $0.99$), the Value at Risk is the smallest number $\ell$ such that the loss exceeds $\ell$ with probability at most $1-\alpha$: $$\text{VaR}_\alpha=\min\{\ell:\ P(L\gt\ell)\le1-\alpha\}.$$ In words: with probability $\alpha$ the loss over the next $h$ days will not exceed $\text{VaR}_\alpha$; on a fraction $1-\alpha$ of occasions it will — by an amount VaR does not state. VaR is the $\alpha$-quantile of the loss distribution, reported as a positive rupee amount (or as a percentage of $V$).
So Priya's "1-day 99% VaR of ₹1.44 lakh" means: on about 99 trading days in 100 her portfolio should lose less than ₹1.44 lakh in a day; on about 1 day in 100 — two or three days a year — it should lose more, and the statement says nothing about how much more. It is a threshold for a bad day, not a worst case.
Historical simulation is the most direct way to get a loss distribution: ask what today's portfolio would have lost on each of the last $N$ days, and treat those $N$ outcomes as equally likely scenarios for tomorrow. It needs no assumption about normality — whatever fat tails and skewness were in the data are carried over automatically.
Let $r_{i,t}$ be the return of asset $i$ on past day $t=1,\dots,N$ and $V_i$ today's holding. Scenario $t$ gives the loss $$L_t=-\sum_i V_i\,r_{i,t}.$$ Sort the $N$ scenario losses from worst to best and let $k=\lfloor N(1-\alpha)\rfloor+1$. The historical VaR is the $k$-th worst loss; the historical expected shortfall (Definition 7.20) is the average of the $k$ worst losses.
With $N=250$ days: the 99% VaR is the 3rd worst loss ($k=\lfloor2.5\rfloor+1$), the 97.5% VaR the 7th worst, the 95% VaR the 13th worst. (Conventions differ by one observation, and spreadsheets' PERCENTILE functions interpolate, so published figures can differ slightly; what matters is to state and keep one rule.)
- Put the last $N+1$ daily closes (or NAVs) of each holding in columns and compute daily returns.
- Scenario loss for each day: $-\sum_i V_i r_{i,t}$ with today's rupee holdings $V_i$, e.g.
=-($H$1*B3+$H$2*C3). - VaR:
=LARGE(D3:D252,k)with $k=\lfloor N(1-\alpha)\rfloor+1$; ES:=AVERAGE(LARGE(D3:D252,{1,2,3}))for $k=3$. - For a 10-day horizon either use overlapping 10-day returns or scale by $\sqrt{10}$ (Theorem 7.19) — and say which.
Parametric VaR and Portfolio VaR with Correlations
The parametric or variance–covariance method assumes the P&L is normally distributed. Then the whole distribution is fixed by its mean and standard deviation, and every quantile is a multiple of the standard deviation — the same idea as the confidence bands of §3.3.
If the portfolio's return over the horizon is normal with mean $\mu$ and standard deviation $\sigma$, then $$\text{VaR}_\alpha=(z_\alpha\,\sigma-\mu)\,V,$$ where $z_\alpha$ is the standard normal quantile: $z_{0.95}=1.645$, $z_{0.975}=1.960$, $z_{0.99}=2.326$. For daily horizons $\mu$ is tiny and usually set to $0$: $\text{VaR}_\alpha=z_\alpha\sigma V$. If daily returns are independent with daily volatility $\sigma_1$, the $h$-day volatility is $\sigma_1\sqrt h$ (§3.1, square-root-of-time rule), so $$\text{VaR}_\alpha(h\text{ days})=\sqrt h\;\text{VaR}_\alpha(1\text{ day})\qquad(\mu=0).$$
Why. The loss is $L=-RV$ with $R\sim N(\mu,\sigma^2)$. Then $P(L\gt\ell)=P\!\left(R\lt-\ell/V\right)=\Phi\!\left(\frac{-\ell/V-\mu}{\sigma}\right)$. Setting this equal to $1-\alpha$ and using $\Phi^{-1}(1-\alpha)=-z_\alpha$ gives $-\ell/V-\mu=-z_\alpha\sigma$, i.e. $\ell=(z_\alpha\sigma-\mu)V$. Over $h$ independent days the variances add, so $\sigma$ becomes $\sigma_1\sqrt h$. ∎
For a portfolio, the standard deviation is not the sum of the holdings' standard deviations: it depends on how they move together. This is exactly the portfolio variance of §4.1, with rupee amounts in place of weights.
