Common Probability Distributions in Finance › Fat-Tailed Distributions

A normal distribution is only one possible model for returns. When an observed or simulated return distribution places more probability in sufficiently remote tails than a fitted normal model, normal tail calculations can misstate extreme-event frequencies and downstream risk measures. The examples below use a seeded Student's t sample to show how to diagnose that mismatch without presenting synthetic observations as market data.

Here, "fat tails" means slower remote-tail decay than a specified normal benchmark. The comparison requires a common location/scale convention and sufficiently large thresholds; it does not order every finite-threshold probability. For models without a finite mean, tail behavior must be described relative to another location parameter rather than an undefined mean. Two useful diagnostics follow:

Kurtosis: The normal distribution has a kurtosis of 3 (or excess kurtosis of 0). Kurtosis is the fourth central moment divided by the variance squared. A higher value signals more weight on large standardized deviations overall, but it does not by itself order every tail probability or identify where probability moved between the center and shoulders.

Tail exponents: For power-law-tailed distributions, the probability of extreme events decays polynomially rather than exponentially. The tail exponent $\alpha$ characterizes how fast this decay occurs. A smaller $\alpha$ means slower polynomial decay and fatter tails, while a larger $\alpha$ means faster polynomial decay and thinner tails within the power-law family. For every finite $\alpha$, however, the asymptotic tail remains heavier than a normal tail.

The seeded sample was generated from a heavy-tailed distribution, so its departure from the normal benchmark is expected rather than empirical evidence about markets. For this seed, absolute moves beyond four sample standard deviations occur 30 times, compared with about 0.38 expected in 6,000 independent normal draws, a ratio of about 79. Nine observations exceed five sample standard deviations, compared with about 0.0034 under that same normal benchmark. The exercise demonstrates how a known heavy-tailed data-generating process changes extreme-event counts; it does not estimate their frequency in an observed market.

The Student's t distribution provides a simple way to model fat tails while retaining many convenient properties of the normal distribution. William Sealy Gosset published small-sample work under the pseudonym "Student" while working as an experimental brewer at Guinness. In finance, a location-scale t model is one candidate for forecasting or fitting heavy-tailed returns; its suitability still needs to be checked.

The probability density function of the standard t distribution with $\nu>0$ degrees of freedom is:

$$f(x) = \frac{\Gamma\left(\frac{\nu+1}{2}\right)}{\sqrt{\nu\pi}\,\Gamma\left(\frac{\nu}{2}\right)} \left(1 + \frac{x^2}{\nu}\right)^{-\frac{\nu+1}{2}}$$

This formula may look complex, but its structure shows the key difference between the t distribution and the normal distribution. While the normal distribution uses the exponential function $e^{-x^2/2}$ which decays very rapidly, the t distribution uses a power function $(1 + x^2/\nu)^{-(\nu+1)/2}$ which decays much more slowly. This slower decay is precisely what produces the heavier tails.

Each component:

• $x$: the value at which we evaluate the density
• $\nu$: degrees of freedom parameter (controls tail thickness; lower $\nu$ means fatter tails)
• $\Gamma(\cdot)$: the gamma function, generalizing the factorial to non-integer values ($\Gamma(n) = (n-1)!$ for positive integers)
• $\frac{\Gamma\left(\frac{\nu+1}{2}\right)}{\sqrt{\nu\pi}\,\Gamma\left(\frac{\nu}{2}\right)}$: a normalization constant ensuring the density integrates to 1
• $(1 + x^2/\nu)^{-(\nu+1)/2}$: the core term that produces polynomial rather than exponential tail decay

The polynomial decay $(1 + x^2/\nu)^{-(\nu+1)/2}$ versus the normal's exponential decay $e^{-x^2/2}$ is why the t distribution has heavier tails. To understand this intuitively, consider what happens as $x$ becomes large. For the normal distribution, doubling $x$ quadruples the magnitude of the exponent because $(2x)^2 = 4x^2$, so the density decays extremely rapidly. For the t distribution, the density is asymptotically proportional to $|x|^{-(\nu+1)}$, so doubling a large threshold reduces the density by about $2^{-(\nu+1)}$. Its one-sided survival probability is proportional to $x^{-\nu}$, so doubling a large threshold reduces that probability by about $2^{-\nu}$. As $\nu \to \infty$, the polynomial term converges to the exponential, and the t distribution approaches the normal.

Key properties of the t distribution:

• Degrees of freedom ($\nu$): Controls tail thickness. Lower values mean fatter tails.
• Convergence: As $\nu \to \infty$, the t distribution approaches the standard normal.
• Variance: For $\nu > 2$, the variance is $\nu/(\nu-2)$, which is always greater than 1. For $1<\nu\leq2$, the mean exists but the variance is infinite. For $0<\nu\leq1$, the mean and therefore the variance are undefined. The second raw moment diverges throughout $0<\nu\leq2$.
• Kurtosis: For $\nu > 4$, excess kurtosis is $6/(\nu-4)$, always positive and decreasing in $\nu$. For $\nu \leq 4$, the kurtosis is undefined because the fourth moment does not exist.

When fitting returns, the degrees of freedom must be estimated or otherwise justified for the dataset and model; there is no universal range for financial returns. Values near 4 illustrate the family's moment boundary: for $0<\nu\leq4$, the fourth moment diverges, while for $\nu>4$ it is finite. The seeded example uses $\nu=4$ deliberately.

