Introduction to extreme value theory

Q: What are the two primary approaches used in Extreme Value Theory (EVT) to model extreme data?

The two primary approaches are the Block Maxima approach (analysing the maximum value within fixed intervals) and the Peaks Over Threshold (POT) approach (analysing all values that exceed a set threshold).

Q: How do the three types of the Generalized Extreme Value (GEV) distribution correspond to the shape parameter $ \gamma $?

$ \gamma = 0 $ corresponds to the light-tailed Gumbel (Type I); $ \gamma > 0 $ corresponds to the heavy-tailed Fréchet (Type II); and \( \gamma Weibull (Type III) with a finite upper bound.

Q: What does the Pickands-Balkema-de Haan Theorem state about excesses over a high threshold?

It states that for a sufficiently high threshold $ u $, the excess distribution $ Y = X - u \mid X > u $ converges to the Generalized Pareto Distribution (GPD) .

Q: How does the shape of the mean excess function $ e(u) $ distinguish between heavy-tailed, exponential, and short-tailed distributions?

$ e(u) $ is increasing in $ u $ for heavy-tailed distributions, constant for exponential distributions, and decreasing for short-tailed distributions.

CS2 – Risk Modelling and Survival Analysis · IFoA (Institute and Faculty of Actuaries) · 7 min read · Updated: 8 Nov 2026

Welcome to the World of the "Unlikely": Introduction to Extreme Value Theory

Hello there! Welcome to one of the most fascinating chapters in your CS2 journey. So far in your studies, you’ve likely spent a lot of time looking at "average" outcomes—the middle of the Bell Curve. But as an actuary, you aren't just worried about the average car crash or the average life expectancy. You are worried about the catastrophes.

Extreme Value Theory (EVT) is the branch of statistics that deals with the "tails" of a distribution. It’s about the 1-in-100-year flood, the massive stock market crash, or the huge insurance claim that could bankrupt a company. This chapter teaches us how to model those rare but high-impact events. Don’t worry if the math looks intimidating at first—we’re going to break it down piece by piece!

1. Why do we need EVT?

Imagine you are building a sea wall. If you only look at the average tide, your wall will be very short, and the first big storm will wash the town away. You need to model the maximum possible tide.

Standard distributions (like the Normal distribution) often "underestimate" how often extreme events happen. They have "thin tails." EVT gives us a mathematically sound way to estimate the probability of events that we may have never even seen in our historical data yet!

Key Concept: The Two Main Approaches

In CS2, we focus on two ways to look at extremes:
1. The Block Maxima Approach: Looking at the biggest value in fixed time intervals (e.g., the highest claim each year).
2. The Peaks Over Threshold (POT) Approach: Looking at every value that "breaks the limit" (e.g., every claim over \$1 million).

2. The Block Maxima and the GEV Distribution

Let’s start with Block Maxima. Suppose we have a sequence of independent and identically distributed (i.i.d.) random variables $ X_1, X_2, \dots, X_n $. We define the maximum as:
$ M_n = \max(X_1, X_2, \dots, X_n) $

The Fisher-Tippett-Gnedenko Theorem tells us that as $ n $ gets very large, the distribution of this maximum (after suitable linear normalisation) converges to the Generalized Extreme Value (GEV) Distribution.

The GEV Formula

The cumulative distribution function (CDF) of the GEV is:
$ H_{\gamma}(x) = \exp \left( -(1 + \gamma x)^{-1/\gamma} \right) $

for $ 1 + \gamma x > 0 $, where $ \gamma $ (gamma) is the shape parameter. This parameter determines how "heavy" the tail is.

The Three Flavors of GEV

Depending on the value of $ \gamma $, the distribution falls into one of three types:
• Type I: Gumbel ($ \gamma = 0 $): Used for "light-tailed" distributions like the Exponential or Normal. Taking the limit as $ \gamma \to 0 $, $ H_0(x) = \exp(-\exp(-x)) $. The tail decays exponentially.
• Type II: Fréchet ($ \gamma > 0 $): This is the "heavy-tailed" version. This is very important for actuaries because it models distributions like the Pareto, where huge losses are more likely.
• Type III: Weibull ($ \gamma < 0 $): This is "short-tailed" with a finite upper limit. Think of it as a maximum possible ceiling—it does not extend to infinity.

Quick Review:
- $ \gamma > 0 $: Heavy tail (Fréchet) - Watch out, big claims ahead!
- $ \gamma = 0 $: Light tail (Gumbel) - Thin tails.
- $ \gamma < 0 $: Finite tail (Weibull) - There is an upper bound.

