Tail risk: what VaR leaves beyond the threshold

A threshold is not a maximum

Value at Risk, or VaR, summarizes a loss quantile for a stated horizon and confidence level. A one-day 95% VaR is a threshold associated with the upper tail of the modeled one-day loss distribution. It is not a maximum possible loss, a promised stop level or a statement that tomorrow's loss is bounded.

The confidence level alone is incomplete. A dollar figure depends on portfolio size, holdings, valuation method and the distribution used. Historical, parametric and simulation approaches can give different answers because they make different assumptions. Always state the sign convention: some systems report a positive loss amount, while others show a negative return threshold.

Ask how severe the worse outcomes are

Expected shortfall examines the average loss in the selected worst portion of outcomes, with appropriate handling at a discrete quantile boundary. It therefore describes tail severity that a quantile alone can conceal. Two portfolios can share a similar VaR while having very different losses beyond that threshold.

This extra information does not make the measure infallible. Tail observations are scarce, so estimates can be unstable. A short sample may contain no crisis resembling the next one. Regulatory market-risk frameworks use expected shortfall in specified models, but an educational calculation is not a regulatory capital assessment and should not borrow authority from that terminology.

Discrete scenarios make the tail definition important. Suppose a loss is zero with probability 97%, $100 with probability 2% and $1,000 with probability 1%. Under the smallest-loss-quantile convention, 95% VaR is zero. The average of the worst 5% of outcomes is nevertheless (0.02 × $100 + 0.01 × $1,000 + 0.02 × $0) / 0.05 = $240. Simply averaging outcomes strictly above VaR gives $400 and answers a different question because those outcomes occupy only 3% of the distribution.

Historical absence is not impossibility

A historical simulation revalues or approximates the current portfolio under past changes. Its transparency is useful, but it cannot automatically represent events outside the sample. Structural changes in leverage, trading venues or the instruments themselves can make a familiar historical shock behave differently today.

Parametric models also require scrutiny. Assuming normally distributed returns can understate certain extreme outcomes and dependencies. Scaling one-day risk by the square root of time relies on restrictive assumptions about dependence and distribution; it is not a universal rule for illiquid positions or nonlinear products. Show sensitivity to alternative assumptions instead of presenting one model as an objective property of the portfolio.

Tail dependence and liquidity belong together

Extreme co-movement can differ from the correlation measured across ordinary days. Several apparently diversifying positions may become hard to sell simultaneously when funding tightens. Losses can then increase collateral demands, which force sales and create additional price pressure. A static return distribution may omit this feedback.

Include scenarios for spread widening, delayed execution and unavailable hedges. For options or leveraged products, revalue the position under changing sensitivities rather than extrapolating one small-move estimate. Operational interruptions can also turn a manageable adverse price move into a larger realized loss. Risk measurement is incomplete if it assumes a perfectly functioning exit process in every tail event.

Use several lenses and inspect exceptions

Pair quantitative tail measures with scenario losses, concentration reports and cash-flow stress tests. Each highlights something different. When realized losses exceed a modeled VaR threshold, investigate whether the event is consistent with the intended exceedance frequency, whether the model was implemented correctly and whether the underlying distribution changed.

Do not recalibrate solely to make recent exceptions disappear. Preserve previous forecasts so the model can be evaluated honestly. A useful tail-risk report explains what is included, what is not, how uncertain the estimate is and which decisions depend on it. The largest danger is often treating a compact statistic as permission to ignore the assumptions that generated it.

Evaluate the forecast before judging the exceptions

A forecast and its realized comparison must refer to the same object. A one-day risk forecast for the positions held at yesterday’s close is not directly tested by a profit-and-loss figure that also includes today’s new trades, withdrawals and unrelated fees. Keep a holdings-based comparison if the question is whether the risk model described those holdings, and a separate actual account result if the question is what the trading process experienced. Both matter, but mixing them obscures the source of an exception.

Under an idealized correctly calibrated continuous-loss model, a 99% one-day VaR threshold is exceeded on 1% of days. Across 250 days, the expected number is 2.5, not exactly two or three. If exceedances were independent with constant probability, the chance of at least one would be 1 − 0.99²⁵⁰, about 91.9%. An exception is therefore not automatically proof that the model failed. Conversely, a small total count does not establish adequate calibration when all the exceptions occur in a short cluster.

Severity is a different diagnostic from frequency. Losing just beyond the threshold and losing ten times the threshold both count as one VaR exception, but they have very different consequences. Inspect their size, the conditions in which they occurred, and the forecasts that preceded them. A model that gets the average count roughly right while missing a concentration of severe funding-related losses has not answered the full risk question.

Sample size becomes particularly restrictive at high confidence levels. In 1,000 equally weighted historical scenarios, the worst 1% contains only ten observations. Replacing one of those observations can materially change the tail average, especially if it is an extreme loss. Changing the horizon or adding stressed scenarios changes the question again. Report the number of effective tail observations, the weighting convention and the sensitivity to a few large outcomes alongside the headline measure; extra decimal places cannot compensate for scarce information.