Portfolio and risk · Measuring return and risk
Value at risk (VaR): the loss a portfolio should rarely exceed
What a 95% value at risk means, three ways to calculate it, a test of SPY’s VaR on the following year, and why expected shortfall answers the question VaR leaves open.
On this page
Value at risk, or VaR, is a loss threshold with a probability attached. A one-day 95% VaR of $1,600 on a $100,000 portfolio means: on 95 days out of 100, the portfolio should lose less than $1,600, and on the other five it should lose more. It compresses the left side of a return distribution into one sentence a non-specialist can act on.
Banks adopted VaR in the 1990s because it put the risk of very different holdings (stocks, bonds, currencies, options) into a single dollar figure. It has been criticized ever since, because of what the sentence leaves out: how much more is lost on those other five days. This lesson explains what VaR claims, calculates it three ways on SPY, tests it against the following year, and introduces expected shortfall, the measure that regulators have moved to.
The idea
Sort every daily return of a portfolio from worst to best and count in from the bad end. The return that has 5% of the days below it and 95% above it is the 95% VaR. That is all the historical method is: a percentile.
In the figure the 25 worst days of 500 lie to the left of the 95% line. VaR reports where the line is. It says nothing about the shape of the red area beyond it: whether those days cluster just past the line or stretch far to the left. Here they stretch. The worst day was almost four times the VaR.
The same statement can be made about any horizon. A monthly VaR uses monthly returns; a ten-day VaR uses ten-day returns. The confidence level and the horizon are choices, and a VaR quoted without both is meaningless.
How it works
There are three standard ways to produce the number.
- Historical assumes the future will be drawn from the same set of days as the sample. It needs no assumption about the shape of returns.
- Parametric (also called variance-covariance) assumes returns follow a normal distribution, so that the average and standard deviation are enough. It is quick and it extends easily to many holdings using their correlations.
- Simulated (Monte Carlo) generates thousands of possible outcomes from a model. Monte Carlo simulation: testing a plan against thousands of futures covers the technique. Its result is only as good as the model.
A worked example on SPY
The data: 500 daily returns of SPY from closing prices, November 19, 2024 to November 17, 2026. The average daily return was 0.050% and the standard deviation 1.066%.
Historical. Five percent of 500 days is 25 days. The return at that boundary is −1.57%. One percent is 5 days, and the boundary is −2.71%.
Parametric. 95%: 0.050% − 1.645 × 1.066% = −1.70%. 99%: 0.050% − 2.326 × 1.066% = −2.43%.
| On a $100,000 portfolio of SPY | Historical | Parametric (bell curve) |
|---|---|---|
| One-day 95% VaR | $1,567 | $1,703 |
| One-day 99% VaR | $2,706 | $2,430 |
| Average loss beyond the 95% VaR | $2,456 | |
| Worst day (April 4, 2025) | $5,854 |
The two methods disagree in opposite directions at the two levels. At 95% the bell curve was too cautious: it set the line at −1.70%, and only 19 days were worse, where 25 would be expected. At 99% it was not cautious enough: its line was −2.43%, and 8 days were worse, where 5 would be expected. That is the signature of fat tails. Real returns have more small days and more extreme days than a normal distribution, and fewer in between. The worst day, −5.85%, was 5.5 standard deviations below the average, something the bell curve treats as close to impossible.
Expected shortfall
VaR answers "how bad is a bad day at the threshold?" Expected shortfall answers the natural follow-up: "when the threshold is broken, how bad is it on average?" For SPY, the 25 days beyond the 95% VaR averaged −2.46%, against a VaR of −1.57%.
Expected shortfall is always at least as large as VaR and it responds to the shape of the tail, which VaR ignores. Two portfolios can share a VaR of 1.6% while one has an expected shortfall of 2% and the other 6%. The international standard for banks' market-risk capital (the Basel Committee's 2019 framework) replaced 99% VaR with expected shortfall at 97.5% for that reason.
Other holdings, same two years
| Symbol | 95% one-day VaR | 99% one-day VaR | Expected shortfall (95%) | Worst day |
|---|---|---|---|---|
| TLT | 1.18% | 1.71% | 1.58% | 3.02% |
| KO | 1.56% | 2.46% | 2.14% | 4.44% |
| SPY | 1.57% | 2.71% | 2.46% | 5.85% |
| QQQ | 2.04% | 3.88% | 3.30% | 6.49% |
| GLD | 2.63% | 4.12% | 3.76% | 10.27% |
| NVDA | 4.27% | 7.36% | 6.27% | 16.97% |
All are losses, from daily closing prices without dividends. Coca-Cola and SPY have almost the same 95% VaR. Gold's worst day was nearly four times its 95% VaR.
