Hubbry Logo
Fat-tailed distributionFat-tailed distributionMain
Open search
Fat-tailed distribution
Community hub
Fat-tailed distribution
logo
7 pages, 0 posts
0 subscribers
Be the first to start a discussion here.
Be the first to start a discussion here.
Fat-tailed distribution
Fat-tailed distribution
Fat-tailed distribution
View on Wikipedia
from Wikipedia

A fat-tailed distribution is a probability distribution that exhibits a large skewness or kurtosis, relative to that of either a normal distribution or an exponential distribution.[when defined as?] In common usage, the terms fat-tailed and heavy-tailed are sometimes synonymous; fat-tailed is sometimes also defined as a subset of heavy-tailed. Different research communities favor one or the other largely for historical reasons, and may have differences in the precise definition of either.

Fat-tailed distributions have been empirically encountered in a variety of areas: physics, earth sciences, economics and political science. The class of fat-tailed distributions includes those whose tails decay like a power law, which is a common point of reference in their use in the scientific literature. However, fat-tailed distributions also include other slowly-decaying distributions, such as the log-normal.[1]

The extreme case: a power-law distribution

[edit]

The most extreme case of a fat tail is given by a distribution whose tail decays like a power law.

The Cauchy Distribution
A variety of Cauchy distributions for various location and scale parameters. Cauchy distributions are examples of fat-tailed distributions.

That is, if the complementary cumulative distribution of a random variable X can be expressed as[citation needed]

as for

then the distribution is said to have a fat tail if . For such values the variance and the skewness of the tail are mathematically undefined (a special property of the power-law distribution), and hence larger than any normal or exponential distribution. For values of the claim of a fat tail is more ambiguous, because in this parameter range, the variance, skewness, and kurtosis can be finite, depending on the precise value of and thus potentially smaller than a high-variance normal or exponential tail. This ambiguity often leads to disagreements about precisely what is, or is not, a fat-tailed distribution. For the moment is infinite, so for every power law distribution, some moments are undefined.[2]

Note
Here the tilde notation "" means that the tail of the distribution decays like a power law; more technically, it refers to the asymptotic equivalence of functions – meaning that their ratio asymptotically tends to a constant.[citation needed]

Fat tails and risk estimate distortions

[edit]
Lévy flight from a Cauchy distribution compared to Brownian motion (below). Central events are more common and rare events more extreme in the Cauchy distribution than in Brownian motion. A single event may comprise 99% of total variation, hence the "undefined variance".
Lévy flight from a normal distribution (Brownian motion).

Compared to fat-tailed distributions, in the normal distribution, events that deviate from the mean by five or more standard deviations ("5-sigma events") have lower probability, meaning that in the normal distribution extreme events are less likely than for fat-tailed distributions. Fat-tailed distributions such as the Cauchy distribution (and all other stable distributions with the exception of the normal distribution) have "undefined sigma" (more technically, the variance is undefined).

As a consequence, when data arise from an underlying fat-tailed distribution, shoehorning in the "normal distribution" model of risk—and estimating sigma based (necessarily) on a finite sample size—would understate the true degree of predictive difficulty (and of risk). Many—notably Benoît Mandelbrot as well as Nassim Taleb—have noted this shortcoming of the normal distribution model and have proposed that fat-tailed distributions such as the stable distributions govern asset returns frequently found in finance.[3][4][5]

The Black–Scholes model of option pricing is based on a normal distribution. If the distribution is actually a fat-tailed one, then the model will under-price options that are far out of the money, since a 5- or 7-sigma event is much more likely than the normal distribution would predict.[6]

Applications in economics

[edit]

In finance, fat tails often occur but are considered undesirable because of the additional risk they imply. For example, an investment strategy may have an expected return, after one year, that is five times its standard deviation. Assuming a normal distribution, the likelihood of its failure (negative return) is less than one in a million; in practice, it may be higher. Normal distributions that emerge in finance generally do so because the factors influencing an asset's value or price are mathematically "well-behaved", and the central limit theorem provides for such a distribution. However, traumatic "real-world" events (such as an oil shock, a large corporate bankruptcy, or an abrupt change in a political situation) are usually not mathematically well-behaved.

Historical examples include the Wall Street crash of 1929, Black Monday (1987), Dot-com bubble, 2008 financial crisis, 2010 flash crash, the 2020 stock market crash and the unpegging of some currencies.[7]

Fat tails in market return distributions also have some behavioral origins (investor excessive optimism or pessimism leading to large market moves) and are therefore studied in behavioral finance.