Let holding $i$ have value $V_i$, return volatility $\sigma_i$ and rupee volatility $a_i=V_i\sigma_i$, and let $\rho_{ij}$ be the correlations (matrix $C$). Under the normal model with zero mean:
- $\sigma_P^2=\sum_i\sum_j a_i a_j\rho_{ij}=\mathbf a^{\mathsf T}C\,\mathbf a$ and $\text{VaR}_P=z_\alpha\sigma_P$.
- For two holdings, with standalone $\text{VaR}_i=z_\alpha a_i$: $$\text{VaR}_P=\sqrt{\text{VaR}_1^2+\text{VaR}_2^2+2\rho\,\text{VaR}_1\text{VaR}_2}\ \le\ \text{VaR}_1+\text{VaR}_2,$$ with equality only when $\rho=1$. The gap is the diversification benefit.
- The component VaR of holding $i$, $\ \text{CVaR}_i=z_\alpha\dfrac{a_i\,(C\mathbf a)_i}{\sigma_P}$, adds up exactly to $\text{VaR}_P$ and shows which holding the risk comes from (the risk contributions of §4.1, in rupees).
Why. The rupee P&L is $\sum_iV_iR_i$, a sum of correlated normal variables, so it is normal with variance $\sum_{i,j}V_iV_j\operatorname{Cov}(R_i,R_j)=\sum_{i,j}a_ia_j\rho_{ij}$ (§3.2, variance of a sum). Multiply by $z_\alpha^2$ for (2); since $\rho\le1$, the expression under the root is at most $(\text{VaR}_1+\text{VaR}_2)^2$. For (3), $\sum_ia_i(C\mathbf a)_i=\mathbf a^{\mathsf T}C\mathbf a=\sigma_P^2$, so the components sum to $z_\alpha\sigma_P$. ∎
| $\rho$ | 0 | 0.6 | 0.9 | 1 |
|---|---|---|---|---|
| 1-day 99% VaR (₹) | 1,13,611 | 1,43,606 | 1,56,461 | 1,60,518 |
Monte Carlo VaR and Expected Shortfall
The normal formula breaks down when the P&L is not a simple sum of normal returns — a portfolio with options (whose value is a curved function of the index, §6.3), or returns with fat tails. Monte Carlo VaR keeps a model for the risk factors but replaces the formula by simulation, exactly as §3.3 simulated price paths.
Choose a joint model for tomorrow's risk-factor returns (volatilities, correlations, and the shape of the distribution — normal, or fat-tailed). Simulate $N$ independent scenarios from it, revalue the whole portfolio in each scenario (fully, including any options), and compute the scenario losses. The Monte Carlo VaR and expected shortfall are the historical-simulation statistics of Definition 7.18 applied to these $N$ simulated losses. Its simulation error shrinks like $1/\sqrt N$ (§3.3), and tail quantiles need many scenarios.
- Draw two independent standard normals $z_1,z_2$ (Box–Muller, §3.3, or
NORM.S.INV(RAND())). - Make them correlated (the two-dimensional Cholesky step): $x_1=z_1$, $\ x_2=\rho z_1+\sqrt{1-\rho^2}\,z_2$. Then $x_1,x_2$ are standard normal with correlation $\rho$.
- Returns $R_i=\sigma_ix_i$ (for fat tails, divide both by a common random factor, e.g. a Student-$t$ mixing variable); P&L $=V_1R_1+V_2R_2$, or the full revaluation of options at the new prices.
- Repeat $N$ times, sort the losses, take the $k$-th worst ($k=\lfloor N(1-\alpha)\rfloor+1$) as VaR and the average of the $k$ worst as ES.
