Fat Tails: Why a Once-in-a-Lifetime Crash Can Arrive Every Few Years
Most numerical risk tools assume returns follow a bell-shaped distribution. Under that assumption, large declines become virtually impossible. Yet markets actually experience such days every few years. The gap between that assumption and reality is called a fat tail.
A Normal Distribution Treats Large Events as Almost Impossible
In a normal, bell-shaped distribution, distance from the mean is expressed in standard-deviation multiples, or sigma (σ). Once an event's sigma size is known, its probability follows automatically. The problem is how extremely fast that probability falls as sigma rises.
|z| > 2 → 4.55% → About once per 22 trading days
|z| > 3 → 0.27% → About once per 370 trading days
|z| > 4 → 0.0063% → About once per 15,800 trading days
|z| > 5 → 0.000057% → About once per 3,489,000 trading days
For only the downside tail, probabilities are half those above.
A five-sigma decline occurs about once per 6,977,000 trading days.
Divide by 365 → About 19,000 years.
─────────────
From two sigma to five sigma,
probability becomes 80,000 times smaller.
Sigma rises only from two to five, yet frequency changes from once in 22 days to once in 19,000 years. The normal distribution's shape makes its tails thin extremely quickly. Risk estimates based on it therefore treat large events as effectively zero.
Converting Sigma into Actual Price Moves
Translate those multiples into prices to see the scale. Assume an asset has a daily-return standard deviation of 3.5%.
2σ = ±7.0% → About once a month
3σ = ±10.5% → Once every year and a half
4σ = ±14.0% → Once every 63 years
5σ = ±17.5% → Once every 19,000 years
─────────────
Trusting a normal distribution implies that a −17.5% day
is something you should never see in your lifetime.
Yet people who have watched markets for a few years
have already seen several such days.
That discrepancy is a fat tail. Actual return distributions are more peaked in the center and much thicker at both ends than a normal distribution. Ordinary days can be quieter than expected, while event days occur much more often and move much further. Black swans covers representative examples.
Measuring Tail Thickness: Kurtosis
Kurtosis summarizes tail thickness. A normal distribution has kurtosis 3; higher values mean more frequent extremes than the normal model.
Kurtosis 3: The normal distribution
A 4σ event once per 15,800 days.
Kurtosis 6: Fat tails
The same 4σ event becomes much more frequent.
Kurtosis 9: Very fat tails
A 4σ event can become an event occurring every few years.
─────────────
Even with identical means and standard deviations,
different kurtosis can produce entirely different account risk.
Mean and standard deviation alone cannot reveal that difference. Two assets may both have daily volatility of 3.5%, while one fluctuates similarly each day and the other stays quiet before occasionally falling 20% at once. Measures summarizing risk with standard deviation miss this distinction. It also explains criticism that the Sharpe ratio understates extreme losses.
Leverage Brings the Tail Directly into the Account
In spot holdings, −17.5% is a painful day. In a leveraged position, it may be the end of the position. Expressing liquidation distance in sigma shows why.
Initial margin: $100; position: $1,000
Maintenance margin rate: 0.5% → $5
Loss buffer:
$100 − $5 = $95
As a price move:
$95 ÷ $1,000 = −9.5%
At 3.5% daily standard deviation:
9.5% ÷ 3.5% = 2.7σ
Under a normal distribution, probability of that adverse daily move in one tail ≈ 0.33%
→ About once per 300 trading days.
With fat tails, this can happen
several times more often than calculated.
─────────────
At 10×, the safety buffer is only 2.7σ.
Even the normal model suggests roughly once a year, and the actual distribution reaches that zone more frequently. More leverage shortens the distance. At 20× it is roughly halved to 1.4σ, a range crossed several times even within one month. See leverage for liquidation distance, and position sizing with ATR for adjusting size to volatility.
Stops Slip During Tail Events
“My stop is −3%, so my maximum loss is −3%” works only under the assumption that prices move continuously. Tail events do not follow that continuity assumption.
