Fat Tails: Why 5-Sigma Days Happen Far More Often Than the Bell Curve Says

Fat tails describe a plain fact: real daily returns contain very large moves tens to thousands of times more often than the normal distribution (the bell curve) predicts. Risk formulas built on the normal distribution therefore say a disaster day is “practically impossible”, when in fact one arrives every few years. This article computes how big the difference is, shows what it does to Stop losses, leverage and high-win-rate systems, and ends with rules you can actually use.

What 5-sigma means

σ (sigma), the standard deviation, is the normal size of the daily swing in returns. If a pair moves 0.5% on a typical day, a 1σ day is 0.5%, a 4σ day is 2% and a 5σ day is 2.5%.

If returns really were normally distributed, the probability of each size would be fixed by the formula. The further from zero, the faster the probability shrinks, until a 5σ day is almost zero. The problem is that markets don’t behave like that.

The numbers: normal versus fat-tailed

We compare the normal distribution with the Student-t distribution, which has heavier tails. The smaller its df (degrees of freedom), the heavier the tails. We scaled every version to the same σ of 1 so they compare directly, and computed everything in node (numerical integration, checked against the closed-form formula for t with 3 df, which matched).

How to get the waiting time: if the chance of a day moving more than kσ, up or down, is p per day, the average wait = 1 ÷ (p × 252 trading days per year).

Example for a normal 4σ day: p = 0.0000633, so 1 ÷ (0.0000633 × 252) = 62.7 years.

Day size	Normal: p per day	Normal: average wait	t 5 df: average wait	t 3 df: average wait
Beyond 3σ	0.27%	1.5 years	0.34 years	0.29 years
Beyond 4σ	0.0063%	63 years	1.1 years	0.64 years
Beyond 5σ	0.000057%	6,900 years	3.0 years	1.2 years
Beyond 6σ	0.0000002%	2 million years	6.9 years	2.1 years

Under the normal formula a 5σ day takes about 6,900 years to arrive. With tails as heavy as a t with 5 df, it takes about 3 years, a difference of more than 2,000 times. For a 10σ day the normal formula gives a wait with 21 digits, while t with 3 df gives about 9.5 years.

Notice something in the first chart: between about 0.7σ and 2.55σ the fat-tailed distribution has fewer days than the normal one. Fat-tailed data looks calm most of the time, and that is exactly what makes people careless.

Kurtosis in plain words

Kurtosis is the number that measures how heavy the tails are. The normal distribution has a kurtosis of 3. Many programs report excess kurtosis, which subtracts 3, so normal shows as 0.

• Student-t with 5 df has a kurtosis of 9 (excess 6). We checked by integration and got a fourth moment of 8.99, matching the formula
• Student-t with 3 df has infinite kurtosis. A value measured from data keeps growing as you collect more days

The picture to remember: high kurtosis does not mean “more volatile”, because σ is the same. It means the same amount of volatility is redistributed. More quiet days, fewer medium days, and bigger explosions.

One important source is that volatility changes from period to period. A worked example: say 90% of days are calm with σ = 0.8 and 10% are panicky with σ = 2, and each regime on its own is perfectly normal. Combined, kurtosis = 3 × (0.9 × 0.8⁴ + 0.1 × 2⁴) ÷ (0.9 × 0.8² + 0.1 × 2²)² = 3 × 1.9686 ÷ 0.976² = 6.2. The tails are fat straight away, even though no single part is abnormal. Read more on how volatility bunches together in Volatility clustering.

Mandelbrot and real-market examples

The mathematician Benoit Mandelbrot published “The Variation of Certain Speculative Prices” in 1963. He studied cotton prices and showed that price changes have far heavier tails than the normal distribution. The same result has since been found again and again in stocks, currencies and commodities. The term Black swan was popularised by Nassim Nicholas Taleb’s 2007 book. It means an event that is rare, has a severe impact, and is explained well only after it happens.

Examples currency traders know (figures are approximate):

• 19 October 1987: the Dow Jones fell 22.6% in one day. If normal daily volatility was around 1%, that is roughly a 20σ day
• 15 January 2015: the Swiss National Bank (SNB) scrapped, without warning, the EUR/CHF floor at 1.20 it had held since 2011. EUR/CHF broke below 1.00 within minutes, a fall of more than 15%. Many client accounts went negative, and some brokers went under
• 7 October 2016: GBP/USD fell about 6% within minutes during thin Asian trading, then recovered most of it. It became known as a flash crash
• August 2024: the unwinding of the yen carry trade. USD/JPY went from around 161 in early July to around 142 on 5 August, and the Nikkei fell more than 12% that day (Carry trade)

For a 20σ day like 1987, the normal formula gives a waiting time with 86 digits, vastly longer than the age of the universe, while t with 3 df gives about 75 years. In real markets, moves the normal formula calls “impossible” turn up every few years in one asset or another.

Consequence 1: your Stop loss can be jumped

A Stop loss is an order that waits for price to touch it. Once touched it becomes a market order and fills at the next available price. If price jumps over the level, at the Monday open, on news, or when liquidity disappears, the real loss is bigger than planned (Order types).

Example: a 10,000-dollar account risking 1% = 100 dollars, with a 20-pip Stop loss on a pair where 1 lot = 10 dollars per pip, such as EUR/USD.

