A normal distribution provides a convenient model of investment returns: most observations sit near the average, and large departures become progressively rarer. Fat tails describe a different pattern, with more probability assigned to extreme outcomes than a comparable normal model would predict.
The distinction matters because a model can describe ordinary trading days reasonably well yet badly misjudge the losses that threaten an account. Within trading probability analysis, the task is not to choose the neatest curve. It is to identify which assumptions remain useful when returns move far from their average.
What a Normal Distribution Assumes About Returns
A normal distribution, also called a Gaussian distribution, is symmetric around its mean. Two parameters describe it: the mean return, μ, and the standard deviation, σ. Standard deviation measures how widely returns vary around that mean.
A standardized return expresses an observation in units of standard deviation:
z = (r − μ) / σ
If the mean daily return is assumed to be zero and daily standard deviation is 1%, a return of −3% has a z-score of −3. That calculation measures distance from the mean. Turning it into a probability requires a distributional assumption.
Under the normal model, the familiar coverage percentages follow the NIST normal distribution probabilities:
| Distance from the mean | Probability inside the interval | Probability outside the interval |
|---|---|---|
| ±1 standard deviation | 68.27% | 31.73% |
| ±2 standard deviations | 95.45% | 4.55% |
| ±3 standard deviations | 99.73% | 0.27% |
The final column combines both tails. The probability of a return below μ − 3σ is approximately 0.135%, not 0.27%.
These are probabilities under the model, not promises about market behavior. Normality alone also says nothing about independence between trading days. That is a separate assumption, required for many calculations involving sequences of returns.
What Fat Tails Mean in Financial Markets
In this context, a fat-tailed return distribution assigns more probability to sufficiently extreme gains or losses than a normal distribution with the same mean and variance. The comparison must hold scale constant. A wider normal distribution is not, just because it is wider, a fat-tailed distribution.
Empirical evidence supports taking this distinction seriously. A study of intraday Dow 30 stock returns from October 2018 to March 2019 found heavy tails and volatility clustering across its sample. The research revisiting financial return patterns also found that not every proposed market pattern held, which cautions against treating all assets and sampling intervals alike.
A deliberately simplified example shows why standard deviation is insufficient. Suppose a hypothetical daily return distribution has three possible outcomes: a 0% return with 96% probability, a +5% return with 2% probability, and a −5% return with 2% probability.
Its mean is zero. Its variance is:
Variance = 0.96 × 02 + 0.02 × 0.052 + 0.02 × (−0.05)2 = 0.0001
The standard deviation is therefore 1%. Yet returns outside ±3% occur with 4% probability, compared with approximately 0.27% under a normal distribution with that same mean and standard deviation.
This is not a proposed market model. It demonstrates that identical averages and volatility can conceal very different extreme outcomes. The bell curve is a benchmark, not a safety certificate.
Fat Tails, Skewness and Kurtosis Are Different
Skewness concerns asymmetry. A distribution with negative skew can have an extended left tail, making large negative outcomes more prominent than large positive ones. Fat tails need not be asymmetric: the hypothetical distribution above is perfectly symmetric.
Kurtosis summarizes the fourth standardized moment, where that moment exists:
Kurtosis = E[(r − μ)4] / σ4
A normal distribution has kurtosis of 3, or excess kurtosis of zero after subtracting 3. The NIST definitions of skewness and kurtosis distinguish these conventions, which matters when comparing software outputs.
Because deviations enter the formula to the fourth power, extreme observations can dominate a sample kurtosis estimate. A high reading warrants investigation, but it does not identify the exact probability of a future crash. Nor does it prove that returns follow a power law or any other particular distribution.
Why a Five-Sigma Move May Expose a Bad Model
A sigma label depends on the volatility estimate used to calculate it. Suppose a model assumes zero mean and estimates daily volatility at 1%. A −5% return is then a five-sigma observation. Against a 3% volatility estimate, the same return is approximately a 1.67-sigma observation.
Neither calculation changes the loss. They change its apparent statistical rarity.
Changing volatility can itself produce fat tails in a pooled return sample. Combining calm and turbulent periods creates a mixture that need not resemble a single normal distribution, even if returns within each state are conditionally normal. Federal Reserve research on changing volatility and return tails provides evidence for this mechanism in high-frequency stock returns.
Volatility clustering adds a time dimension: large absolute returns can occur in groups rather than being scattered independently through the sample. A model that treats every day as an independent draw from one fixed distribution misses that structure.
This does not justify raising the volatility estimate after every loss and declaring the model repaired. The useful test is whether estimates made before each observation produce sensible forecasts afterward. Otherwise, a model can explain every surprise without having anticipated any of them.
