How does Extreme Value Theory (POT method) improve VaR estimation in the tails?

Extreme Value Theory is the mathematical framework designed specifically for modeling rare, catastrophic events. The Peaks-Over-Threshold (POT) method is the practical version you'll see on the FRM exam.

The Problem with Normal VaR:

The normal distribution has thin tails — it assigns negligibly small probability to events beyond 4-5 standard deviations. But financial returns have fat tails: events like the 2008 crisis or the March 2020 crash occur far more often than the normal distribution predicts.

The POT Approach:

1. Choose a high threshold u (e.g., losses greater than the 95th percentile)
2. Collect exceedances: all observations where Loss > u
3. Fit a Generalized Pareto Distribution (GPD) to (Loss - u) for those exceedances

The GPD has two parameters:

• xi (shape): Controls tail fatness. xi > 0 means fat tails (Frechet-type). xi = 0 is exponential tails. xi < 0 means bounded tails.
• beta (scale): Controls the spread of exceedances

The GPD CDF:

G(x) = 1 - (1 + xi * x / beta)^{-1/xi}

Worked Example — Sterling Risk Advisory:

Sterling fits GPD to daily portfolio losses exceeding the 95th percentile threshold of $2.1M:

• xi = 0.25 (fat-tailed)
• beta = $0.8M
• Threshold u = $2.1M
• n = 1000 total observations, N_u = 50 exceedances

99% VaR using POT:

VaR_q = u + (beta/xi) x [(n/N_u x (1-q))^{-xi} - 1]

VaR_0.99 = 2.1 + (0.8/0.25) x [(1000/50 x 0.01)^{-0.25} - 1]

= 2.1 + 3.2 x [(0.2)^{-0.25} - 1]

= 2.1 + 3.2 x [1.4953 - 1]

= 2.1 + 3.2 x 0.4953

= 2.1 + 1.585

= $3.685M

Normal VaR comparison:

If the portfolio has mean daily P&L of $0 and std dev of $1.2M:

VaR_0.99 = 2.326 x $1.2M = $2.791M

The EVT estimate is 32% higher — the normal distribution dangerously underestimates tail risk.

When to Use EVT vs. Normal:

Key Exam Points:

• xi > 0 is the typical finding for financial data (fat tails)
• The threshold choice is a bias-variance tradeoff: too low includes non-tail data, too high leaves too few observations
• EVT makes no assumption about the full distribution — only the tail behavior

Explore more quantitative methods in our FRM Part I course.