Risk & portfolios

An introduction to value at risk

Risk & portfolios · Published 2026-10-26

A Japanese abacus lying below an electronic calculator with an abacus built into it
An abacus and a calculator-abacus in use at a shop in Otaru, Japan. Photo: Syced, CC0.

Value at risk (VaR) condenses the risk of a portfolio into a single number: a loss that should be exceeded only rarely over a given period. Since the 1990s it has been used by banks, asset managers and regulators. It is easy to state, easy to misunderstand, and important to know well. This article explains how it is calculated, how it is checked, and where it falls short.

The definition

VaR has three ingredients: a time horizon, a confidence level and a currency amount. A one-day 99 per cent VaR of one million euros means that, under the model's assumptions, the portfolio should lose more than one million euros on only about one trading day in a hundred. Put another way, VaR is a quantile of the distribution of profit and loss.

VaR says nothing about how much is lost on the bad days beyond the threshold. On the one day in a hundred, the loss might be slightly more than one million or many times more. That limitation is central to the criticisms below.

Origins

VaR spread in the early 1990s. J.P. Morgan developed an internal daily risk report and in 1994 published its methodology and volatility data as RiskMetrics, with a technical document released in more detail in 1996. The Basel Committee's 1996 Market Risk Amendment allowed banks to use internal VaR models, at a 99 per cent confidence level over ten days, to set capital requirements for trading positions.

Method 1: historical simulation

Take the portfolio's current positions and apply to them the actual daily price changes from a past window, for example the last 500 trading days. That produces 500 hypothetical daily profit-and-loss outcomes. Sort them; the 99 per cent VaR is the loss at the 1st percentile, roughly the fifth-worst outcome.

Historical simulation makes no assumption about the shape of the distribution and naturally captures fat tails and correlations present in the window. Its weakness is that it assumes the future resembles the chosen window. A calm window produces a low VaR just before a storm.

Method 2: parametric (variance-covariance)

Assume returns are normally distributed. Estimate the portfolio's standard deviation σ from the volatilities and correlations of its components. Then VaR at confidence level c is approximately:

VaR = zc × σ × portfolio value

where zc is the normal quantile: about 1.645 for 95 per cent and 2.326 for 99 per cent. For example, a portfolio worth 10 million with daily volatility of 1.2 per cent has a one-day 99 per cent parametric VaR of about 2.326 × 0.012 × 10,000,000 ≈ 279,000. This is an arithmetic illustration, not a real portfolio.

The method is fast and transparent, and RiskMetrics popularised it with exponentially weighted volatility estimates. But normality understates the frequency of extreme moves in most financial returns, and the linear approach handles options poorly.

Method 3: Monte Carlo simulation

Specify a statistical model for the risk factors, such as a multivariate distribution with fat tails or a GARCH volatility model. Simulate thousands of scenarios, revalue the portfolio fully in each, and read the VaR off the simulated distribution. Monte Carlo handles non-linear instruments and complex dependencies, at the cost of computing time and of depending heavily on the chosen model.

Scaling across horizons

Regulators and investors often want VaR over several days. Under the assumption of independent, identically distributed returns, volatility scales with the square root of time, so a ten-day VaR is about √10 ≈ 3.16 times the one-day VaR. Real returns show volatility clustering and sometimes autocorrelation, so the square-root rule can understate multi-day risk in turbulent markets.

Checking whether VaR works

A VaR model can be tested by counting exceptions: days when the actual loss exceeded the VaR. A 99 per cent model should produce about 2.5 exceptions in 250 trading days. Paul Kupiec's 1995 test checks whether the number of exceptions is statistically consistent with the confidence level; Peter Christoffersen's 1998 tests also check whether exceptions cluster together, which would show that the model reacts too slowly. The Basel framework uses a "traffic light" system: a bank with too many exceptions in a year faces higher capital multipliers.

The weaknesses

Expected shortfall

Expected shortfall (ES), also called conditional VaR, is the average loss on the days when the loss exceeds the VaR threshold. It answers the question VaR leaves open: when things go badly, how badly on average? It is coherent in the sense of Artzner and colleagues. In its Fundamental Review of the Trading Book, finalised in 2019, the Basel Committee replaced VaR with expected shortfall at a 97.5 per cent confidence level for calculating market-risk capital under internal models, while keeping VaR-based backtesting.

Using VaR sensibly

VaR remains useful as a common language and a daily monitoring tool. Its value increases when it is reported alongside expected shortfall, stress tests based on historical crises and hypothetical scenarios, and the actual distribution of exceptions. For a strategy researcher, computing VaR and ES from a backtest is a simple way to translate volatility into money and to check whether the chosen position sizes match the risk appetite discussed in choosing your risk.

Sources