Understanding Tail Risk and the 95% Confidence Band

A 95% band does not mean the other 5% is impossible. It means you should expect to be outside it roughly one day in twenty — and in real markets, rather more often than that.

What the number literally claims

A 95% confidence band asserts that, under the model's assumptions, about 95% of outcomes should fall inside it and about 5% outside. Over 252 trading days, that implies roughly 12 or 13 breaches a year.

Breaching the band is therefore not evidence the model is broken. Breaches are a required feature. A model whose 95% band was never breached would be badly miscalibrated — it would be far too wide to be useful.

Why real markets breach more often than predicted

The arithmetic behind most confidence bands assumes a normal or lognormal distribution. Real return distributions have fat tails: extreme moves occur considerably more often than the bell curve allows.

So a nominal 95% band computed on normal assumptions typically contains somewhat less than 95% of real outcomes, and — more importantly — the moves that fall outside it are larger than the model implies. The failure is not just frequency, it is severity.

Volatility clustering makes breaches arrive together

Breaches are not evenly distributed through time. Because volatility clusters, they arrive in bunches during stressed periods and are almost absent through calm stretches.

This matters more than the annual count. Twelve breaches spread evenly across a year is a manageable pattern; twelve breaches concentrated into three weeks is a drawdown. Any risk assessment built on average frequency without accounting for clustering will understate how bad a bad period can be.

Using bands without being fooled by them

Treat the outer band as a soft boundary, not a wall. It marks where outcomes become unusual, not where they stop.

Be especially wary of bands estimated during calm periods. If the volatility input was measured across a quiet stretch, the resulting band describes a regime that may no longer apply, and it will be breached far more often than its nominal confidence level suggests.

The practical discipline is to size positions so that a breach is survivable rather than catastrophic. Since fat tails guarantee breaches will be both more frequent and larger than the model states, the model should inform position size, never define the maximum loss you are prepared to absorb.