Extreme Value Theory

You can now spot a fat tail and read its index $\alpha$. But how do you say something rigorous about a loss bigger than anything in your data — a 1-in-200-year event from 20 years of history? That sounds like fortune-telling, yet there’s a branch of statistics built precisely for it: Extreme Value Theory (EVT). Just as the central limit theorem says sums converge to the normal regardless of the underlying distribution, EVT says maxima and exceedances converge to a small, universal family regardless of the underlying distribution. That universality is what lets you extrapolate honestly past your data. This lesson covers EVT’s two pillars — block maxima → GEV and peaks-over-threshold → GPD — unified by one number, the shape parameter $\xi$.

Pillar 1: block maxima and the GEV

Analogy. Forget the daily noise; ask only “what was the worst day each year?” Line up those annual worst-days across decades and they form their own distribution — the distribution of maxima. EVT says that distribution has a universal shape, no matter what the daily returns looked like underneath.

Definition. Chop your data into equal blocks (say, one year each) and take the maximum of each block. As the block size grows, the (properly rescaled) distribution of these maxima converges to the Generalized Extreme Value (GEV) distribution, whose CDF is $G(x) = \exp\!\left\{-\left[1 + \xi\left(\tfrac{x-\mu}{\sigma}\right)\right]^{-1/\xi}\right\},$ with location $\mu$, scale $\sigma > 0$, and the all-important shape parameter $\xi$. This is the Fisher–Tippett–Gnedenko theorem — the EVT analogue of the CLT.

The shape $\xi$ splits the GEV into three classic families:

The connection to last lesson: for $\xi > 0$, the tail index is $\alpha = 1/\xi$. A shape of $\xi = 0.33$ means $\alpha = 3$. So $\xi$ and $\alpha$ are two dialects of the same fact about tail heaviness.

Pillar 2: peaks-over-threshold and the GPD

Block maxima are intuitive but wasteful — they throw away the second-worst day of a year even if it was a near-catastrophe, and keep the best-of-a-calm-year even if it was nothing. The modern workhorse uses the data better.

Analogy. Instead of “worst day each year,” ask “every day that breached this alarm level — how far past it did we go?” You keep all the big events, not just one per block, and you study the overshoots.

Definition. Fix a high threshold $u$. Look at the exceedances $X - u$ for all observations with $X > u$. The Pickands–Balkema–de Haan theorem says that as $u$ rises, the distribution of these overshoots converges to the Generalized Pareto Distribution (GPD): $F(y) = 1 - \left(1 + \frac{\xi y}{\beta}\right)^{-1/\xi}, \qquad y = x - u \ge 0,$ with scale $\beta > 0$ and the same shape parameter $\xi$ as the GEV. The two theorems are deeply linked: if block maxima are GEV with shape $\xi$, the threshold exceedances are GPD with the identical $\xi$. The shape parameter is the invariant that survives both framings.

Same three regimes:

• $\xi > 0$: heavy tail, the GPD becomes a power law ($\alpha = 1/\xi$).
• $\xi = 0$: the GPD reduces to a plain exponential tail.
• $\xi < 0$: a tail with a finite right endpoint at $u - \beta/\xi$.

This POT (peaks-over-threshold) approach is what banks and insurers actually use to model operational losses, catastrophe claims, and market tail risk, because it squeezes far more information out of limited data than block maxima do.

The shape parameter ξ: the one number that classifies any tail

Both pillars hand you the same $\xi$, and it is the single most important output of an EVT analysis. It answers the question that no amount of staring at a histogram can: is the tail bounded, exponential, or heavy?

• $\xi < 0$ — bounded tail. There is a hard maximum the variable cannot exceed. Useful for quantities with a natural ceiling (e.g. a percentage that can’t exceed 100, a bounded payoff).
• $\xi = 0$ — exponential tail. Light but unbounded; extremes are rare and shrink fast. The normal and lognormal land here.
• $\xi > 0$ — heavy tail. Unbounded with a power-law decay; the tail index is $\alpha = 1/\xi$. Larger $\xi$ = heavier tail = scarier. Financial returns typically estimate $\xi$ between about 0.2 and 0.5.

Worked example — what ξ buys you. Suppose POT on daily losses gives $\xi = 0.25$, $\beta = 0.8\%$, threshold $u = 2\%$, and exceedances occur on a fraction $\zeta = 0.05$ of days. You can now extrapolate a far-tail quantile the data never reached. The EVT formula for the $p$-quantile (VaR) is $\text{VaR}_p = u + \frac{\beta}{\xi}\left[\left(\frac{1-p}{\zeta}\right)^{-\xi} - 1\right].$ For a 99.9% level ($p = 0.999$, so $1-p = 0.001$): $\frac{1-p}{\zeta} = \frac{0.001}{0.05} = 0.02, \qquad 0.02^{-0.25} \approx 2.66.$ So $\text{VaR}_{0.999} = 0.02 + \frac{0.008}{0.25}(2.66 - 1) = 0.02 + 0.032 \times 1.66 \approx 0.02 + 0.053 = 0.073$, i.e. about a 7.3% one-day loss. You estimated a 1-in-1000-day loss from a few years of data — the central magic trick of EVT, made possible because the GPD form is guaranteed by theory, not assumed.

The threshold choice: the unavoidable judgment call

Both EVT methods share one nagging decision — where does the “tail” begin? For POT it’s the threshold $u$; for block maxima it’s the block size. The trade-off is identical and inescapable:

• Too high a threshold (too large a block): few exceedances survive, so the GPD/GEV fit is unbiased but noisy — high variance, wide confidence intervals.
• Too low a threshold (too small a block): you’ve reached into the body, where the limiting tail family doesn’t apply, so the fit is low-variance but biased — precisely wrong.

The standard remedy is the mean-excess plot (and the Hill plot from last lesson): for a true GPD tail, the average overshoot above $u$ is linear in $u$. So you raise $u$ and look for the point where the mean-excess plot straightens into a line — that’s where the GPD approximation kicks in, and you set your threshold there. It’s part science, part craft, and reasonable analysts disagree. The honest takeaway: EVT gives you a principled form for the tail, but it cannot conjure certainty from the handful of data points that, by definition, the tail contains.

Next up — expected shortfall beyond VaR — we turn EVT’s tail model into a better risk number. VaR points at where the tail starts but refuses to say how deep it goes; expected shortfall averages the whole tail, satisfies the coherence axioms VaR violates, and pairs beautifully with the GPD you just fitted.