Extreme Values Do Not Behave Like Ordinary Averages

Statistical intuition is built largely around averages. If independent observations have a finite variance, the sample mean becomes approximately Gaussian after centring and scaling, almost regardless of the detailed shape of the original distribution. This is one of the reasons mean and variance are so useful. Many distributions that look different at the observation level behave similarly after enough additive aggregation.

Maxima do not behave this way. The largest observation in a sample is determined by the tail, and two distributions that agree closely around their centres can produce completely different maxima. Matching means and variances does not fix the probability of an event five or ten standard deviations away. A model can fit the bulk of the data extremely well and still be unusable for design loads, flood levels, insurance losses, component stresses, heat extremes, market crashes or any other problem in which the scientific target lies in the tail rather than near the centre.

Extreme-value theory exists because the asymptotic problem is different. Instead of asking how sums behave, it asks how maxima and high threshold exceedances behave. The answer is strikingly restrictive. Under broad conditions, normalized maxima converge to one three-parameter family, the generalized extreme-value distribution, and high threshold excesses converge to the generalized Pareto family. The same shape parameter appears in both results and determines whether the underlying tail is heavy, exponentially light or bounded.

The practical value of the theory is not that it somehow observes rare events that have never happened. It provides a principled structure for extrapolating from the upper part of the data, while making the assumptions and uncertainty of that extrapolation explicit.

Averages and maxima require different normalizations

Let

be independent observations with common distribution function $F$. For the sample mean,

the central-limit theorem says, under familiar finite-variance conditions,

The mean aggregates small contributions from the entire distribution. The limiting Gaussian law is therefore controlled by first and second moments rather than by the detailed tail.

Now consider the maximum,

Its distribution is exact:

Write the upper-tail probability as

For a high level $x$ with small $\bar F(x)$,

A non-degenerate maximum therefore lives near levels $x_n$ for which

is of order one. This is already enough to show why the tail matters. The maximum is located where the survival probability is approximately

not where the bulk of the data lie. As $n$ grows, the relevant region moves further into the tail.

For a normal distribution, this movement is slow, roughly of logarithmic order after appropriate transformation. For a power-law tail, the maximum grows polynomially with sample size. For a distribution with a finite upper endpoint, the maximum approaches that endpoint instead of diverging. These are fundamentally different asymptotic regimes, and they cannot be distinguished by the mean and variance alone.

The extremal types theorem produces one family with three regimes

Suppose there exist constants

and

such that

converges in distribution to a non-degenerate limit. The Fisher-Tippett-Gnedenko extremal types theorem says that the limit must belong to the generalized extreme-value family. In its usual location-scale form,

with

and support restricted by

The limit as

is the Gumbel form,

The parameter $\mu$ locates the extreme distribution and $\sigma$ sets its scale. The most important parameter for extrapolation is $\xi$, the extreme-value shape parameter.

When

the distribution is of Fréchet type. The upper tail is heavy and unbounded. Pareto-type distributions fall in this class. If

then the extreme-value index is

When

the distribution is of Gumbel type. The upper tail is unbounded but lighter than a power law. Gaussian and exponential distributions both lie in this domain of attraction even though their ordinary distributions are very different.

When

the distribution is of Weibull type for maxima. The underlying distribution has a finite upper endpoint. Uniform and many bounded physical quantities are examples.

This classification is one of the deepest simplifications in extreme-value theory. The detailed form of $F$ can vary enormously, but the asymptotic behaviour of its maxima is governed by the tail class and, in particular, by $\xi$.

The same simplification also explains why a central fit can be misleading. A Gaussian model and a heavy-tailed model can agree remarkably well over the region where most observations occur while belonging to different extreme-value domains. Once extrapolation moves beyond the observed bulk, the asymptotic difference dominates.

The first two moments do not determine the tail

Consider two distributions with mean zero and variance one. The first is standard normal,

The second is a Student distribution with four degrees of freedom, rescaled to unit variance,

Because a Student $t$ distribution with four degrees of freedom has variance $2$, both $X_N$ and $X_T$ have

and

Near the centre, both are symmetric distributions of comparable scale. Their extreme tails are not remotely similar. The normal distribution is in the Gumbel domain with

while the $t_4$ tail is regularly varying with exponent $4$, giving

The difference becomes visible in high quantiles.

Upper probability	Standard normal quantile	Unit-variance $t_4$ quantile
0.95	1.645	1.507
0.99	2.326	2.649
0.999	3.090	5.072
0.9999	3.719	9.216
0.99999	4.265	16.498

At the 95th percentile the two distributions are close. At the 99.999th percentile the heavy-tailed quantile is almost four times as large.

