WhitmanTrading

What Is Tail Risk?

Tail risk is the exposure to rare events producing outsized losses, occupying the extreme end of a return distribution. It is systematically underestimated because standard models assume a normal distribution, while real financial returns produce extreme moves far more often than that assumption allows.

Tail risk is the part of the distribution where the large losses live. It matters more than anything else in risk management and it is the part we have the least data about.

How it works

A price series with a rare large loss at the edge.
Tail risk is the rare, large loss. Illustrative chart - not real market data.

Plot every return and most cluster near the middle. The tails are the far ends — the handful of observations representing enormous moves in either direction.

A steady series where models understate extremes.
Standard models assume it is rarer than it is. Illustrative chart - not real market data.

A normal distribution says extremes are vanishingly rare. A five-standard-deviation move should occur roughly once in several thousand years under that assumption.

A rising series with more extremes than expected.
Real returns have fatter tails than normal ones. Illustrative chart - not real market data.

Financial markets produce them every few years. The assumption is not slightly wrong; it is wrong by orders of magnitude at the extremes, which is where it matters.

A falling series where losses exceed the modelled worst case.
So the worst case is worse than modelled. Illustrative chart - not real market data.

Why it resists measurement

A choppy series with too few extreme observations.
The data needed to measure it is the data nobody has. Illustrative chart - not real market data.

Estimating a rare event needs observations of it. A once-in-twenty-years event appears perhaps four times in eighty years of data, and four observations support no confident estimate of anything.

A slow series where extremes appear only over decades.
And different again over a long horizon. Illustrative chart - not real market data.

And markets change. Eighty-year-old data describes a market with different participants, instruments, regulation and speed, so the old observations may not describe the current system at all.

A calm series where no extreme is visible.
A quiet stretch hides what it measures. Illustrative chart - not real market data.

So every tail estimate is an extrapolation. Extreme value theory does this more carefully than a normal assumption does, and it is still fitting a curve past the edge of the data.

A worked example

Take this site’s shared series. The median bar range is 0.493, the ninetieth percentile is 1.101, and the largest single bar is 2.338.

A normal distribution fitted to the bulk would not predict that largest bar. It is 4.7 times the median, and the gap between the ninetieth percentile and the maximum is where the whole problem lives.

A falling series with a stop level marked.
A stop fills where the market is. Illustrative chart - not real market data.

A position sized against the ninetieth percentile survives most things. Against the maximum observed, it survives everything in the record — and the record is not the same as everything possible.

Which is the uncomfortable conclusion. The largest move in any dataset is, by definition, the largest so far, and there is no statistical reason to treat it as a ceiling.

The correlation problem

Tails arrive together. In ordinary conditions, holdings diversify each other; in a severe event, correlations rise toward one and the diversification largely disappears.

Which is exactly backwards from what is needed. The protection works when it is not required and weakens when it is, because the same shock is driving everything.

And leverage converts a survivable tail into a fatal one. A 30% decline is unpleasant unmargined and terminal at three times leverage, and the position is closed by somebody else at the worst moment.

So the practical defence is structural rather than statistical. Size positions so an extreme is survivable, avoid forced selling, and hold enough liquidity to never be a seller at the bottom — none of which requires estimating a probability.

The original data

On this site’s shared series: median bar range 0.493, ninetieth percentile 1.101, largest bar 2.338. ATR14 has a median of 0.5994 and a ninetieth percentile of 0.7954. Drawdown reaches a maximum of 3.76% with 95% of bars below a prior peak and a longest stretch of 73 bars.

Notice what those figures do not contain. Nothing resembling a crisis, because the series is an ordinary one — which is precisely the point. Most data is ordinary, tails are not in it, and models calibrated on ordinary data have nothing to say about them.

A candlestick chart annotated with the cost of a round trip.
A round trip costs a share of a bar. Illustrative chart - not real market data.

And the round trip cost of 0.0098 — about 2% of the median bar — is the cost of acting. In a tail event real spreads widen by multiples, so the exit that was cheap when modelled is expensive when used.

A price series with volume shown beneath.
Volume and price measure different things. Illustrative chart - not real market data.

What can actually be done about it

Hold less leverage than feels necessary. This is the only defence that works regardless of what the event turns out to be, and it costs return in every year nothing happens.

Buy explicit protection. Deep out-of-the-money puts pay in a crash and bleed premium continuously, which is a known cost for an uncertain benefit.

Hold genuinely uncorrelated assets. Genuinely is the difficult word, since most correlations that look low in calm periods rise in stressed ones.

And accept that none of these is free. Every tail defence costs something in normal conditions, which is why they are abandoned during long calm periods — and why the abandonment tends to complete itself shortly before it matters.

Where the phrase gets misused

Not every large loss is a tail event. A concentrated position falling 40% because the company disappointed is an ordinary outcome of a concentrated position, and calling it a tail event relabels a sizing decision as bad luck.

And not every tail event is unforeseeable. Many were discussed in advance by people who were ignored, which is a governance problem rather than a statistical one.

The useful test is whether the loss was in the modelled distribution. If the model said it was essentially impossible and it happened, that is a tail event; if the model said it was unlikely and somebody sized as though it were impossible, that is something else.

The distinction matters because the remedies differ. One calls for more humility about models; the other calls for acting on what the model already said.

When it fails

The characteristic failure is a risk model that passed every test. The value-at-risk figure was calculated correctly, backtested against years of history, and approved. It said a 2% daily loss was the 99th percentile outcome, and then an 8% day arrived. The model was not broken — it answered the question it was asked, which was about the distribution’s bulk, and the tail was never inside the data it was fitted to. Passing a backtest on ordinary data is not evidence about extraordinary days, and treating it as such is the single most common error in institutional risk management.

A candlestick series with a gap through a level.
A gap skips the level entirely. Illustrative chart - not real market data.

A second failure is assuming diversification holds. Correlations converge in exactly the events being defended against.

A third is treating the worst observed move as a maximum. It is the largest so far.

A fourth is sizing to survive the average bad day rather than the rare one.

A declining series cut short at a decision point.
Nothing extreme in ten years. Reduce the hedge? Illustrative chart - not real market data.

And a fifth is dropping protection during calm. The cost of a hedge is most visible precisely when it has not been needed for a long time.

Downside risk covers measuring only the losing side of a distribution. Tail risk parity covers sizing positions against this exposure. And volatility risk covers the ordinary variation these events sit beyond.

What I actually do

The honest position on tail risk is that it cannot be measured well, because measuring it properly requires observations of events that have happened a handful of times. Anybody quoting a precise probability for a once-in-fifty-years event is extrapolating, and should say so.

— Michael Whitman

This page is educational, not financial advice. Test every idea on your own charts before risking money.