Why step 2 works. $\operatorname{Var}(x_2)=\rho^2+(1-\rho^2)=1$ and $\operatorname{Cov}(x_1,x_2)=\operatorname{Cov}(z_1,\rho z_1)=\rho$, because $z_1$ and $z_2$ are independent. For many assets the same idea uses the Cholesky factor $L$ of the correlation matrix ($C=LL^{\mathsf T}$, a linear-algebra tool): $\mathbf x=L\mathbf z$. ∎
| # | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 |
|---|---|---|---|---|---|---|---|---|---|---|
| $x_2$ | −0.008 | −1.830 | 1.026 | −0.104 | 0.180 | −0.966 | −1.118 | −1.450 | 1.232 | 0.442 |
| P&L | 16,872 | −1,26,930 | 40,566 | −24,864 | 49,380 | −1,14,306 | −31,338 | −83,550 | 1,01,112 | 5,022 |
VaR answers "how bad is a 1-in-100 day?" but not "when such a day comes, how bad is it on average?". Two portfolios can have the same VaR and very different tails. Expected shortfall looks beyond the threshold.
The expected shortfall (ES, also called conditional VaR or tail VaR) at level $\alpha$ is the average loss on the occasions when the loss reaches or exceeds VaR: $$\text{ES}_\alpha=E\left[L\,\middle|\,L\ge\text{VaR}_\alpha\right].$$ For a continuous loss distribution this is the average of the worst $(1-\alpha)$ fraction of outcomes; always $\text{ES}_\alpha\ge\text{VaR}_\alpha$. (For discrete distributions refined definitions average exactly the worst $1-\alpha$ of probability; for the continuous models and large samples used here the two agree.)
- If the loss is normal with mean $0$ and standard deviation $\sigma_P$ (rupees), then $$\text{ES}_\alpha=\frac{\varphi(z_\alpha)}{1-\alpha}\,\sigma_P,$$ where $\varphi$ is the standard normal density. The multipliers are $2.063$ at 95%, $2.338$ at 97.5% and $2.665$ at 99%; in particular the 97.5% ES ≈ the 99% VaR ($2.338$ against $2.326$).
- For a fat-tailed Student-$t$ distribution with 4 degrees of freedom, scaled to the same $\sigma_P$, the multipliers become: VaR $1.507,\ 1.963,\ 2.649$ and ES $2.265,\ 2.824,\ 3.692$ at 95%, 97.5%, 99%. Fat tails raise the far quantiles and raise ES even more.
- Expected shortfall is subadditive: $\text{ES}(A+B)\le\text{ES}(A)+\text{ES}(B)$ for any two portfolios. VaR is not, in general (Practice corner E6).
Why (1). With $L=\sigma_PZ$, $Z\sim N(0,1)$: $$E[Z\mid Z\ge z_\alpha]=\frac{1}{1-\alpha}\int_{z_\alpha}^{\infty}z\,\varphi(z)\,dz=\frac{\big[-\varphi(z)\big]_{z_\alpha}^{\infty}}{1-\alpha}=\frac{\varphi(z_\alpha)}{1-\alpha},$$ because $\frac{d}{dz}\varphi(z)=-z\varphi(z)$ — an integral that can be done exactly. The Student-$t$ numbers come from the same integral with the $t$ density (computed numerically). Subadditivity is a known theorem about averages of tails, quoted without proof. ∎ Because ES "sees" the whole tail and adds up sensibly across desks, international bank-capital rules have moved from 99% VaR towards 97.5% ES for market risk; VaR remains widespread because it is simpler to estimate and to check.
Checking VaR, Stress Tests and the Limits of VaR
A VaR model makes a testable promise: losses beyond the 99% VaR should happen on about 1% of days, scattered at random. Counting those days — backtesting the VaR — is the first check every risk team runs (the same word as in §7.3, now applied to a risk model rather than a trading rule).
An exception (breach) is a day on which the actual loss exceeds the VaR that was computed for that day in advance. Backtesting a VaR model compares the number of exceptions over a period — and how they are spread in time — with what the confidence level implies.
