On this page
Value at risk, or VaR, is a loss threshold tied to a specified horizon and confidence level. A stated VaR does not say the worst loss possible, nor the average loss after the threshold is crossed. Read it as a quantile of a modelled loss distribution, then ask how the distribution was built and what happens beyond that boundary.
Translate the headline into a probability statement
Suppose a portfolio's illustrative one-day 95 per cent VaR is 2 million currency units. Under the particular model and portfolio used, the probability of a one-day loss no greater than 2 million is at least 95 per cent. The complementary tail has about five per cent probability of a worse loss in a continuous distribution. It does not mean the desk will lose exactly 2 million on five per cent of days, and it is not a promise about the next day. VaR is conditional on the positions, risk-factor behaviour and horizon assumed in the calculation.
The Bank for International Settlements describes VaR in terms of a specified confidence level and holding period in its market-risk application guidance. The GARP learning objectives are the candidate's syllabus anchor for the FRM curriculum. Their role here is to locate the skill, not to make a risk estimate accurate. A number without a horizon is incomplete: a one-day threshold and a one-month threshold answer different questions, even when stated for the same portfolio.
For a loss variable, find the smallest boundary whose cumulative probability reaches the selected confidence level. In a continuous distribution that is a straightforward percentile. For a discrete distribution with jumps, the chosen quantile convention matters: report the smallest loss whose cumulative probability is at least the required confidence. Stating the rule avoids ambiguity at a probability mass. Profit and loss can be framed with opposite signs, so write down whether positive numbers denote losses before identifying the tail.
Build a discrete example from invented outcomes
Consider an entirely invented portfolio with four possible one-day losses. There is an 88 per cent probability of no loss, an eight per cent probability of a 1 million loss, a three per cent probability of a 4 million loss, and a one per cent probability of a 15 million loss. These probabilities sum to 100 per cent; they are invented to expose the quantile, not estimated from a real portfolio. For a 95 per cent VaR, the cumulative probability at zero loss is 88 per cent. Including the 1 million outcome brings it to 96 per cent. Therefore the smallest loss threshold covering at least 95 per cent of outcomes is 1 million.
The tail beyond the 1 million boundary still contains the 4 million and 15 million losses, with combined probability four per cent. A reader who says that the worst one-day loss is 1 million has discarded precisely the outcomes risk management is concerned about. For a 99 per cent quantile, cumulative probability reaches 99 per cent at the 4 million loss, so the 99 per cent VaR under this explicit convention is 4 million. The 15 million loss remains beyond it. Raising confidence moves the boundary further into the tail but does not measure the severity of what remains outside.
The two VaR figures have different meanings; neither is a forecast of typical loss. The one-day expected loss in this artificial distribution would be a probability-weighted average of all four outcomes, a separate calculation. Expected shortfall instead asks about the average loss in a specified worst tail, with careful treatment of discrete probability at a boundary. The key conceptual difference is that VaR stops at a quantile whereas expected shortfall looks into tail severity. GARP's research discussion of conditional risk measures explicitly distinguishes those ideas.
Understand what changes the answer
Historical simulation ranks losses from selected past scenarios. Its apparent simplicity hides choices about observation window, position mapping and whether the selected history contains a relevant stress. A quiet history can produce a low VaR without proving the future is safe. A parametric method instead imposes a distributional structure, often a normal approximation for changes in a largely linear portfolio. That can make a calculation fast and transparent, but skewed payoffs or unusually heavy tails can make the assumed shape misleading. Monte Carlo simulation samples from an explicit model and inherits that model's assumptions. The method name never removes model risk.
Confidence level and horizon also affect comparability. A higher confidence level generally moves the quantile to a larger loss, while a longer horizon typically calls for modelling the portfolio and risk factors over that interval. Multiplying a one-day VaR by a square-root-of-time factor is an approximation that requires assumptions about independent, stable changes and an appropriate treatment of drift and positions. It is not a universal conversion. If a question changes the horizon, inspect the underlying assumptions before applying any shortcut.
Portfolio composition matters just as much. A VaR for a fixed set of positions is not automatically the VaR after a desk rebalances or an option's exposure changes with market moves. Basis risk, non-linear instruments and changing correlations can make a simple local approximation fragile. The BIS guidance notes that backtesting compares actual outcomes with the model's intended coverage and that shifts in market behaviour can make historical inputs less appropriate. A backtest exception is a diagnostic, not proof by itself of a single cause; data errors, model limits and genuine tail events need different responses.
Why a clean VaR can conceal a dangerous tail
Imagine two invented portfolios whose 95 per cent one-day VaR is the same 1 million. Beyond that threshold, one loses at most 2 million in its rare states, while the other sometimes loses 20 million. A VaR-only comparison declares them equal at the selected percentile despite radically different damage after crossing it. This is why VaR should be accompanied by a tail-sensitive measure and stress scenarios when interpreting exposure. The exercise does not require a candidate to know a regulatory capital formula; it requires recognising the missing information in the single number.
Consider a risk manager comparing a current portfolio with a proposed hedged version under the same invented loss scenarios. If both report a 95 per cent VaR of 1 million, the manager should still ask whether the hedge changes losses beyond the boundary, whether its payoff survives a liquidity shock, and whether the assumed position can actually be held over the stated horizon. A lower VaR from a calmer calibration window is not evidence that a hedge worked. Recalculate both positions on comparable scenarios and inspect selected stresses separately. This is a decision about the information the statistic can support: the quantile helps compare a chosen boundary, while scenario losses reveal what happens when that boundary fails.
A second trap is to treat the confidence level as an assurance about realised frequency. The statement is generated under a model. If volatility, correlations, liquidity or positions change, the actual exceedance rate can change too. Even with a perfectly calibrated continuous model, some observations are expected to exceed the threshold. BIS discusses exactly this feature of backtesting: exceptions can occur even when a model covers its stated proportion of outcomes. Conversely, a short run with no exceptions cannot certify that the model is sound in every market state.
The FRM program page situates this topic among the risk tools candidates need, and the methodology page explains Omni's approach to study. In an exam stem, write down portfolio, loss sign, time horizon, confidence level and method before calculating. Then translate the result into a sentence that names all those qualifiers. This habit prevents the familiar error of reporting a threshold as the worst possible outcome.
A practical interpretation checklist
Ask what loss distribution is being used, how current positions are represented, and whether the percentile is based on a normal approximation, historical outcomes or simulation. Check whether the stated horizon matches the risk decision. Locate the threshold as a quantile and keep losses beyond it visible. If the decision depends on how bad a breach can become, VaR alone cannot answer it. If the answer asks whether modelled coverage is borne out, look to backtesting, while remembering that too few observations and changing markets limit what it can establish. VaR communicates one useful boundary; a sound reading keeps its assumptions attached.
Sources
- Application guidance | BIS · bis.org · Retrieved
- 2026 FRM Learning Objectives | GARP · garp.org · Retrieved
- Conditional Risk Measurement | GARP · garp.org · Retrieved
- Value at Risk | CFA Institute · rpc.cfainstitute.org · Retrieved



