Extreme Value Theory

Most statistics is about the typical — the average, the spread, the central bulk of a distribution. But the events that shape our lives and our risk budgets are the extremes: the hundred-year flood, the record-breaking heatwave, the once-in-a-generation crash. And here's the cruel twist — these are exactly the events ordinary statistics, built around the centre, describes worst. Extreme value theory (EVT) is the specialised branch built for the tails: the maths of estimating how likely a rare extreme is, even one more severe than anything yet recorded.

It's a topic close to my CSIRO climate-risk work, where the whole question is the probability of extremes. This page is why the tails need their own theory, the two frameworks for modelling them (GEV and the Generalised Pareto), how the famous "1-in-N-year" event is computed, and the serious caveat that climate change has thrown at the whole enterprise. It builds on the probability page.

The tail is the point

The defining feature of EVT is that it deliberately throws away the bulk of the data and studies only the extremes — because the centre of a distribution tells you almost nothing about its tail. Two datasets can have identical means and variances but wildly different chances of a catastrophic outlier. For flood defences, insurance, infrastructure, and hazard planning, it's the tail probability — not the average — that determines whether you're prepared or exposed.

And the goal is genuinely audacious: estimate the probability of an event more extreme than any observed so far. You have 50 years of records and need the 200-year flood. That's extrapolation beyond the data — which is impossible in general, except that EVT provides a remarkable theoretical reason it can sometimes be done.

Why normal statistics fail in the tails

The instinct is to fit a familiar distribution (a normal) to all the data and read off the tail. This fails badly: a fitted normal is shaped to match the centre where most data sits, and it systematically underestimates the chance of extreme events, because real-world tails are often far heavier than the normal's thin, fast-decaying one. Using the bulk to predict the tail is how "impossible" 10-sigma events keep happening.

The deeper issue is conceptual: a "1-in-100-year" event is not 100× rarer than a typical year in any simple linear sense — the relationship between magnitude and rarity in the tail follows its own law. EVT's contribution is to identify what that law is, so you model the extremes with the right family of distributions rather than forcing the wrong one.

Block maxima & the GEV distribution

The first framework is block maxima: divide the record into blocks (e.g. years) and keep only the maximum of each (the hottest day of each year). Then comes the beautiful result that makes EVT work — the extremal types theorem (Fisher–Tippett): no matter what the original distribution is, the distribution of those block maxima converges to a single family, the Generalised Extreme Value (GEV) distribution.

Fit the GEV to your block maxima, and you have a model of the extremes you can push beyond the observed range — the 200-year event from 50 years of annual maxima.

Peaks over threshold & the Generalised Pareto

Block maxima is wasteful — it keeps one value per year and discards the second-worst day even if it was also extreme. The peaks-over-threshold (POT) approach uses the data better: pick a high threshold and model every exceedance over it. The companion theorem (Pickands–Balkema–de Haan) says these threshold exceedances converge to the Generalised Pareto Distribution (GPD).

Two ways to capture the tail. Block maxima keeps the single largest value per block (one per year) and fits a GEV. Peaks-over-threshold keeps every value above a high threshold and fits a Generalised Pareto — using far more of the extreme data from the same record.

POT is usually preferred in practice (in hydrology, finance, climate) precisely because it extracts more information from the tail — more exceedances mean more data to pin down the distribution, from the same record. The trade-off is choosing the threshold: too low and you contaminate the tail with non-extreme data; too high and you're back to too few points.

The shape parameter: how heavy is the tail?

Both the GEV and GPD have a critical shape parameter $\xi$ (xi) that controls the tail's character — arguably the single most important number in the analysis. Its sign sorts the world into three tail types:

$\xi = 0$ — Gumbel: a light, exponential tail (extremes get rarer fast; e.g. roughly normal-ish data).
$\xi > 0$ — Fréchet: a heavy tail with no upper bound (extreme events much more likely than intuition suggests — financial losses, some rainfall). The dangerous case.
$\xi < 0$ — Weibull: a tail with a finite upper limit (there's a hard physical maximum).

Estimating $\xi$ tells you whether you live in a world where the worst is bounded or where there's always a bigger catastrophe lurking — a distinction that completely changes how much margin to build in.

Return levels: the '1-in-N-year' event

The headline output of EVT is the return level — the magnitude expected to be exceeded once on average every $N$ years (the "100-year flood"). The companion idea is the return period: an event with a return period of $N$ years has, each year, roughly a $1/N$ probability of occurring:

T = \frac{1}{p} \quad\Longleftrightarrow\quad p = \frac{1}{T}

The honest limits

EVT is powerful and unusually honest about its own fragility:

Common pitfall

You're extrapolating beyond the data, so the uncertainty on a 500-year level from 50 years of records is enormous — always report confidence intervals, and treat a point estimate as a centre of a wide range, not a fact. Results are sensitive to the threshold / block choice. And — most seriously for climate — classical EVT assumes stationarity: that the distribution of extremes isn't changing. Under climate change it is, which means the familiar "return period" and "return level" can be outright misleading — yesterday's 1-in-100-year event may be tomorrow's 1-in-20. Modern practice uses non-stationary EVT (letting the parameters trend with time or covariates), but the uncertainty grows further. The discipline is to quantify the extremes and be loud about how uncertain the quantification is.

Where it shows up in my work

The probability of the rare

EVT is core to the climate-risk work I did at CSIRO — the entire question there is the probability and magnitude of extremes (extreme heat, rainfall, hazard events), which is exactly what ordinary statistics handles worst and EVT is built for. The key discipline it instils is to model the tail directly (GEV or POT) rather than fit a distribution to the bulk and underestimate the catastrophe.

Two cautions travel with it everywhere: the "1-in-N-year" communication trap (it's an annual probability, not a schedule — a real risk-communication issue tying to briefing clearly), and non-stationarity under climate change, which means a return level computed on historical data can be dangerously out of date (the drift problem in a different guise). It pairs with time series (the underlying record) and probability (the tail laws) — and is one of the most consequential tools when the rare event is the whole point.

Refresh in 60 seconds

EVT models the tail, not the bulk — because the centre of a distribution barely constrains its extremes, and normal-fit tails underestimate catastrophes.
Block maxima → GEV: maxima converge to the Generalised Extreme Value distribution regardless of the parent (the CLT analogue for maxima — Fisher-Tippett).
Peaks-over-threshold → GPD: model every exceedance over a high threshold — uses more tail data; preferred in practice (mind the threshold choice).
The shape parameter ξ sets the tail type: Gumbel (light), Fréchet (heavy, unbounded — dangerous), Weibull (bounded).
Return level / period: the "1-in-N-year" event = a 1/N annual probability, NOT a schedule (two can hit back-to-back).
Limits: huge extrapolation uncertainty (report CIs), threshold sensitivity, and non-stationarity — climate change breaks the stationary assumption.

The block-maxima/GEV and POT/GPD frameworks, the shape-parameter tail types, return-period interpretation, and the non-stationarity caution reflect current extreme-value references alongside climate-risk work.