If a VaR model at level $\alpha$ is correct and days are independent, the number $X$ of exceptions in $n$ days is binomial $B(n,\,p)$ with $p=1-\alpha$: $$E[X]=np,\qquad \text{SD}(X)=\sqrt{np(1-p)},\qquad P(X\ge k)=\sum_{j\ge k}\binom nj p^j(1-p)^{n-j}.$$ For $n=250$ and $\alpha=99\%$: $E[X]=2.5$, $\text{SD}\approx1.57$, $P(X\ge5)\approx10.8\%$, $P(X\ge7)\approx1.4\%$, $P(X\ge10)\approx0.03\%$. Bank supervisors have long used a "traffic-light" reading of these counts — green for 0–4 exceptions, amber for 5–9, red for 10 or more in 250 days. Exceptions that come in clusters are a second warning sign: a correct model's exceptions are independent.
Why. Each day is an independent trial with "success" (an exception) probability $1-\alpha$ — that is what the VaR promises — so the count is binomial and the mean and variance are $np$ and $np(1-p)$. ∎ Note what the test can and cannot do: with only 2.5 expected exceptions a year, a model that is somewhat too optimistic is hard to catch in one year, while a model that is far off (or slow to react, Interactive 2) shows up quickly.
Backtesting checks VaR against ordinary bad days. It cannot tell you what happens in a crisis the data have never contained. For that, risk managers imagine the crisis and revalue the portfolio in it.
- A stress test (scenario analysis) specifies large, joint moves of the risk factors — "the index falls 15%, bank shares 25%, gold rises 6%, interest rates rise 1%" — and computes the portfolio's loss under them, without attaching a probability.
- Scenarios may be hypothetical (designed by judgement), historical (replaying the factor moves of a past crisis on today's portfolio), or sensitivity tests (one factor at a time, or stressed volatilities and correlations).
- A reverse stress test starts from an unacceptable loss and asks which scenario would cause it.
Stress losses are computed with the simplest tools of the course: a share's move is its value times the shock; an index-linked holding moves by about its beta times the index move (§3.2); a debt fund's NAV moves by about $-D_{\text{mod}}\times\Delta y$ for a change $\Delta y$ in yields (§2.3, modified duration).
- "VaR is the most I can lose." It is a loss exceeded on $1-\alpha$ of days — 2 or 3 days a year at 99% — and says nothing about how far beyond it those losses go (Example 1: VaR ₹6 lakh, worst case ₹15 lakh).
- Trusting the normal formula in the tail. With fat tails the 99% VaR and especially the ES are larger than $2.33\sigma$ suggests (Examples 2 and 6).
- Calibrating on calm data. A VaR estimated from a quiet year is low just before the storm, and exceptions then come in clusters (Interactive 2).
- Adding standalone VaRs, or assuming today's correlations hold. Portfolio VaR needs correlations (Theorem 7.20), and correlations tend to jump towards 1 in sell-offs (§3.2), wiping out the diversification benefit.
- Scaling by $\sqrt{h}$ blindly. The rule assumes independent days, a constant portfolio and the ability to exit; trends in a crisis, positions that cannot be sold quickly, and options make the 10-day VaR larger.
- "VaR always rewards diversification." VaR is not subadditive: combining two positions can raise VaR above the sum (Practice corner E6). Expected shortfall does not have this flaw.
- Too few scenarios. A 99% VaR from 250 days rests on 2–3 observations; Monte Carlo with a few hundred paths is just as noisy (Example 5).
- Treating a VaR as a measurement rather than a model output. Different methods and windows give different numbers for the same portfolio; the method should be stated with the figure.
- "A stress test is a forecast." It is a "what if": it has no probability, and its value depends on imagining the right scenario — reverse stress tests help.
- "Low VaR means a safe investment." A one-day VaR says nothing about long-run losses from inflation, fees, concentration in one company or a slow bear market (§3.1, drawdowns).
The Menon family's (hypothetical) ₹1 crore is spread over four holdings. Illustrative daily volatilities and correlations, estimated from simulated history:
| holding | value | daily σ | ₹ σ a day | correlations (index, bank, gold, debt) | |||
|---|---|---|---|---|---|---|---|
| Nifty 50 index fund | ₹40 lakh | 1.1% | 44,000 | 1 | 0.6 | −0.1 | 0 |
| Alpha Bank shares | ₹15 lakh | 1.8% | 27,000 | 0.6 | 1 | −0.1 | 0 |
| gold ETF | ₹15 lakh | 0.9% | 13,500 | −0.1 | −0.1 | 1 | 0.1 |
| short-duration debt fund | ₹30 lakh | 0.1% | 3,000 | 0 | 0 | 0.1 | 1 |
- Portfolio volatility (Theorem 7.20): $\sigma_P=\sqrt{\mathbf a^{\mathsf T}C\mathbf a}\approx$ ₹64,018 a day (0.64% of ₹1 crore).