3. Peaks Over Threshold (POT) and the GPD

While Block Maxima only looks at one value per block, the POT approach is more efficient because it uses all data points that are "extreme."

If we pick a high threshold $ u $, we are interested in the excess values: $ Y = X - u $, given that $ X > u $.

The Pickands-Balkema-de Haan Theorem states that for a sufficiently high threshold $ u $, the distribution of these excesses converges to the Generalized Pareto Distribution (GPD).

The GPD Formula

The CDF of the GPD for excesses $ y \ge 0 $ is given by:
$ G_{\gamma, \beta}(y) = 1 - \left(1 + \frac{\gamma y}{\beta}\right)^{-1/\gamma} $ (for $ \gamma \neq 0 $)
$ G_{\gamma, \beta}(y) = 1 - \exp\left(-\frac{y}{\beta}\right) $ (for $ \gamma = 0 $)

Here, $ \beta > 0 $ is the scale parameter and $ \gamma $ is the same shape parameter found in the corresponding GEV!

Did you know?
There is a direct link between GEV and GPD. If the block maxima of a dataset follow a GEV distribution with shape parameter $ \gamma $, then the excesses over a high threshold for that same underlying distribution follow a GPD with the exact same $ \gamma $.

4. Measuring and Comparing Tail Weights

In CS2, you need to be able to calculate and interpret mathematical measures of tail weight to compare which distribution has a heavier tail. Here are the four primary tools:

1. Limiting Ratio of Survival Functions

To compare the tail weight of two distributions $ X_1 $ and $ X_2 $ with survival functions $ \bar{F}_1(x) = P(X_1 > x) $ and $ \bar{F}_2(x) = P(X_2 > x) $, examine the limit of their ratio as $ x \to \infty $:
$ \lim_{x \to \infty} \frac{\bar{F}_1(x)}{\bar{F}_2(x)} $
• If the limit is $ \infty $: $ X_1 $ has a heavier tail than $ X_2 $.
• If the limit is 0: $ X_1 $ has a lighter tail than $ X_2 $.
• If the limit is a finite positive constant $ c > 0 $: $ X_1 $ and $ X_2 $ have tails of equivalent weight.

2. The Hazard Rate (Force of Mortality)

The hazard rate is defined as $ h(x) = \frac{f(x)}{\bar{F}(x)} $. Examining how $ h(x) $ behaves as $ x \to \infty $ tells us about the tail:
• Heavy-tailed: $ \lim_{x \to \infty} h(x) = 0 $ (e.g., Pareto, Log-normal). The conditional probability of failure decreases as value grows.
• Medium / Exponential tail: $ \lim_{x \to \infty} h(x) = \lambda > 0 $ (e.g., Exponential distribution with parameter $ \lambda $).
• Light-tailed: $ \lim_{x \to \infty} h(x) = \infty $ (e.g., Normal, Gamma with shape parameter $ > 1 $).

3. The Mean Excess Function

The mean excess function is the expected amount by which a claim exceeds $ u $, given that it exceeds $ u $:
$ e(u) = E[X - u \mid X > u] = \frac{\int_u^{\infty} \bar{F}(x) \, dx}{\bar{F}(u)} $
• Heavy tail: $ e(u) $ is increasing in $ u $ (e.g., Pareto, where $ e(u) = \frac{u + \lambda}{\alpha - 1} $ for $ \alpha > 1 $).
• Memoryless / Light tail: $ e(u) $ is constant (e.g., Exponential, $ e(u) = 1/\lambda $).
• Short tail / Finite limit: $ e(u) $ is decreasing in $ u $.

4. Existence of Moments

Heavy-tailed distributions typically have moments that do not exist beyond a certain order. For example, for a Pareto distribution with shape parameter $ \alpha $, $ E[X^k] $ exists only for $ k < \alpha $. Distributions where all positive moments $ E[X^k] $ exist (e.g., Normal, Gamma, Exponential) have lighter tails than distributions where higher moments are infinite.

5. Practical Challenges: Choosing the Threshold

When applying the POT approach, how high should the threshold $ u $ be?
• If $ u $ is too high: You won’t have enough data points, leading to high parameter variance.
• If $ u $ is too low: The asymptotic theorem does not hold, leading to biased estimates.

The Mean Excess Plot: Actuaries plot sample mean excess against various values of $ u $. If the empirical plot becomes approximately linear and upward-sloping above a certain point, that indicates a Pareto/GPD tail and helps select the appropriate threshold $ u $.