In marketing, the familiar 80-20 rule frequently found (e.g. "20% of customers account for 80% of the revenue") is a manifestation of a fat tail distribution underlying the data.[8]

The "fat tails" are also observed in the record industry, especially in phonographic markets. The probability density function for logarithm of weekly record sales changes is highly leptokurtic and characterized by a narrower and larger maximum, and by a fatter tail than in the normal distribution case. On the other hand, this distribution has only one fat tail associated with an increase in sales due to promotion of the new records that enter the charts.[9]

See also

[edit]

References

[edit]
[edit]
Revisions and contributorsEdit on WikipediaRead on Wikipedia

Fat-tailed distribution

View on Grokipedia
from Grokipedia
A fat-tailed distribution, also known as a , is a in statistics and characterized by tails that decay more slowly than an exponential rate, resulting in a higher likelihood of extreme values or outliers compared to light-tailed distributions like the normal distribution. This slower decay often follows a power-law form, where the survival function Fˉ(x)∼x−αL(x)\bar{F}(x) \sim x^{-\alpha} L(x) for large xx, with α>0\alpha > 0 as the tail index and L(x)L(x) a slowly varying function. Key properties of fat-tailed distributions include regular variation in the tails, subexponentiality (where the tail of the sum of independent variables behaves like the tail of the maximum), and the potential for infinite moments such as variance when α<2\alpha < 2 or even the mean when α≤1\alpha \leq 1. These distributions exhibit scale invariance, meaning rescaling the variable proportionally scales the tail probabilities, and they often display higher kurtosis, reflecting greater peakedness and tail heaviness relative to the normal distribution. In contrast to exponential-tailed distributions, where the conditional mean exceedance remains constant, fat-tailed ones show an increasing conditional mean exceedance for large thresholds, indicating escalating risk of extremes. Common examples include the Pareto distribution, with survival function Fˉ(x)=(xm/x)α\bar{F}(x) = (x_m / x)^\alpha for x≥xm>0x \geq x_m > 0, which models phenomena like wealth inequality and city sizes; the , a with α=1\alpha = 1 and no defined mean or variance; and the with low , widely used in robust statistics. Other instances are the (with heavier right tail) and Lévy-stable distributions, which capture asymmetric extremes. These distributions arise in diverse applications, such as financial returns (where they explain market crashes), insurance claims for , network traffic modeling, and biological processes like species abundance. Historically, the study of fat-tailed distributions traces back to Vilfredo Pareto's 1896 observation of power-law wealth distributions in , later generalized by Maurice Fréchet and in during the early . extended their relevance to in the 1960s, challenging Gaussian assumptions by demonstrating power-law behaviors in cotton prices and stock returns, influencing modern and geometry applications.

Fundamentals

Definition

A is a in which the probability mass in the tails—corresponding to extreme deviations from the —is greater than that of a , resulting in a higher likelihood of rare but large events occurring. These distributions exhibit slower decay in the tails compared to the seen in thin-tailed distributions like the Gaussian, leading to more frequent outliers that can significantly impact expectations and risk assessments. The term "fat-tailed distribution" gained prominence in the through the work of mathematician Benoît Mandelbrot, who analyzed speculative prices and argued that financial variations followed distributions with heavier tails than the Gaussian assumption prevalent in statistics at the time, such as those resembling Pareto laws rather than . Mandelbrot's analysis of cotton prices from 1816 to 1940 highlighted how such distributions better captured the empirical reality of large fluctuations, challenging the adequacy of normal models for . In real-world contexts, fat-tailed distributions manifest in phenomena like income inequality, where a small number of individuals hold disproportionately large shares of , far exceeding what a would predict for extreme values. Similarly, stock market crashes illustrate this property, as extreme price drops occur more often than expected under Gaussian assumptions, underscoring the role of fat tails in financial volatility. The tails of a fat-tailed distribution can be understood through the cumulative distribution function (CDF), which describes the probability that a random variable falls below a given value; in these cases, the CDF approaches 1 (or 0 for the lower tail) more gradually for large values, reflecting elevated probabilities in the extremes without implying infinite support.