The log-scale plot shows density heights, not probabilities beyond a threshold. At $x=5$, the unit-variance t distribution with 3 degrees of freedom has about 633 times the normal density. Integrating beyond 5 gives a different comparison: its one-sided survival probability is about 5,650 times the normal survival probability. The normal probability is about 1 in 3.5 million observations; the scaled t probability is about 1 in 600. For independent daily draws and 252 trading days per year, the expected waiting times are roughly 13,900 trading years and 2.4 trading years. These are model-based averages, not event schedules.

We can fit a location-scale t distribution to the seeded heavy-tailed sample and compare the log-likelihoods evaluated at the returned numerical fits. This does not establish a global maximum for the unrestricted t likelihood. Because the sample was generated from a t distribution, it is a reproducibility check and model-comparison demonstration, not evidence about observed market returns. For $\nu>2$, the t standard deviation is $s\sqrt{\nu/(\nu-2)}$, where $s$ is its scale parameter; scale itself is not the standard deviation. For $0<\nu\leq2$, no finite standard deviation exists.

Power-law notation provides a general way to describe polynomial tail decay. Student's t distributions already have power-law tails, but other regularly varying models can use a different exponent or slowly varying factor:

$$P(X > x) \sim Cx^{-\alpha} \quad \text{as } x \to \infty, \qquad C>0,\ \alpha>0$$

This asymptotic relationship describes the right-tail survival function for large thresholds. It includes the polynomial right tail of a Student's t distribution as a special case. For a nonnegative variable, the exponent controls positive moments; for a signed variable, existence of a full moment also depends on the left tail, which this one-sided relationship does not specify.

The components of this relationship are:

• $P(X > x)$: the probability that the random variable exceeds value $x$ (the survival function)
• $x$: a large threshold value
• $\alpha$: the tail exponent, which controls how fast probabilities decay. Some authors also call $\alpha$ a tail index; in another common extreme-value-theory convention, the tail index is $\xi=1/\alpha$
• $\sim$: denotes asymptotic equivalence. The ratio of the left side to $Cx^{-\alpha}$ approaches 1 as $x \to \infty$

A larger $\alpha$ means faster polynomial decay and a smaller $\alpha$ means slower decay. As an illustration, if $\alpha=3$ and the asymptotic approximation is appropriate at both thresholds, doubling a sufficiently large threshold reduces its exceedance probability by approximately $2^{-3}=1/8$. This is much slower than the exponential falloff of normal tails.

For comparison, the one-sided standard-normal survival probability falls by a factor of about $5.1\times 10^{10}$ from 4σ to 8σ, roughly eleven orders of magnitude. Under the illustrative cubic law, the corresponding asymptotic reduction is only a factor of 8. Which approximation is appropriate must be established from a specified model or dataset rather than inferred from this contrast alone.

The implications of power-law tails are serious for risk management:

• Moment thresholds require both tails: For $p>0$ and a nonnegative variable with $P(X>x)\sim Cx^{-\alpha}$, the right-tail contribution to $E[X^p]$ is finite for $p<\alpha$ and diverges for $p\geq\alpha$. For a signed return, the displayed right-tail relation alone cannot guarantee a full mean or variance because the left tail is unspecified. If both tails have such power-law asymptotics with exponents $\alpha_L$ and $\alpha_R$, absolute moments exist for $p<\min(\alpha_L,\alpha_R)$ and diverge at or above that minimum. In particular, when both exponents lie strictly between 2 and 4, the mean and variance are finite, while the fourth moment is not.
• Slow convergence: Under i.i.d. sampling with finite variance ($\alpha > 2$ in this tail model), the Central Limit Theorem applies, but its normal approximation can remain poor at practically available sample sizes.
• Tail concentration and dependence: With power-law marginal tails, a small number of observations can dominate a sample risk estimate. Diversification may be weaker when those tails combine with strong common or tail dependence, but marginal tail thickness alone does not determine diversification benefits.

Fat tails change how we should think about risk:

VaR model error: A fitted normal model can underestimate sufficiently extreme heavy left-tail quantiles, but it need not underestimate VaR at every confidence level. The direction and magnitude depend on the fitted distributions, skewness, and confidence level. The seeded example below slightly overestimates 95% VaR and underestimates 99% VaR.

Expected shortfall: Under the continuous-loss convention used here, expected shortfall is the mean loss conditional on being at or beyond the VaR threshold. Terminology such as "Conditional VaR" varies across sources, and distributions with atoms require a quantile-integral definition to handle the boundary consistently. For continuous heavy-tailed models, ES can be substantially larger than VaR.

The code treats the sample as an empirical distribution with equal probability on each observation. VaR is its inverse-CDF loss quantile. ES averages exactly the worst $1-\alpha$ probability mass, using a fractional boundary weight when necessary rather than including every loss tied at VaR. For example, losses $0,0,1$ at 50% confidence give VaR $0$ and ES $2/3$, not the full-sample mean $1/3$. Finite-sample VaR depends on the quantile convention; the inverse-CDF convention here differs from a linearly interpolated percentile. The ratio ES/VaR is undefined when VaR is zero.