Heavy-tailed distributions form a broad class of probability distributions that contrast sharply with light-tailed distributions—the family of distributions whose tails decay at least as fast as an exponential or Gaussian distribution. In heavy-tailed distributions, tail probabilities decay much more slowly, giving extreme events a non-negligible probability of occurring.
Defining Heavy Tails
A random variable with cumulative distribution function and survival function is said to have a heavy right tail if its moment generating function is infinite for all :
Equivalently, the tail decays slower than any exponential distribution .
Heavy Tails vs. Fat Tails
While the terms are sometimes used interchangeably, in quantitative finance and extreme value theory a useful distinction is made:
- Heavy-Tailed Distributions (Sub-exponential decay): Any distribution whose tail is not exponentially bounded (e.g., Log-normal, Weibull with shape , Pareto).
- Fat-Tailed Distributions (Power-law decay): A stricter subset of heavy-tailed distributions whose survival function decays asymptotically as a power law (regularly varying tail):
where is a slowly varying function and is the tail index. For a fat-tailed distribution, all moments of order are infinite:
For example, the Log-normal distribution is heavy-tailed (slower than exponential) but not fat-tailed because all of its moments are finite. By contrast, the Student- (with degrees of freedom), Cauchy (), and Pareto distributions are genuinely fat-tailed.
Central Limit Theorem Implications
When the tail index satisfies , the variance is infinite () or borderline, and the classical Central Limit Theorem no longer guarantees rapid convergence of sample means to a Gaussian distribution:
- Slow Convergence of Sample Moments: Even when (finite variance, infinite kurtosis), sample variance and empirical risk metrics converge extremely slowly and remain dominated by single extreme observations.
- Generalized Central Limit Theorem: For , properly normalized sums of i.i.d. random variables converge not to a Normal distribution, but to an -stable (Lévy stable) distribution.
- Risk Modeling Takeaway: Assuming Gaussianity in financial returns or energy price spikes leads to severe underestimation of tail events ( moves).