Entry: $100; stop order: $97, or −3%
Ordinary Day
$100 → $99 → $98 → Fill at $97
Actual loss: −3%
Tail-Event Day
$100 → $90 in one jump
The stop fills at $90, not $97.
Actual loss: −10%
At 10× leverage:
−10% × 10 = Beyond the entire margin.
─────────────
A stop limits losses,
but cannot limit them exactly in a tail event.
Calculating maximum loss solely from stop distance is therefore optimistic. The risk reference should be how much the account loses during a gap, not only the stop price. Stop-loss orders covers order types and execution; slippage covers fills away from the intended price.
Tails Exist on the Profit Side Too
Fat tails exist in both directions. Just as large losses cluster, large gains also concentrate in a few days. That complicates the conclusion that you should frequently stay out because markets are risky.
Across 1,000 trading days, suppose much of the total gain
came from the best ten days, just 1% of the period.
Remove those ten days:
Cumulative return → Much smaller or negative.
Remove the worst ten days instead:
Cumulative return → Unrealistically good.
─────────────
Both calculations are possible only afterward.
You cannot know that morning which days will be those ten.
The response cannot simply be to select the big days: their unknowability is part of the tail problem. What remains is entering at a size that does not end the account whichever day arrives. See risk of ruin for the point of irrecoverable damage, drawdown for the return needed to recover, and the Kelly criterion for a sizing upper bound.
How Backtests Erase Tails
The problem also follows historical testing. Tail events are rare, so results can change completely depending on how many fall inside the test period.
Test period: Two years, about 730 trading days
Expected number of 4σ events: 0.05
→ Most two-year periods contain none.
Consequently:
A period without a tail event → Results look excellent.
A period with one → Results abruptly deteriorate.
The same strategy can yield opposite conclusions
when only the test period changes.
─────────────
Report Together
The worst day included in the period.
Its loss and maximum drawdown.
Whether gap execution was modeled.
When a backtest looks good, inspect its worst day before average returns. Performance from a period without the worst conditions has yet to be tested by them. The backtesting guide covers the overall process, walk-forward analysis covers shifting the test period, and multiple-testing traps explains retaining only good results after repeated trials. Survivorship bias addresses counting only survivors.
Key Points
② A 5σ decline is theoretically once in 19,000 years.
③ Real markets experience such moves every few years.
④ That gap is the fat-tail problem.
⑤ Kurtosis measures thickness; the normal value is 3.
⑥ Identical mean and standard deviation can conceal different tail risks.
⑦ The example's 10× liquidation distance is 2.7σ.
⑧ At 20× it is about 1.4σ, a range crossed several times a month.
⑨ A −3% stop can fill at −10% in a gap.
⑩ Large gains also concentrate in a few days.
⑪ Those days cannot be identified in advance.
⑫ A backtest needs its worst day reported alongside it.
Calculate risk from the worst day, not the average day. Sizing around ordinary volatility may work ordinarily, but one tail-event day can invalidate all those calculations.
Notice
The 3.5% daily standard deviation, 0.5% maintenance rate, $100 margin and $1,000 position, −9.5% and 2.7σ liquidation distance, $97 stop and $90 gap fill, kurtosis values 3, 6, and 9, best ten of 1,000 trading days, two-year test, and 0.05 expected 4σ events are hypothetical examples illustrating how fat tails affect an account, not measurements from a particular asset or account. Sigma probabilities under a normal distribution are mathematically defined; the point is that they cannot be applied unchanged to actual markets. How much extreme-event frequency rises with kurtosis depends on the distribution, so those descriptions indicate direction only. Maintenance rates and liquidation procedures differ by exchange, asset, and margin mode; check actual liquidation distance in your account. Past distributions are not guaranteed to persist. Leveraged trading can lose all principal. Investment decisions and their consequences are your responsibility.
NOONOO TRADING invites you to follow live trading in our free chat.
Start in the bot📈 OKX trading fee discount for new registrations
Register for the OKX Fee Discount →