Extreme Value Theory for Tail Risk

A risk manager wants the loss so bad it happens only once in a thousand trading days, about once every four years. Historical simulation with a two-year window has 500 observations and cannot see past its single worst day. Fitting a normal distribution to the whole sample tells you almost nothing about the tail, because the bulk of ordinary days determines the fit and the tail is thin by assumption, not by evidence. You need a method that uses the ordinary statistics of the extremes themselves, not the body of the distribution, to say something honest about losses you have rarely or never observed.

The engineer's flood problem

Civil engineers who design a levee do not model the entire distribution of the river's daily flow and read off a quantile. Daily flow is dominated by ordinary weather and tells you almost nothing about flood behavior. Instead they collect the worst flood of each year for as many years as they have records, and fit a distribution to just those maxima. There is a mathematical reason this works: no matter what the underlying daily-flow distribution looks like, the distribution of yearly maxima settles, as more years accumulate, into one of a small family of shapes. You do not need to know the parent distribution. You only need to know that you are looking at extremes.

Extreme value theory (EVT) applies the same idea to portfolio losses. It does not try to describe an ordinary day. It describes the tail on its own terms.

The two ways to isolate the tail

Block maxima. Split history into blocks (months or years), keep only the worst loss in each block, and fit those maxima to the Generalized Extreme Value (GEV) distribution. The Fisher–Tippett theorem guarantees that, whatever the daily loss distribution is, properly normalized block maxima converge to a GEV distribution as the block size grows. In words: the shape of "the worst thing that happened in a year" doesn't depend on the fine detail of an ordinary day, only on how fat the tail is.

Peaks over threshold (POT). Instead of throwing away every day but the yearly worst, pick a high threshold $u$ and keep every day that crosses it. The Pickands–Balkema–de Haan theorem says that, for a threshold high enough, the sizes of the exceedances above $u$ converge to a Generalized Pareto Distribution (GPD):

$P(L - u > x \mid L > u) \approx \left(1 + \xi \frac{x}{\beta}\right)^{-1/\xi}.$

In words: given that a loss has already crossed the threshold, the extra amount it exceeds the threshold by follows a two-parameter curve. $\xi$ (the shape parameter, or tail index) controls how fat the tail is: $\xi > 0$ means genuinely fat-tailed (finance usually lands here, $\xi$ around 0.2–0.4 for equities), $\xi = 0$ is exponential-thin, $\xi < 0$ is bounded. $\beta$ (the scale parameter) sets the units, roughly how large a typical exceedance is. POT is the version almost everyone uses in finance, because it keeps more of the data than block maxima throws away.

Once $\xi$ and $\beta$ are fit, VaR and expected shortfall beyond the threshold follow from closed formulas: with $n$ total observations, $n_u$ of them exceeding $u$, and target confidence $\alpha$,

$\text{VaR}_\alpha = u + \frac{\beta}{\xi}\left[\left(\frac{n}{n_u}(1-\alpha)\right)^{-\xi} - 1\right].$

In words: start at the threshold, then add a correction that depends on how far into the tail you are asking, how fat the tail is ($\xi$), and how spread out the exceedances were ($\beta$). This is the formula that lets you extrapolate to confidence levels beyond what your raw data ever showed.

Worked example: fitting the tail by hand

A desk has 2,000 days of P&L. It sets the threshold at the 95th percentile of losses, $u =$ $1.0m, leaving $n_u = 100$ exceedance days. The average amount by which those 100 days exceeded $1.0m is $0.55m, and a quick method-of-moments fit gives $\xi = 0.25$, $\beta = 0.60$.

We want the 99.5% VaR ($1 - \alpha = 0.005$):

$\text{VaR} = 1.0 + \frac{0.60}{0.25}\left[\left(\frac{2000}{100}(0.005)\right)^{-0.25} - 1\right] = 1.0 + 2.4\left[(0.1)^{-0.25} - 1\right].$

$(0.1)^{-0.25} \approx 1.778$, so $\text{VaR} \approx 1.0 + 2.4(0.778) = 1.0 + 1.87 =$ $2.87m.

Only 10 of the 2,000 days (0.5%) should exceed this. Notice that 99.5% is deep in the tail, well past what a 100-day exceedance sample could ever show you directly by counting, and yet the GPD formula produces a specific number by extrapolating the fitted curve, which is exactly what EVT is for.

Worked example: comparing to a normal assumption

The same book has a daily volatility of $0.40m. A normal model puts the 99.5% VaR at $2.576 \times 0.40 =$ $1.03m. EVT says $2.87m, almost triple. The gap is the fat tail the normal distribution assumes away: $\xi = 0.25 > 0$ means genuinely heavier-than-exponential decay, and the difference between the two numbers is the entire reason EVT exists as a discipline separate from ordinary VaR.

Drag the standard deviation on the plot above and watch how thin the tails stay under any normal parameterization; no choice of a single width parameter can reproduce the slower, fatter decay a fitted GPD tail shows in real market data.

What this means in practice

• Threshold choice is the whole ballgame. Too low, and the GPD approximation is poor because you're including days that aren't really "extreme." Too high, and too few exceedances remain to fit reliably. Practitioners typically use 5–10% of the sample and check stability by re-fitting at several thresholds.
• EVT complements, not replaces, VaR/historical simulation. It is specifically for the far tail, beyond where you have direct data, typically 99% and above.
• Multivariate extensions exist (extreme dependence, tail copulas) to ask whether extremes hit multiple assets together, directly relevant to Correlation Breakdown in Crises.
• Regulatory and internal capital models increasingly use EVT-derived tail estimates for stress capital, because they extrapolate more honestly than a normal or even a Student-t fit to the whole sample.

Practice

1. A threshold at the 90th percentile gives $n_u = 200$ out of $n = 2000$, with $\xi = 0.3$, $\beta = 0.45$, $u =$ $0.7m. Compute the 99% VaR.
2. Explain in one sentence why block maxima "wastes" data compared to peaks-over-threshold.
3. If $\xi$ estimated from 2019–2021 data is 0.05 but the true post-2022 regime has $\xi = 0.3$, is the resulting VaR too high or too low, and why?