The contrast is even stronger if we fix a level rather than a probability. At

the standard normal exceedance probability is approximately

For the unit-variance $t_4$ distribution,

The heavy-tailed model assigns the same event a probability about

times larger.

Both models have exactly the same mean and variance. No moment calculation based only on those two quantities can determine which tail is correct.

The difference also predicts very different sample maxima. A rough location for the maximum of $n$ independent observations is the quantile

For

the relevant upper probability is approximately

The normal model places the maximum near $4.27$, while the standardized $t_4$ model places it near $16.50$. A central model that cannot distinguish these two tails cannot answer a question about the maximum reliably.

This is the core reason EVT is not simply ordinary distribution fitting with more emphasis on large observations. The scientific object has changed. The relevant asymptotic information is in the tail class.

Block maxima and threshold exceedances are two views of the same tail

The generalized extreme-value limit motivates the block-maxima method. Divide a sequence into blocks, such as years, seasons or batches, retain the maximum from each block, and fit a GEV model to those maxima. The method aligns directly with the limiting theorem and is often natural when the scientific quantity itself is an annual or batch maximum.

Its cost is information loss. If a year contains several exceptional events, only the largest survives. If the data record is short, the number of block maxima can be very small.

Peaks over threshold use more of the tail. Choose a high threshold $u$ and study the conditional excess

The Pickands-Balkema-de Haan theorem states that if the underlying distribution belongs to a maximum domain of attraction, then for sufficiently high $u$ the excess distribution is approximately generalized Pareto,

on the support where

When

the limiting form is exponential,

The important fact is that the GEV and GPD share the same shape parameter $\xi$. Block maxima and threshold exceedances are therefore not unrelated modelling tricks. They are two manifestations of the same underlying extreme-value limit theory.

If the target is a return level or a high quantile and enough raw data are available, peaks over threshold often use tail information more efficiently because all sufficiently large exceedances contribute. Block maxima can be preferable when natural blocks define the scientific quantity, when within-block observations are difficult to compare, or when only maxima have been recorded historically.

The choice should follow the data-generating process rather than fashion. The operational-risk article Extreme Value Theory: Estimating the Tail You Have Not Seen develops the peaks-over-threshold approach through a full simulation. The present article focuses instead on the mathematical structure that justifies both approaches.

Threshold selection is the central bias-variance trade-off

The generalized Pareto approximation is asymptotic. It becomes more defensible as the threshold moves further into the tail. Raising the threshold therefore reduces approximation bias. It also leaves fewer exceedances and increases statistical uncertainty.

This produces the defining practical trade-off in peaks-over-threshold analysis. A threshold that is too low includes observations that do not yet behave like the limiting tail model. A threshold that is too high produces a model based on very little data.

The GPD itself gives useful diagnostics. Suppose exceedances above $u$ follow a generalized Pareto model with shape $\xi$ and scale $\beta_u$. If we raise the threshold to

the excess distribution remains GPD with the same shape parameter and updated scale

The stability of $\xi$ across a range of high thresholds is therefore evidence that the tail approximation has entered a coherent regime.

For

the mean excess exists. Under the GPD,

Using the threshold-stability relation,

which is linear in the threshold. This motivates the mean residual life plot. A roughly linear region supports, but does not prove, the use of a GPD above those thresholds.

Neither diagnostic supplies a unique mechanical threshold. In finite samples the estimates fluctuate, different diagnostics can disagree, and slowly converging distributions can retain bias far into the tail. A defensible analysis therefore reports sensitivity to a range of thresholds rather than presenting one threshold as though it were observed directly from nature.

The same logic appears in tail-index estimators such as the Hill estimator for positive $\xi$. Using more upper order statistics reduces variance and increases the risk of including non-tail observations. Using fewer order statistics reduces bias asymptotically and increases variance. The tuning parameter changes form, but the underlying problem is the same.

Return levels are probabilities, not schedules

Extreme-value results are often communicated through return levels. Suppose annual maxima follow a GEV distribution $G$. A $T$-year return level $z_T$ is conventionally defined so that the probability of exceeding it in any one year is

Equivalently,

For

the GEV quantile formula gives

For

the limit becomes

The shape parameter has an increasingly large influence as $T$ grows. To see this, set

and

At a 100-block return period, the return levels are approximately

for

for

and

for

At a 1000-block return period, the same three models give approximately

Small uncertainty in the shape parameter therefore becomes large uncertainty in far-tail extrapolation. This is exactly where narrow standard errors for central parameters can become misleading.

The phrase "100-year event" is also easy to misinterpret. It does not mean the event occurs once every hundred years on a deterministic schedule. If annual exceedances were independent with probability

per year, the probability of at least one exceedance during the next hundred years would be

More generally,

as $T$ grows. A return period is the reciprocal of an exceedance probability under a specified model. It is not a calendar promise.

