Choosing Extreme-Value Models from Tail Behavior

Why the Tail Guides the Choice

The largest observations in a large sample come from the far tail of the distribution, not from its center. Consequently, a familiar-looking density or a finite-sample histogram may say little about which model describes extreme maxima. The relevant clues are how quickly the upper-tail probability decays and whether the distribution has a finite upper endpoint.

For independent observations with a common distribution, three broad tail patterns lead to three extreme-value families: exponential-type decay is associated with the Gumbel family, regularly varying heavy tails with the Fréchet family, and sufficiently regular behavior near a finite upper endpoint with the reverse-Weibull family. These are domain-of-attraction statements: they concern limits after suitable centering and scaling, not an assertion that the original observations themselves follow one of the limiting distributions.

In the generalized extreme-value (GEV) parameterization, the shape parameter \(\xi\) distinguishes the three types: \(\xi=0\) is Gumbel, \(\xi>0\) is Fréchet-type, and \(\xi<0\) is reverse-Weibull-type. Tail behavior helps choose among these possibilities, but the word “type” is important: a limit model describes normalized maxima in a large-sample approximation.

A Tail-Probability Principle for Maxima

Write \(\overline F(u)=1-F(u)\) for the upper-tail probability. The exact distribution of a maximum of independent observations is \(P(M_n\le u)=F(u)^n\). Thus, the chance that one observation exceeds a high threshold controls the chance that the maximum does not.

Proof. Set \(q_n=\overline F(u_n(x))\) for a fixed \(x\). The assumed convergence implies \(nq_n\to\tau(x)\), so \(q_n\to0\). Independence gives \(P(M_n\le u_n(x))=(1-q_n)^n\). For \(0\le q<1\), \(\log(1-q)=-q+O(q^2)\) as \(q\to0\). Moreover, \(nq_n^2=(nq_n)q_n\to0\), since \(nq_n\) converges and \(q_n\to0\). Therefore \[ n\log(1-q_n)=-nq_n+O(nq_n^2)\longrightarrow-\tau(x). \] Exponentiating proves the result. \(\square\)

This principle gives a practical route to a limit: select thresholds for which \(n\overline F(u_n(x))\) has a nontrivial limit. The limiting maximum distribution then follows without approximating the whole distribution \(F\).

Indeed, multiplying the two tail relations gives \(n\overline F(b_n+a_nx)\to e^{-x}\). Apply the theorem with \(u_n(x)=b_n+a_nx\). The condition says that, on the scale \(a_n\), moving the high threshold by \(a_nx\) changes the tail by an asymptotic factor \(e^{-x}\).

Exponential-Type Tails: The Gumbel Choice

An exponential-type tail decreases rapidly without ending at a finite upper bound. The Gumbel family is a natural candidate because suitable small shifts of a high quantile produce the tail-ratio condition above. A useful sufficient pattern is \(\overline F(x)\sim c x^\beta\exp(-\lambda x^p)\) as \(x\to\infty\), with \(c,\lambda,p>0\) and \(\beta\) real. This pattern is not a definition of the entire Gumbel domain; it is a tractable sufficient case.

For a high level \(b_n\) chosen so \(n\overline F(b_n)\to1\), use \(a_n=(\lambda p b_n^{p-1})^{-1}\). For fixed \(x\), \(a_n/b_n\to0\), and expansion of the power gives \[ \lambda\big((b_n+a_nx)^p-b_n^p\big)\longrightarrow x. \] Also, \((b_n+a_nx)^\beta/b_n^\beta\to1\). Taking the ratio of the proposed tail forms therefore gives \(\overline F(b_n+a_nx)/\overline F(b_n)\to e^{-x}\). The corollary yields a Gumbel limit whenever the stated tail approximation and quantile condition hold.

Normal and lognormal distributions are also standard examples associated with Gumbel limits, although their tail calculations differ from the exponential example. “Gumbel” should not be read as “the original data have an exponential distribution”: it is the limiting family for normalized maxima.