- VaR. 1-day 99%: $2.326\times64{,}018\approx$ ₹1,48,927; 10-day: $\times\sqrt{10}\approx$ ₹4,70,949. The standalone VaRs add to ₹2,03,555, so diversification saves ₹54,628.
- Where the risk is (component VaR): index fund 63.2%, bank shares 34.3%, gold 2.2%, debt fund 0.3%. The bank shares are 15% of the money but a third of the risk; gold, slightly negatively correlated, adds almost nothing.
- Beyond VaR. Normal 99% ES: $2.665\times64{,}018\approx$ ₹1,70,621. With Student-$t_4$ tails and the same volatility: VaR $\approx$ ₹1,69,583, ES $\approx$ ₹2,36,353.
- Correlation and volatility stress. If the index–bank correlation rises to 0.9 and gold turns positively correlated (0.3) with both, the 1-day 99% VaR rises to ₹1,73,743; if the two equity volatilities also rise 2.5-fold, to ₹4,13,922 — almost three times the calm figure.
- Scenario. Index −15%, Alpha Bank −25%, gold +6%, yields +1% with the debt fund's modified duration 1.5 (NAV −1.5%): P&L $=-6{,}00{,}000-3{,}75{,}000+90{,}000-45{,}000=$ −₹9,30,000, about 6 times the 1-day VaR and twice the 10-day VaR.
- Reverse stress. If the bank shares fall 1.5 times as much as the index and gold and debt do not move, an index fall of $x$ costs $40x+15(1.5x)=62.5x$ lakh; a ₹10 lakh loss needs $x=16\%$.
What the family learns. On an ordinary bad day (1 in 100) they should expect to lose around ₹1.5 lakh, and around ₹1.7–2.4 lakh on the average such day, depending on how fat the tails are. A crisis scenario costs about ₹9 lakh — nearly 10% of their wealth — and almost all of it comes from the two equity holdings, about 40% from a single bank. VaR, ES and the stress test are three different questions; a sensible risk report shows all three, states its assumptions, and does not pretend that any of them is the worst case.
Exercises
All volatilities, correlations and returns are illustrative. Use $z_{0.95}=1.645$, $z_{0.975}=1.960$, $z_{0.99}=2.326$; zero mean daily returns unless stated; and the historical convention of Definition 7.18.
- 1 Tomorrow's P&L (₹ thousand) is +20 (prob. 0.30), +5 (0.30), −5 (0.20), −15 (0.12), −30 (0.05), −60 (0.03). Find the 90%, 95% and 99% VaR and the 95% expected shortfall.
- 2 With 500 days of history, which ordered loss is the 99% historical VaR? The 95% VaR?
- 3 On a ₹20 lakh portfolio, the 8 worst of 250 daily returns are (%): −5.2, −4.1, −3.3, −3.0, −2.8, −2.6, −2.5, −2.4. Find the 99% and 97.5% historical VaR and ES.
- 4 Explain in one sentence what "1-day 95% VaR of ₹2 lakh" means. On how many of 252 trading days should the loss exceed ₹2 lakh?
- 5 Why can a historical VaR estimated from a calm year be dangerously low? Suggest one remedy.
- 6 A 1-day 99% VaR is ₹1.5 lakh. Find the 10-day VaR by the square-root rule, and give two reasons it may understate the true 10-day risk.
- 7 Under the normal model with zero mean, a portfolio's 1-day 99% VaR is ₹1,00,000. Find its 1-day 95% VaR.
- 1 ₹10 lakh in a share with daily volatility 2%. Find the 1-day 99% and 95% VaR.
- 2 A fund's annual volatility is 20%. Find its daily volatility (252 days) and the 1-day 99% VaR of ₹25 lakh in it.
- 3 ₹6 lakh in asset A (daily σ 1.5%) and ₹4 lakh in B (daily σ 2.0%), correlation 0.3. Find the 1-day 99% portfolio VaR, the sum of the standalone VaRs and the diversification benefit.
