Quant Analytics: from scratch to pro · Portfolio, sizing and risk metrics · Intermediate
Value at Risk and expected shortfall
Value at Risk names a loss level that is exceeded on a small share of days. Expected shortfall averages the losses beyond it. Banking regulation moved from the first to the second for internal models, for reasons this lesson works through.
In this lesson
- Compute parametric value at risk and expected shortfall under a normal assumption
- Compute both from a sample of historical returns
- Scale a one-day figure to a longer horizon and say when that fails
- Explain why expected shortfall replaced value at risk in the Basel internal models approach
Read first: Volatility: from daily moves to an annual figure, Return distributions and fat tails
Risk reports need one number a manager can compare with a limit. Value at Risk (VaR) is the most widely used. It answers a narrow question, and the answer is easy to misread. Expected shortfall (ES) answers the question people usually mean.
Definitions
The one-day VaR at confidence level p (say 99%) is the loss that is exceeded on a fraction 1 - p of days (1% here). Equivalently, it is the p quantile of the loss distribution. See value at risk.
Expected shortfall is the average loss on the days when the loss exceeds the VaR. It looks into the tail instead of stopping at its edge. See expected shortfall.
Parametric version
If daily returns are normal with mean zero and volatility sigma, then:
VaR_p = z_p x sigma x V and ES_p = sigma x pdf(z_p) / (1 - p) x V
where V is the portfolio value, z_p the normal quantile and pdf the standard normal density.
Worked example
A portfolio is worth 1,000,000 and has daily volatility of 1.2%.
- 99% VaR:
z = 2.326, so VaR = 2.326 x 1.2% x 1,000,000 = 27,916. - 97.5% VaR:
z = 1.960, so VaR = 23,520. - 97.5% ES:
1.2% x pdf(1.960) / 0.025 x 1,000,000= 28,054. - 99% ES: 31,983.
Under a normal distribution the 97.5% ES (28,054) is close to the 99% VaR (27,916). The two figures are close in scale, but only under a normal distribution. For ten days, square-root scaling gives a 99% VaR of 27,916 x sqrt(10) = 88,279.
Historical version
Without a distribution assumption, sort past returns. With 1,000 days, the 99% VaR is the 10th worst loss and, by one common convention, the ES is the average of the ten worst. Suppose the ten worst losses are 6.0%, 5.1%, 4.4%, 4.0%, 3.7%, 3.5%, 3.4%, 3.3%, 3.2% and 3.1%. VaR is 3.1% and ES is their average, 3.97%. The historical method inherits the sample: a calm window shows small tails, and ten observations give a noisy average.
Why regulators moved to expected shortfall
VaR is a threshold. Two portfolios can share the same VaR and differ widely in what happens past it. VaR also fails the subadditivity test in the framework of Artzner, Delbaen, Eber and Heath (1999): the VaR of a combined portfolio can exceed the sum of the parts' VaRs, which contradicts the idea that diversification reduces risk. Expected shortfall satisfies that property. The Basel Committee's revised market risk standard (published January 2019) uses expected shortfall at a 97.5% confidence level for internal models. See the Basel history and tail risk.
Checking a VaR model
At 99% over 250 days a sound model produces about 2.5 exceptions (days when the loss exceeds VaR). Kupiec (1995) formalised the test: compare the count of exceptions with the expected count. With few observations the test has little power; a model that is wrong can pass.
Common mistakes
- Reading VaR as the worst case. It is the boundary of the worst 1% (or 2.5%), not its size.
- Assuming normality for assets with fat tails. See return distributions.
- Scaling to long horizons with the square root when volatility clusters.
- Using a calm window for the historical estimate. The tail is then understated.
- Comparing VaR figures that use different confidence levels, horizons or windows.
In practice
State the level, the horizon, the method and the window every time. Show ES next to VaR, and a stress scenario next to both, since neither describes a loss the sample never contained.
Normal VaR and ES multipliers against confidence level
- VaR
- expected shortfall
Worked numbers
Each result is recomputed from the formula shown.
| What | Formula | Result |
|---|---|---|
| 99% normal quantile | phi_inv(0.99) | 2.326 |
| 99% one-day VaR on 1,000,000, volatility 1.2% | z99 * 0.012 * 1000000 | 27,916 |
| 97.5% normal quantile | phi_inv(0.975) | 1.960 |
| 97.5% VaR | z975 * 0.012 * 1000000 | 23,520 |
| 97.5% expected shortfall | 0.012 * pdf(z975) / 0.025 * 1000000 | 28,054 |
| 99% expected shortfall | 0.012 * pdf(z99) / 0.01 * 1000000 | 31,983 |
| 10-day 99% VaR by square-root scaling | var99 * sqrt(10) | 88,279 |
| average of the ten worst daily losses, in percent of value | (6.0 + 5.1 + 4.4 + 4.0 + 3.7 + 3.5 + 3.4 + 3.3 + 3.2 + 3.1) / 10 | 3.97 |
Check yourself
A portfolio of 500,000 has daily volatility of 1%. What is its one-day 99% normal VaR?
2.326 x 1% x 500,000 = about 11,632. On about 1 day in 100, the loss is expected to be at least this large.
Does a 99% VaR of 11,632 say the loss cannot exceed that figure?
No. It is a threshold, not a maximum. It says nothing about how large losses are once past it.
Why do two portfolios with the same 97.5% VaR have different expected shortfall?
Expected shortfall averages the losses beyond the VaR level. A portfolio with a fatter or longer tail has the same threshold but a larger average loss past it.
You have 1,000 daily returns. How do you read the historical 99% VaR and ES?
Sort the losses. VaR is the 10th worst loss, ES is the average of the 10 worst.
Why is square-root-of-time scaling of a one-day VaR to ten days only an approximation?
It assumes independent days with constant volatility. Volatility clustering and autocorrelation break both assumptions, usually making the real ten-day tail heavier.
Key terms
Checked facts
Every figure below carries how it was checked: Primary checked (read in the cited primary, official, academic or data source), Two sources agree, or Approximate (an order of magnitude only).
- Primary checked
Under the Basel Committee's revised market risk framework, internal models measure risk with expected shortfall at a 97.5% confidence level, replacing value at risk.
Sources
- Minimum capital requirements for market risk (d457). Basel Committee on Banking Supervision, Bank for International Settlements, 2019. Accessed 9 Nov 2026
- Coherent Measures of Risk. Mathematical Finance 9(3), 203-228 (Philippe Artzner, Freddy Delbaen, Jean-Marc Eber, David Heath), 1999. Accessed 9 Nov 2026
- Techniques for Verifying the Accuracy of Risk Measurement Models. Finance and Economics Discussion Series 95-24, Board of Governors of the Federal Reserve System (Paul H. Kupiec), 1995. Accessed 9 Nov 2026