Longer horizons
A common shortcut scales a one-day VaR by the square root of the number of days: ten-day VaR = one-day VaR × √10. For SPY that gives 1.57% × 3.16 = 4.95%. Measured directly on every ten-day stretch in the sample, the 95% VaR was 3.52%, and the worst ten days lost 13.72%. The shortcut was too high for the threshold and far too low for the extreme, because days are not independent: losses cluster in some weeks and reverse in others.
On monthly data, with a much longer history, SPY's total returns from March 1993 to October 2026 give a one-month 95% VaR of 6.91%, a 99% VaR of 10.44%, an expected shortfall of 9.31% and a worst month of 16.52%. A 60/40 mix of SPY and AGG over 2005 to 2026 had a one-month 95% VaR of 3.96%, an expected shortfall of 5.96% and a worst month of 10.82%.
Using it: testing a VaR on the following year
A VaR is a forecast, and forecasts can be checked. Split the two years in half. Estimate the VaR from the first 250 days, then count how often the next 250 days broke it.
- From the first 250 days (to November 18, 2025): 95% VaR −1.67%, 99% VaR −3.70%.
- In the next 250 days: 8 losses beyond the 95% VaR, where 12.5 were expected, and 1 beyond the 99% VaR, where 2.5 were expected.
The VaR was too conservative, and the reason is visible in the volatility. The first year included April 2025 and had a daily standard deviation of 1.23%. The second was calmer, at 0.87%. A VaR is a rear-view measurement: it reported the risk of the year that had just ended.
That error is harmless in this direction. It is dangerous in the other. A VaR estimated during 2006, a quiet year, would have been broken again and again in 2008. The measure is lowest just before it is needed and highest just after.
Where it fails or misleads
- It is silent about the tail. "Losses should exceed $1,567 on one day in twenty" does not say whether the twentieth day costs $1,600 or $16,000. Expected shortfall and stress tests exist to fill that gap.
- It is not a worst case. A 99% daily VaR should be exceeded two or three times a year. Calling it a "maximum loss" is the most common misreading.
- It depends on the sample. A short or calm window gives a small VaR. A window that happens to contain a crisis gives a large one for exactly as long as the crisis stays in the window, then drops overnight.
- The normal assumption understates extremes. Parametric VaR at 99% was too optimistic on SPY and is worse for single stocks, options and anything with jumps.
- Square-root scaling assumes independent days. Real losses cluster.
- It can discourage diversification. For some lopsided positions, VaR adds up badly: two positions that each show a VaR of zero can show a positive VaR together. Artzner and co-authors' 1999 paper made this point and set out the properties a "coherent" risk measure should have. Expected shortfall meets them and does not have the flaw.
- It measures market moves, not other failures. Being unable to sell, a broker failing, or a margin call that forces a sale at the low are outside it. See Liquidity risk: the cost of selling when you have to.
Compared with its neighbours
| Measure | Statement it makes | Sees the size of extreme losses? |
|---|---|---|
| Volatility | Typical size of a return in either direction | No |
| Value at risk | A loss exceeded with a stated small probability | No: only where the tail begins |
| Expected shortfall | The average loss once that threshold is passed | Yes, on average |
| Maximum drawdown | The worst cumulative loss over many periods | Yes, for one past episode |
| Stress test | The loss in a specified scenario, with no probability | Yes, for the scenario chosen |
How Finabase does it
Finabase does not calculate value at risk or expected shortfall for a symbol or a portfolio, and its screens make no probability statements about losses.
The inputs are available. A symbol's History tab lists daily closing prices that can be
exported; from them the daily percent changes and their 5th percentile are a short
spreadsheet exercise (PERCENTILE.INC(range, 0.05)). A portfolio page shows each position's
market value and today's gain or loss, and its history chart plots total value over a chosen
range. The market heatmap shows the size of today's moves across an index, which gives a
sense of whether today is an ordinary day or a tail day.
For a rough portfolio VaR without software, multiply the portfolio's value by its largest holding's or its benchmark's VaR and treat the result as an order of magnitude. A portfolio that is mostly a stock index will have a one-day 95% VaR in the region of 1.5% to 2% of its value in ordinary markets and a multiple of that in a crisis.