- 4 Repeat part 3 with correlation −0.5 and with correlation 1.
- 5 In part 1, include a daily mean return of 0.05%. How much does the 99% VaR change?
- 6 Find the component VaRs of A and B in part 3 and their shares of the total.
- 7 ₹10 lakh of shares (daily σ 1.2%) is hedged by selling index futures of notional $H$ (index daily σ 1.0%, correlation 0.9). Find the $H$ that minimises the portfolio variance, and the 1-day 99% VaR with and without the hedge. (This is a calculus minimisation; compare §6.1 on hedging with futures.)
- 1 A portfolio's daily P&L is normal with mean 0 and standard deviation ₹50,000. Find its 99% VaR and 99% ES.
- 2 Show numerically that for a normal loss the 97.5% ES is almost equal to the 99% VaR.
- 3 With $\rho=0.8$, independent standard normals $z_1=1.2$, $z_2=-0.5$ are turned into correlated $x_1,x_2$. Find $x_2$.
- 4 In 1,000 Monte Carlo scenarios the 11th worst loss is ₹1,32,000 and the average of the 11 worst is ₹1,68,000. Give the 99% VaR and ES. How would you check whether 1,000 scenarios are enough?
- 5 A 99% VaR model had 8 exceptions in 250 days. Find $P(X\ge8)$ under a correct model and the traffic-light zone.
- 6 A 95% VaR had 38 exceptions in 500 days. Find the expected number, the standard deviation and the z-score. Verdict?
- 7 A portfolio: ₹20 lakh index fund, ₹10 lakh bank shares that move 1.6 times the index, ₹10 lakh debt fund with modified duration 3. Find the loss if the index falls 10% and yields rise 0.5%.
- 8 Two independent loans each lose ₹1 lakh with probability 3% and nothing otherwise. Find the 95% VaR of each loan and of the two together. What does this show?
E1. NISM-style A ₹1 crore portfolio has daily return volatility 1% and normally distributed returns with zero mean. Its 1-day 99% VaR is closest to (A) ₹1.00 lakh (B) ₹1.64 lakh (C) ₹2.33 lakh (D) ₹23.3 lakh.
E2. NISM-style A 1-day VaR is ₹2 lakh. Assuming independent daily returns with zero mean, the 25-day VaR at the same confidence is (A) ₹50 lakh (B) ₹10 lakh (C) ₹5 lakh (D) ₹2 lakh.
E3. NISM-style Which statement about the 99% expected shortfall of a portfolio is correct? (A) It is always smaller than the 99% VaR (B) It is the average loss on days when the loss reaches or exceeds the 99% VaR (C) It is the largest possible loss (D) It ignores losses larger than VaR.
E4. Case study An investor holds ₹40 lakh in an equity fund (daily σ 1.2%) and ₹20 lakh in a gold ETF (daily σ 0.9%), correlation −0.1. (a) Find the 1-day and 10-day 99% VaR and the diversification benefit. (b) Stress scenario: equity −20%, gold +8%. Find the P&L and compare with (a).
E5. Case study A 99% VaR model is backtested over 500 days: 11 exceptions, of which 7 fall in one 30-day period. (a) Find the expected number and $P(X\ge11)$ under a correct model. (b) The worst 250-day window contains 9 exceptions: which traffic-light zone? (c) What does the clustering suggest, and what would you change?
E6. Interview Two independent bonds each default with probability 4%, losing ₹100 on default and nothing otherwise. What is the 95% VaR of each bond, and of a portfolio holding both? What property of a risk measure does this violate, and which measure fixes it?
Suggested plan — 2 lectures × 60 min.
- Three ways to compute VaR. Open with Priya's broker screen and ask the class what "1-day 99% VaR ₹1.44 lakh" means; collect answers ("the maximum loss" will appear). Definition 7.17 with Example 1; historical simulation (Definition 7.18) and Example 2 — point out the fat-tail pattern. Normal VaR (Theorem 7.19), Example 3; portfolio VaR with correlations (Theorem 7.20) and Example 4's correlation table. Project Interactive 1: press Play and watch the diversification benefit shrink as ρ rises; then switch to fat tails at 99%.