Mathematical Characterization

A fat-tailed distribution is mathematically characterized by the behavior of its survival function, denoted as Fˉ(x)=P(X>x)\bar{F}(x) = P(X > x) or more generally P(∣X∣>x)P(|X| > x) for symmetric cases, which exhibits slower decay than exponential rates typical of thin-tailed distributions. Specifically, the tails are said to be fat if, for large xx, P(∣X∣>x)∼c/xαP(|X| > x) \sim c / x^\alpha where c>0c > 0 is a constant and α>0\alpha > 0 is the tail index, a parameter that quantifies the heaviness of the tails; smaller values of α\alpha indicate heavier tails, with α<2\alpha < 2 often marking distributions where the variance is infinite. This asymptotic approximation is formalized through the concept of regular variation, where the tail function Fˉ(x)\bar{F}(x) is regularly varying at infinity with index −α-\alpha if it satisfies lim⁡t→∞Fˉ(tx)Fˉ(x)=t−α\lim_{t \to \infty} \frac{\bar{F}(tx)}{\bar{F}(x)} = t^{-\alpha} for all t>0t > 0. This limit condition captures the power-law decay precisely and is a cornerstone for analyzing extreme value behavior in fat-tailed models. Fat-tailed distributions often belong to the broader class of subexponential distributions, defined by the property that for independent and identically distributed random variables X1,…,XnX_1, \dots, X_n with common distribution FF, the tail of their sum satisfies Fˉn(x)∼nFˉ(x)\bar{F}_n(x) \sim n \bar{F}(x) as x→∞x \to \infty, where Fˉn\bar{F}_n is the survival function of the sum; this implies that large deviations are dominated by the largest single term rather than collective moderate ones. Such distributions encompass regularly varying tails but extend to other slowly decaying forms, providing a unified framework for tail risk assessment.

Properties

Moments and Kurtosis

In fat-tailed distributions, particularly those exhibiting regularly varying tails with index −α-\alpha where α>0\alpha > 0, the existence of statistical moments is governed by the tail index α\alpha. The kk-th absolute moment E[∣X∣k]E[|X|^k] is finite if and only if α>k\alpha > k. For instance, when α≤2\alpha \leq 2, the variance is infinite, rendering second-order statistics like standard deviation undefined or unreliable for inference. This condition arises because the tail probability Fˉ(x)=P(X>x)\bar{F}(x) = P(X > x) decays as x−αL(x)x^{-\alpha} L(x), where L(x)L(x) is a slowly varying function, leading to divergence of the moment integral for α≤k\alpha \leq k. A precise characterization of moment existence for such distributions is given by the tail integral condition: the kk-th moment is finite if ∫x∞tk dF(t)<∞\int_x^\infty t^k \, dF(t) < \infty, which holds precisely when α>k\alpha > k. Equivalently, for nonnegative random variables, E[Xk]=k∫0∞xk−1Fˉ(x) dx<∞E[X^k] = k \int_0^\infty x^{k-1} \bar{F}(x) \, dx < \infty under the same tail index requirement, as established by Karamata's theorem for regularly varying functions. These properties highlight why fat-tailed models, common in financial returns and insurance claims, often lack finite higher moments, complicating traditional parametric assumptions. Kurtosis, a measure of tail heaviness, is defined for distributions with finite fourth moments as κ=E[(X−μ)4]σ4\kappa = \frac{E[(X - \mu)^4]}{\sigma^4}, with excess kurtosis given by κ−3\kappa - 3. Values of excess kurtosis greater than 0 (i.e., κ>3\kappa > 3) indicate leptokurtosis, characteristic of fat-tailed distributions where extreme deviations are more probable than under normality. However, kurtosis is undefined when the variance is infinite (α≤2\alpha \leq 2), as the normalizing denominator σ4\sigma^4 does not exist, and even when α>4\alpha > 4 (ensuring finite fourth moments), it may not fully capture tail behavior in subexponential classes. In practice, computing empirical from finite samples systematically underestimates the true tail fatness of such distributions, as extreme events occur infrequently and dominate higher moments but are unlikely to be observed in limited data. For example, in power-law ed data with α≈3\alpha \approx 3, sample kurtosis converges slowly, often requiring impractically large datasets (e.g., 101110^{11} observations) to approximate theoretical values, leading to apparent near-normality in short histories. This bias arises from the rarity of tail realizations, exacerbating errors in models reliant on historical moments.