6. Common Mistakes to Avoid

1. Mixing up GEV and GPD: Remember: GEV is for Block Maxima (the largest observation per block). GPD is for Exceedances over a threshold.
2. Forgetting the condition: The GPD formula applies to the excess $ Y = X - u $, not the original loss amount $ X $.
3. Misinterpreting $ \gamma $: In the standard formulation, $ \gamma > 0 $ corresponds to heavy tails (Fréchet / Pareto-type), $ \gamma = 0 $ to light tails (Gumbel / Exponential-type), and $ \gamma < 0 $ to bounded upper tails (Weibull-type).

7. Summary and Key Takeaways

• The Goal: EVT provides asymptotic models for the extreme tails of loss distributions.
• Block Maxima: Modelled using the Generalized Extreme Value (GEV) distribution (Gumbel, Fréchet, Weibull).
• Peaks Over Threshold: Excesses over high thresholds are modelled using the Generalized Pareto Distribution (GPD).
• Shape Parameter ($ \gamma $): Identical in both GEV and GPD; $ \gamma > 0 $ signals heavy tails.
• Comparing Tail Weights: Evaluated using limits of survival function ratios $ \lim \bar{F}_1(x)/\bar{F}_2(x) $, hazard rate behaviour $ \lim h(x) $, the slope of the mean excess function $ e(u) $, and the existence of higher moments.

References

CS2 Risk Modelling and Survival Analysis Syllabus for the 2027 Examinations — Institute and Faculty of Actuaries

Quick check

Can you answer these now?

Open each question to check the key ideas from this chapter.

What are the two primary approaches used in Extreme Value Theory (EVT) to model extreme data?

The two primary approaches are the Block Maxima approach (analysing the maximum value within fixed intervals) and the Peaks Over Threshold (POT) approach (analysing all values that exceed a set threshold).

What does the Fisher-Tippett-Gnedenko Theorem state regarding normalised block maxima?

It states that as the block size $ n \to \infty $, the distribution of the normalised block maximum $ M_n = \max(X_1, \dots, X_n) $ converges to the Generalized Extreme Value (GEV) distribution.

How do the three types of the Generalized Extreme Value (GEV) distribution correspond to the shape parameter $ \gamma $?

$ \gamma = 0 $ corresponds to the light-tailed Gumbel (Type I); $ \gamma > 0 $ corresponds to the heavy-tailed Fréchet (Type II); and $ \gamma < 0 $ corresponds to the short-tailed Weibull (Type III) with a finite upper bound.

What does the Pickands-Balkema-de Haan Theorem state about excesses over a high threshold?

It states that for a sufficiently high threshold $ u $, the excess distribution $ Y = X - u \mid X > u $ converges to the Generalized Pareto Distribution (GPD).

What is the key link between the shape parameter $ \gamma $ in the GEV distribution and the GPD?

The shape parameter $ \gamma $ is identical in both distributions; if block maxima converge to a GEV with parameter $ \gamma $, excesses over a high threshold converge to a GPD with the exact same $ \gamma $.

How can the limiting ratio of survival functions $ \lim_{x \to \infty} \frac{\bar{F}_1(x)}{\bar{F}_2(x)} $ be used to compare tail heaviness?

If the limit is $ \infty $, $ X_1 $ has a heavier tail than $ X_2 $; if $ 0 $, $ X_1 $ has a lighter tail; and if a positive constant $ c > 0 $, both distributions have equivalent tail weight.

How does the asymptotic behavior of the hazard rate $ \lim_{x \to \infty} h(x) $ characterize the tail of a distribution?

A limit of $ 0 $ indicates a heavy tail, a constant $ \lambda > 0 $ indicates a medium/exponential tail, and $ \infty $ indicates a light tail.

How does the shape of the mean excess function $ e(u) $ distinguish between heavy-tailed, exponential, and short-tailed distributions?

$ e(u) $ is increasing in $ u $ for heavy-tailed distributions, constant for exponential distributions, and decreasing for short-tailed distributions.

What is the primary trade-off and practical diagnostic tool when selecting the threshold $ u $ in the POT approach?

The trade-off is between bias (threshold too low) and variance (threshold too high); it is guided by a mean excess plot, which should become linear and upward-sloping above a valid threshold $ u $.

Ready to test yourself?

Turn these notes into exam-style practice. Get unlimited AI questions on this topic with instant marking and explanations.

Practise This Topic

More CS2 – Risk Modelling and Survival Analysis chapters

* The content provided by thinka is generated by AI and may not always be accurate or up-to-date. Please use it as a supplementary resource and verify with official materials.

Put These Notes into Practice

Reading the notes is a great start. Now practise with unlimited AI-generated questions and get instant feedback. 100,000+ students are already improving their grades.

Start Practising Now View Pricing

Done reading? Test yourself with AI practice questions

Practise This Topic Now