Dependence changes the effective number of extremes

The simplest extreme-value theory assumes independence. Environmental, financial, engineering and biological extremes often occur in clusters. Heat waves persist for several days. Market volatility creates runs of large losses. Storms can generate multiple high measurements within one event. Treating every exceedance as independent can therefore exaggerate the effective amount of tail information.

Under suitable dependence conditions, a stationary sequence can still have an extreme-value limit, but a new quantity appears: the extremal index

If levels $u_n$ are chosen so that

then a common limiting form is

When

the maximum behaves asymptotically like that of an independent sequence with the same marginal tail. Values

indicate clustering of extremes.

Under specific conditions, the reciprocal

can be interpreted in relation to an asymptotic mean cluster size, but that interpretation should not be applied mechanically in every dependence structure. The important point is that dependence changes how many effectively independent extreme episodes the record contains.

In practice this motivates declustering, cluster maxima, run-length methods or explicit time-series extreme-value models. The correct choice depends on whether the scientific target is the probability of individual exceedances, event clusters, episode maxima or cumulative impact.

Dependence also changes uncertainty. A record containing 500 threshold exceedances may contain much less than 500 independent pieces of tail information if those exceedances arrived in a small number of storms or stress episodes.

EVT does not eliminate extrapolation risk

Extreme-value theory reduces the tail to a structured asymptotic problem, but it does not make far-tail inference easy. The rarest levels remain extrapolations beyond the observed record. The shape parameter is estimated from relatively little data, and far quantiles are highly sensitive to its uncertainty.

A complete EVT analysis therefore needs more than a maximum-likelihood estimate and one return level. Threshold or block sensitivity should be reported. Profile likelihood or bootstrap intervals should accompany remote return levels. Dependence should be assessed. Changes over time, covariate effects and non-stationarity should be considered when the physical process is evolving.

Stationarity is especially important. A 100-year return level estimated from a stationary model assumes that the distribution governing future extremes remains the same. If temperature, exposure, technology, land use, population, regulation or system capacity changes over time, the return level itself may be time-dependent. A precise stationary estimate can then answer a historical question more accurately than a future one.

There is also model risk in the asymptotic approximation. A distribution can belong to the Gumbel domain and still converge so slowly that finite-sample maxima behave poorly relative to the limiting GEV. Threshold estimates can drift over the entire observed range. Different plausible tail models can produce similar likelihoods near the data and sharply different extrapolations outside them.

This is not a reason to avoid EVT. It is a reason to report tail inference with the same seriousness given to any other inverse extrapolation problem. The further the return level lies beyond the record, the more the conclusion depends on asymptotic structure rather than direct observation.

The main practical error is to hide that dependence by fitting a familiar central distribution to all observations and reading off an extreme quantile. Such a model can appear more stable because most of the data constrain its centre tightly. That stability is irrelevant if the assumed tail class is wrong.

Extreme-value theory makes the extrapolation assumption explicit. It asks which tail domain the process belongs to, estimates the associated shape, and uses only the part of the data informative about that structure. The result is usually less comfortable than a central fit because uncertainty grows rapidly as we move deeper into the tail. That discomfort is often the correct statistical answer.

Averages and maxima do not obey the same asymptotic law. The central-limit theorem tells us why many sums become Gaussian. Extreme-value theory tells us why maxima converge to a different three-regime family and why high threshold excesses inherit the same tail shape through the generalized Pareto distribution. Mean and variance can describe the centre while saying almost nothing about the event that matters most.

When the scientific question is about the largest observation rather than the average one, the tail is not an afterthought. It is the model.

References

Balkema, A. A., & de Haan, L. (1974). Residual life time at great age. The Annals of Probability, 2(5), 792–804.

Beirlant, J., Goegebeur, Y., Segers, J., & Teugels, J. (2004). Statistics of Extremes: Theory and Applications. Wiley.

Coles, S. (2001). An Introduction to Statistical Modeling of Extreme Values. Springer.

Embrechts, P., Klüppelberg, C., & Mikosch, T. (1997). Modelling Extremal Events for Insurance and Finance. Springer.

Fisher, R. A., & Tippett, L. H. C. (1928). Limiting forms of the frequency distribution of the largest or smallest member of a sample. Proceedings of the Cambridge Philosophical Society, 24(2), 180–190.

Leadbetter, M. R., Lindgren, G., & Rootzén, H. (1983). Extremes and Related Properties of Random Sequences and Processes. Springer.

Pickands, J. (1975). Statistical inference using extreme order statistics. The Annals of Statistics, 3(1), 119–131.

Resnick, S. I. (1987). Extreme Values, Regular Variation, and Point Processes. Springer.