Tail Decay Behavior

Fat-tailed distributions exhibit tail decay that occurs more slowly than the characteristic of thin-tailed distributions such as or exponential. Instead, their tails often follow a power-law or form, where the probability of extreme values diminishes gradually, increasing the likelihood of outliers relative to exponential models. This qualitative difference underscores why fat tails are associated with higher risks in domains like and , as retain substantial probability mass far into the tails. A key visual distinction arises in log-log plots of the (complementary ), where power-law tails manifest as straight lines with a constant negative , reflecting the consistent rate of decay. In contrast, exponential tails curve downward more sharply on the same scale, highlighting the slower, more persistent decay of fat-tailed structures. These plots provide an intuitive graphical method for detecting power-law behavior without relying on parametric assumptions. Many fat-tailed distributions possess the subexponential property, meaning that for large thresholds, the tail probability of the sum of independent variables is dominated by the largest single variable rather than the collective contribution of all. This "catastrophe principle" implies that extreme outcomes in aggregates, such as portfolio returns or aggregate losses, are driven primarily by the most severe individual event, diverging from the averaging behavior in light-tailed cases. Graphical tools like quantile-quantile (Q-Q) plots and plots further aid in identifying fat tails by comparing empirical data against reference distributions. In Q-Q plots, fat-tailed data show upward deviations in the extremes compared to normal quantiles, while on a log-log scale reveal straight-line linearity for power-law tails or persistent high probabilities beyond exponential expectations. These visualizations emphasize the qualitative "heaviness" without requiring moment computations. Tail heaviness in fat-tailed distributions is qualitatively classified into heavy-tailed, where lower-order moments like the may exist but higher ones such as variance do not, and super-heavy-tailed, where all moments are infinite due to even slower decay. Super-heavy tails amplify the dominance of extremes, making traditional statistical summaries unreliable, while heavy tails still allow partial moment-based . This classification highlights varying degrees of tail persistence across applications.

Examples

Power-Law Distributions

Power-law distributions represent the archetypal example of fat-tailed distributions, characterized by tails that decay polynomially rather than exponentially, leading to a higher probability of extreme events compared to distributions with thinner tails. In these distributions, the probability of observing a value xx scales as P(X>x)∝x−αP(X > x) \propto x^{-\alpha} for large xx, where α>0\alpha > 0 is the tail index that governs the heaviness of the tail: smaller values of α\alpha indicate heavier tails, with α≤2\alpha \leq 2 often resulting in infinite variance and α<1\alpha < 1 yielding an infinite mean. This polynomial decay distinguishes power-laws from thinner-tailed alternatives and makes them prevalent in natural and social phenomena exhibiting scale-free behavior. The continuous form of power-law distributions is exemplified by the Pareto distribution, named after economist Vilfredo Pareto who observed it in income data. The PDF of the Pareto distribution (Type I) is f(x)=αxmαxα+1,x≥xm>0,f(x) = \frac{\alpha x_m^\alpha}{x^{\alpha+1}}, \quad x \geq x_m > 0, where xmx_m is the minimum value () and α\alpha is the shape parameter or tail index. The corresponding (CDF) is F(x)=1−(xmx)α,x≥xm,F(x) = 1 - \left( \frac{x_m}{x} \right)^\alpha, \quad x \geq x_m, which directly shows the power-law tail behavior in the survival function 1−F(x)∝x−α1 - F(x) \propto x^{-\alpha}. The tail index α\alpha critically determines the distribution's heaviness; for instance, if α<1\alpha < 1, the expected value is infinite, amplifying the impact of rare large events. In discrete settings, power-law distributions manifest as Zipf's law, first empirically described by linguist George Zipf in word frequencies and city sizes. The probability mass function takes the form P(X=k)∝1/kα+1P(X = k) \propto 1/k^{\alpha+1} for positive integers k≥1k \geq 1, where the exponent α+1\alpha + 1 (often denoted as ss) typically ranges from 1 to 2 in empirical data. This discrete power-law is commonly applied in rankings, such as the frequency-rank plots of word occurrences in texts or website traffic, where the most frequent item appears roughly twice as often as the second, and so on. To simulate samples from a power-law distribution, the transformation method leverages the inverse CDF: generate a uniform random variable U∼Uniform(0,1)U \sim \text{Uniform}(0,1), then set X=xm(1−U)−1/αX = x_m (1 - U)^{-1/\alpha} for the Pareto case, which ensures the generated values follow the desired power-law tail. This approach is efficient and exact for continuous power-laws, though for discrete variants like Zipf's, a similar inversion or rejection sampling may be used to approximate the proportionality.

Other Common Distributions

The Student's t-distribution is a symmetric fat-tailed distribution commonly used in statistical modeling, particularly for small-sample inference and robust estimation where data may exhibit heavier tails than the . Its probability density function is given by f(x;ν)=Γ(ν+12)νπ Γ(ν2)(1+x2ν)−ν+12,f(x; \nu) = \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}},
Add your contribution
Related Hubs
User Avatar
No comments yet.