Heavy Tails: The Fréchet Choice

A common mathematical description of a heavy upper tail is regular variation: \(\overline F(x)=x^{-\alpha}L(x)\), where \(\alpha>0\) and \(L\) is slowly varying, meaning \(L(tx)/L(x)\to1\) for each fixed \(t>0\). The tail then has a power-law component; the factor \(L\) can vary slowly but does not change the principal scaling exponent.

Proof. For each fixed \(x>0\), regular variation gives \[ \frac{\overline F(b_nx)}{\overline F(b_n)} =x^{-\alpha}\frac{L(b_nx)}{L(b_n)} \longrightarrow x^{-\alpha}. \] Multiplying by \(n\overline F(b_n)\to1\) yields \(n\overline F(b_nx)\to x^{-\alpha}\). The tail-probability theorem with \(u_n(x)=b_nx\) gives the stated limit for \(x>0\). For \(x<0\), \(P(M_n/b_n\le x)\to0\) because \(b_n\to\infty\) and the limiting law is supported on positive values; at \(x=0\), the same conclusion follows since \(P(M_n\le0)=F(0)^n\to0\) when the regularly varying unbounded upper tail has \(F(0)<1\). \(\square\)

The limit \(\exp(-x^{-\alpha})\) for \(x>0\) is a Fréchet distribution. In the GEV parameterization, its positive shape corresponds to \(\xi=1/\alpha\), after an appropriate location and scale representation.

Finite Upper Endpoints: Check the Endpoint Tail

Suppose the distribution has a finite upper endpoint \(x_F\), meaning \(P(X\le x_F)=1\) and \(P(X>x_F-\varepsilon)>0\) for every \(\varepsilon>0\). A finite endpoint makes the reverse-Weibull family a candidate, but the endpoint alone does not determine the limit. What matters is how probability accumulates just below it.

For fixed \(x<0\), the threshold \(x_F+a_nx\) lies below the endpoint, and \[ n\overline F(x_F+a_nx)\sim nc(-a_nx)^\alpha=(-x)^\alpha, \] because \(cn a_n^\alpha=1\). The tail-probability theorem gives the claimed limit for \(x<0\); at \(x=0\), the maximum is at most \(x_F\) with probability one, so the limit is \(1=\exp(0)\). For \(x>0\), the threshold eventually exceeds \(x_F\), making the probability exactly one. This proves the result. \(\square\)

The displayed limit is a reverse-Weibull form. Its corresponding GEV shape is negative, specifically \(\xi=-1/\alpha\), with an affine change of location and scale to the standard GEV representation.

Choosing Carefully: Clues, Not Shortcuts

The three tail patterns provide a disciplined starting point: an exponential-type tail suggests Gumbel; a regularly varying, unbounded tail suggests Fréchet; and a finite endpoint with power-like probability near it suggests reverse-Weibull. Each suggestion should be checked against the actual asymptotic condition and an appropriate normalization. In applications, the data can be too limited to reveal the far tail, so a plausible family is not automatically a confirmed one.

A particularly important caution is that a finite endpoint by itself is not enough to conclude reverse-Weibull attraction. The corollary requires power-like behavior near the endpoint. For example, a bounded distribution whose probability of lying within \(t\) of its endpoint decreases faster than every power of \(t\) need not satisfy that condition; some such distributions are instead in the Gumbel domain. The endpoint locates the boundary, while the tail’s rate of approach to that boundary determines which limit is appropriate.

• Do not classify from the density’s center. Extreme maxima depend on upper-tail behavior.
• Do not equate “light-tailed” with exactly exponential. Many different distributions can share the Gumbel limit.
• Do not infer a Fréchet limit from a few large observations. Regular variation is an asymptotic property, not a visual impression.
• Do not select reverse-Weibull from bounded support alone. Verify how the tail behaves near the endpoint.