Frequently asked questions
What does a 95% VaR of $5,000 mean?
That, based on the data and method used, the portfolio is expected to lose more than $5,000 in about one period out of twenty and less than that in the other nineteen. It does not say how much more.
Is VaR the most I can lose?
No. It is a threshold that is expected to be broken regularly: about 12 or 13 days a year at 95%, two or three at 99%. SPY's worst day in the sample was 3.7 times its 95% VaR.
What is the difference between VaR and expected shortfall?
VaR is the loss at the edge of the tail. Expected shortfall is the average loss inside the tail. The second is always larger and is the better guide to how bad a bad day is.
Which confidence level should be used?
95% describes a bad but routine period and can be estimated from modest amounts of data. 99% looks further into the tail and needs far more data: with 500 days, it rests on five observations. Higher confidence levels from short samples are guesses.
Why did VaR models fail in 2008?
Mostly because they were estimated from the calm years before, assumed normal returns, and ignored that markets would become illiquid and correlated at once. The measure did what it was built to do with the inputs it was given.
Summary
Value at risk states a loss that should be exceeded only a set fraction of the time, over a stated horizon. For SPY over two years of daily data the 95% figure was 1.57% and the 99% figure 2.71%, by counting; the bell-curve formula gave 1.70% and 2.43% and was wrong in both directions. The average loss beyond the 95% line was 2.46%, and the worst day 5.85%. VaR is a useful summary of ordinary bad days and a poor guide to extraordinary ones: it says where the tail begins and nothing of what is in it, and it shrinks in calm markets. Read it with expected shortfall and a stress test, and never as a maximum.
Check your understanding
4 questions on this lesson. Each answer comes with the reason, and nothing is scored.
A portfolio has a one-day 99% VaR of $3,000. Over a year of about 250 trading days, how often should it lose more than $3,000 in a day?
- Never; $3,000 is the maximum daily loss
- About two or three times
- About 25 times
Show the answer
B. About two or three times A 99% VaR is exceeded on about 1% of days: 250 × 0.01 = 2.5. VaR is a threshold that is expected to be broken at a known rate, not a ceiling.
Two portfolios both have a one-day 95% VaR of 1.6%. Portfolio A’s expected shortfall is 2.0% and Portfolio B’s is 5.5%. What is the difference between them?
- None; equal VaR means equal risk
- When B has a day worse than its VaR, the loss is on average far larger
- B has fewer bad days
Show the answer
B. When B has a day worse than its VaR, the loss is on average far larger VaR marks where the worst 5% of days begin. Expected shortfall is the average of those days. B’s tail extends much further, which VaR alone cannot show.
On two years of SPY data, the bell-curve (parametric) 99% VaR predicted 5 days beyond its threshold, and there were 8. What does that indicate?
- An arithmetic error
- Real returns have more extreme days than a normal distribution assumes
- The sample was too long
Show the answer
B. Real returns have more extreme days than a normal distribution assumes Stock returns have fat tails: more very small and very large moves than the normal curve. A method that assumes normality understates losses at high confidence levels.
A VaR is estimated from a calm year and then applied to the next year, in which a crisis begins. What is likely to happen?
- Losses beyond the VaR occur far more often than the confidence level implies
- The VaR adjusts itself in advance
- Nothing; VaR does not depend on the estimation period
Show the answer
A. Losses beyond the VaR occur far more often than the confidence level implies VaR is computed from past returns. A calm sample gives a small VaR, which the following turbulent period breaks repeatedly. The measure is lowest just before it is needed.
Sources and further reading
- J.P. Morgan/Reuters, RiskMetrics — Technical Document, 4th ed. (1996)
- Philippe Jorion, Value at Risk: The New Benchmark for Managing Financial Risk, 3rd ed. (McGraw-Hill, 2007)
- Philippe Artzner, Freddy Delbaen, Jean-Marc Eber and David Heath, “Coherent Measures of Risk,” Mathematical Finance, vol. 9, no. 3 (1999)
- Basel Committee on Banking Supervision, Minimum capital requirements for market risk (2019) — replaces VaR with expected shortfall at 97.5%
- Finabase chart data — SPY, QQQ, NVDA, TLT, GLD and KO daily closes, November 2024 to November 2026, and SPY and AGG monthly bars and dividends through October 2026; statistics computed by Finabase