- Beyond VaR. Monte Carlo (Definition 7.19) with Example 5 done by the class in pairs; expected shortfall (Definition 7.20, Theorem 7.21) and Example 6. Backtesting (Theorem 7.22), Example 7, and Interactive 2 with "calm, then a sudden crisis": compare the historical and EWMA methods and count the clustered exceptions. Stress tests (Definition 7.22), Example 8, the 📈 Market case and Practice corner E6.
Class activity. Groups each receive a different hypothetical portfolio (weights in an index fund, a single share, gold and a debt fund) with a volatility/correlation table. Each computes the 1-day 99% VaR, the component VaRs and one stress scenario of their own design, then presents "the one number" and "the number VaR doesn't show".
Misconceptions to watch for. VaR as maximum loss; adding VaRs; assuming normal tails; square-root scaling as a law; "more exceptions than expected" as bad luck; stress tests as forecasts; ES as "the same as VaR".
Suggested homework. I(1, 3, 6), II(3, 4, 6, 7), III(1, 4, 5, 7, 8); Practice corner E4, E5.
Project. Build a one-page risk report for a two- or three-fund portfolio: historical, parametric and Monte Carlo VaR, expected shortfall, a VaR backtest and one stress scenario.
Why it matters. Computing the same "risk" three ways and getting three numbers is the fastest way to understand that VaR is a model output with assumptions — and to read any risk report critically.
What to do.
- Data. Download two to three years of daily NAVs of two or three mutual funds of different types from your fund house's or AMFI's website (or simulate them as in §3.3). Compute daily returns in columns B, C, D.
- Holdings. Put hypothetical rupee amounts in H1:H3. Scenario P&L in E:
=$H$1*B3+$H$2*C3+$H$3*D3. - Historical VaR and ES on the last 250 rows:
=-SMALL(E_last250,3)for 99% and=-AVERAGE(SMALL(E_last250,{1,2,3}))for ES; for 95% use 13. - Parametric VaR: build the covariance matrix with
=COVARIANCE.S(B:B,C:C)etc., then $\sigma_P$ with=SQRT(MMULT(TRANSPOSE(H1:H3),MMULT(Cov,H1:H3)))and VaR $=2.326\sigma_P$. Compute the component VaRs. - Monte Carlo in Python:
import numpy as np; R=np.loadtxt('returns.csv',delimiter=','); V=np.array([...]); S=np.cov(R.T); rng=np.random.default_rng(1); X=rng.multivariate_normal(np.zeros(len(V)),S,100000); L=-(X@V); q=np.quantile(L,0.99); print(q, L[L>=q].mean()). Repeat with Student-$t$ shocks: divide each row bynp.sqrt(rng.chisquare(4,100000)/2)[:,None](this keeps the variance). - Backtest: for each day after the first 250, compute the historical VaR from the previous 250 days and mark an exception if that day's loss is larger. Count them and compute $P(X\ge k)$ with
=1-BINOM.DIST(k-1,n,0.01,TRUE). - Stress: write one hypothetical scenario (equity −15%, debt yields +1% with the fund's stated modified duration, gold +5%) and compute its P&L.
Hand in: a table with the three VaRs and two ESs, the component-VaR shares, the backtest count with its p-value, the stress loss, and a paragraph explaining why the numbers differ.
Going further. Replace the equally weighted volatility by EWMA ($\sigma_t^2=\lambda\sigma_{t-1}^2+(1-\lambda)r_{t-1}^2$, $\lambda=0.94$) and compare the backtests; add a put option on the index fund to the portfolio and see why only Monte Carlo with full revaluation handles it (§6.3 for the option price).
🧪 Self-check quiz
Q1. "1-day 99% VaR = ₹1 lakh" means
Q2. Two positions each have a 1-day 99% VaR of ₹1 lakh and their returns are uncorrelated (normal model). The VaR of the combined position is about
Q3. For a normally distributed loss, the 99% expected shortfall is about how many times the 99% VaR?
Q4. A correct 1-day 99% VaR model is backtested over 250 trading days. The expected number of exceptions is
Q5. Which is the main purpose of a stress test?