1. Lot = 100 ÷ (20 × 10) = 0.5 lot
2. Price jumps 200 pips from your entry before there is any price to close at
3. Loss = 0.5 × 10 × 200 = 1,000 dollars = 10% of the account, not 1%

Your real risk per trade is “position size × the gap that can happen”, not “position size × Stop loss distance”. Read more on gaps in Weekend gaps and News spikes and slippage.

Consequence 2: VaR understates exactly where it matters

VaR (Value at Risk) answers “what is the most I lose in a day, with 99% confidence”. The problem is that VaR only tells you where the door is, not how deep the room behind it goes.

Measure (loss side, in σ)	Normal	t 5 df	t 3 df
VaR 99%	2.33	2.61	2.62
VaR 99.9%	3.09	4.57	5.90
Average loss on days worse than VaR 99%	2.67	3.45	4.04

At 99% the three differ by only about 12%, which is why people believe the normal formula “works”. At 99.9%, t with 3 df is almost twice as far, and the average loss on the worst 1% of days (called Expected shortfall) is clearly bigger. The risk that wipes out an account lives in the part VaR doesn’t look at.

Consequence 3: leverage plus fat tails is a shortcut to zero

Effective leverage is the total value of your positions ÷ your equity. The loss as a percentage of equity = leverage × the percentage price move.

Effective leverage	Price moves 2%	6% (like GBP 2016)	15% (like CHF 2015)
5×	−10%	−30%	−75%
10×	−20%	−60%	−150%
20×	−40%	−120%	−300%
30×	−60%	−180%	−450%

Anything beyond −100% means a negative balance. On a day like that, Stop out can’t close you in time either, because there are no prices in between to close at. Whether you have negative balance protection depends on the broker and the rules of each country, so check your account terms yourself. Margin and Stop out mechanics are covered in Leverage, margin and stop out.

Fat tails don’t make you lose more often. They make the big loss bigger than you think, and leverage multiplies that one loss.

Consequence 4: picking up pennies in front of a steamroller

“Picking up pennies in front of a steamroller” describes a strategy that collects small, steady gains in exchange for tail risk that strikes rarely but crushingly. Selling far out-of-the-money options, Grid and Martingale are all examples (Grid and Martingale).

We simulated the simplest version in node:

• Every day the “insurance” seller collects one small premium
• If the day moves more than 2.5σ, the seller pays out the part beyond 2.5σ
• The seller prices the premium at 1.5 times its fair value under the normal formula, believing they have a 50% edge
• 2,520 days (10 years), 10,000 paths, once with normal returns and once with Student-t 5 df returns of the same σ

Results (unit = 1 day’s premium):

	Normal world	Fat-tailed world, t 5 df
Days with no payout (wins)	98.8%	97.7%
Average payout per day	0.67	3.19
Expected total over 10 years	+840	−5,500
Paths ending below zero	2.7%	100%
Worst single day (median)		753 days of premium

The average payout in the fat-tailed world is 4.78 times what the normal formula expects, and with t at 3 df it is 7.74 times. The daily win rate barely changes, which is why looking at win rate will never show you the problem.

In real option markets, dealers already charge more for far out-of-the-money options than the normal formula would, which shows up as the volatility smile (FX options and implied volatility). So this example doesn’t say selling options always loses. It shows that anyone who thinks about risk in bell-curve terms underprices tail risk by several times. A Grid with no Stop loss is selling this kind of insurance without pricing it at all.

Rules you can actually use

1. Size for the gap, not just the stop. Set a ceiling for what you could lose if price jumps, say 5% of the account, then estimate a plausible gap, say 300 pips. On a 10,000-dollar account the maximum size = 500 ÷ (300 × 10) ≈ 0.166 lot. With a 20-pip Stop loss that trade normally risks only 0.166 × 10 × 20 ≈ 33 dollars, or 0.33%, less than the familiar 1%, and that is the price of surviving the rare day
2. Avoid structures with unlimited loss, such as a Grid or Martingale without a basket Stop loss, or selling options with no protective leg. If you can’t say the most you could lose on the worst day, don’t open it
3. Cap effective leverage. From the table, 10× leverage plus a CHF 2015 day is a negative balance. Set the cap from the worst day that has happened in that market, not from a normal day
4. Cut size before likely gaps: holding over the weekend, before big news, before central bank meetings, or in a currency held at a peg. The longer a peg has been held, the harder the jump tends to be when it is released
5. Test your system with real bad days. Insert a 10σ gap into a backtest or a Monte Carlo test and see whether the account survives. If it doesn’t, the system just hasn’t met its day yet

The right question is not “will a day like this happen” but “if it happened tomorrow, could I keep trading?”

To work out the long-run chance of blowing up an account, see Risk of ruin.

Common mistakes

• Treating σ as a ceiling. σ is an average size, not a boundary. 5σ to 10σ days happen in every market
• Measuring σ in a calm period. If the recent past was calm, σ is small and the next explosion will look like “10σ”, when it was the calm that was unusual
• Trusting the win rate. In the simulation above the strategy won 97.7% of days and lost on every path. Win rate tells you nothing about the tails (Win rate vs break-even)
• Thinking a Stop loss caps your loss. It caps it on days with continuous prices. On a day price jumps, it is just a market order
• Using a short backtest as proof of safety. Three years without a black swan doesn’t mean the system can handle one. It just hasn’t met one yet