How to Assess Whether Returns Have Fat Tails
Begin by defining the observations. Daily asset returns, hourly returns and completed trade profits are different datasets. A strategy’s payoff distribution also reflects its entry rules, exits and exposure, not just the asset’s price changes.
Keep the return definition consistent. Simple and logarithmic returns behave differently when compounded, but taking logarithms does not automatically make returns normal.
A histogram offers a starting point, although its bins can hide sparse extreme observations. A normal quantile plot, often called a Q–Q plot, compares ordered sample returns with theoretical normal quantiles. With sample returns on the vertical axis, a roughly straight center accompanied by a lower tail below the line and an upper tail above it suggests heavier tails.
Count threshold exceedances too. Compare the observed frequency beyond three estimated standard deviations with the normal benchmark, while recognizing that parameters estimated from the same sample introduce uncertainty.
Finally, examine returns through time. Compare calm and stressed periods and inspect whether large absolute returns cluster. A single pooled chart discards the order of observations.
Small samples create a stubborn problem: they contain little evidence about rare events. A record with no severe losses does not establish that severe losses are impossible. Treat apparent outliers carefully; correct data errors, but do not delete genuine losses simply because they spoil the fit.
Choosing Models Without Pretending the Tail Is Known
A normal model remains useful as a transparent baseline. The question is whether another approach improves the decisions that depend on extreme outcomes, rather than just improving the fit near the center.
A Student’s t distribution offers heavier tails, controlled by its degrees-of-freedom parameter. It is symmetric unless extended, so it cannot by itself represent an asymmetric loss distribution. Its scale parameter is also not its standard deviation. For degrees of freedom ν greater than 2, variance equals s2ν/(ν − 2), where s is the scale parameter. Comparisons should match variance where it exists.
Historical simulation instead uses observed returns without imposing a normal curve. Its weakness is the historical sample: ordinary resampling cannot generate a one-period return more extreme than its most extreme observation.
Models with changing volatility address variation through time. They can also use fat-tailed shocks, because changing volatility and non-normal shocks are separate modeling choices.
These choices carry directly into Monte Carlo simulations of returns and drawdowns. Independent normal draws, resampled historical returns and volatility-dependent draws answer different questions. More simulations reduce simulation noise; they do not repair an unsuitable return model.
Compare plausible models alongside explicit stress scenarios. If an exposure looks acceptable only under the thinnest assumed tail, that dependence deserves attention before the exposure is increased.
Extreme Price Moves Are Not the Same as Account Losses
A return distribution describes changes in an asset or strategy. Account losses also depend on the amount exposed and how positions can be closed.
Consider a hypothetical $100,000 account holding $300,000 of stock exposure financed partly through borrowing. A 5% decline produces a $15,000 loss before financing and transaction costs: 15% of account equity. A statistical description of the asset’s move does not replace this exposure calculation.
Exit assumptions matter as well. For U.S. stock stop orders, the stop price triggers a market order rather than guaranteeing an execution price. A stop-limit order constrains the price but may remain unfilled. These distinctions appear in the SEC investor bulletin on stop orders.
A hypothetical backtest that caps every loss at 2% assumes away any worse execution. If that cap is not enforceable under the strategy’s actual trading conditions, the simulated left tail can be too short even when the underlying price model is reasonable.
Stress testing should therefore translate asset moves into account losses, including execution assumptions. Testing a price shock without testing the position that absorbs it leaves the practical question unanswered.
Measure Loss Severity Beyond a Probability Threshold
A tail probability describes how often a threshold might be crossed. It does not describe how severe losses become after that crossing.
Value at Risk, or VaR, identifies a loss quantile over a stated horizon. A hypothetical one-day 99% VaR of $10,000 means the model assigns approximately 1% probability to a greater loss, assuming a continuous loss distribution. It does not mean $10,000 is the maximum possible loss.
Expected shortfall looks beyond that threshold. For a continuous loss distribution, 99% expected shortfall is the average loss within the worst 1% of outcomes. The Basel Committee’s explanation of tail risk measures highlights this distinction between a threshold and the severity of losses beyond it.
Expected shortfall still depends on the assumed tail and the data used to estimate it. Replacing one risk measure with another does not remove model uncertainty.
The practical response is to connect distribution analysis with position sizing and capital allocation. Compare ordinary volatility, stressed volatility and losses beyond the historical record. Check whether the account could withstand those scenarios without relying on an exact forecast of their probability.
The aim is not to predict the next extreme move precisely. It is to avoid making survival depend on a normality assumption that has not earned that responsibility.