FRM Exam Part II · Parametric Approaches (II): Extreme Value

Generalized Extreme Value (GEV) Distribution and Block Maxima

Q: Which GEV type fits financial losses?

Fréchet is the usual fit because returns have fat tails, like the Student t. Gumbel is a thinner-tailed approximation. Weibull is rare because losses rarely have a hard upper limit.

Q: How do you estimate GEV parameters for VaR?

Split the losses into blocks, take each block maximum, and fit μ, σ and ξ, usually by maximum likelihood. Then convert the VaR confidence level to p = α^n and invert the GEV to get the quantile.

Q: How do you choose the block size?

There is a trade-off. Bigger blocks fit the GEV theory better but give fewer maxima and noisier estimates. Smaller blocks give more data but more bias.

Updated 11 October 2026 · Fact-checked

The GEV distribution describes the limiting behaviour of block maxima, such as the largest daily loss in each month. It has a location, a scale and a tail index ξ. If ξ > 0 it is Fréchet (fat tail), if ξ = 0 it is Gumbel (thin tail), and if ξ < 0 it is Weibull (bounded tail). You fit it to the maxima and invert it for VaR.

Understand Generalized Extreme Value (GEV) Distribution

Most risk measures focus on the centre of a distribution. Extreme value theory (EVT) focuses on the tail. The block maxima method splits your data into equal blocks, such as months or years, and takes the largest loss in each block. These maxima are then the sample you model.

The key result is like a central limit theorem for maxima. If the maxima of large blocks are suitably normalised, they converge to the generalized extreme value (GEV) distribution, whatever the parent distribution was. You do not need to know the exact distribution of daily losses.

The GEV has three parameters. The location μ is the centre of the maxima. It is not the mean. The scale σ > 0 is their dispersion. It is not the standard deviation. The tail index ξ controls tail shape. A larger ξ means a heavier tail.

The sign of ξ gives three cases. Fréchet: ξ > 0, heavy tail, with power-law decay. Typical for financial returns such as the Student t. Gumbel: ξ = 0, exponential decay. The normal and lognormal fall here. Weibull: ξ < 0, finite upper bound. Rare for financial losses. The tail index is sometimes quoted as α = 1/ξ for the Fréchet case, so a smaller α means a fatter tail.

There is a trade-off in choosing block size. Large blocks fit the theory better (less bias) but leave few maxima (high variance). Small blocks give more data but the GEV fits less well. Block maxima also waste data, because only one point per block is used. That is why peaks over threshold is often preferred.

Key formulas to remember

GEV cumulative distribution (ξ ≠ 0): F(x) = exp{ −[1 + ξ(x − μ) ÷ σ]^(−1/ξ) }, where 1 + ξ(x − μ) ÷ σ > 0; x is the block maximum of losses. μ is location, σ > 0 is scale, ξ is the tail index.
Gumbel case (ξ = 0): F(x) = exp{ −exp[−(x − μ) ÷ σ] }; Limit of the GEV as ξ → 0. Thin, exponentially decaying tail.
Tail index cases: ξ > 0: Fréchet; ξ = 0: Gumbel; ξ < 0: Weibull; Fréchet is the usual case for financial losses.
Quantile of block maxima (ξ ≠ 0): x_p = μ − (σ ÷ ξ) × [1 − (−ln p)^(−ξ)]; p is the cumulative probability of the maximum. Obtained by inverting F(x) = p.
Gumbel quantile: x_p = μ − σ × ln(−ln p); Use when ξ = 0.
Link to VaR for the original data: If a block has n observations and the VaR confidence is α, then p = α^n; The maximum of n observations is below x only if all n are. So P(max ≤ x) = F_loss(x)^n. Assumes independent, identically distributed losses.

How to solve Generalized Extreme Value (GEV) Distribution questions

Use this order for any GEV or block maxima question. It works for conceptual, parameter and VaR questions.

1Identify the data. Confirm you are working with losses and that block maxima, not all observations, are being modelled.
2Note the block size n (for example 20 trading days for a month) and the number of blocks. Remember the trade-off: larger blocks mean less bias but fewer maxima (higher variance). Smaller blocks mean more maxima but more bias.
3Read the parameters μ, σ and ξ. Check σ > 0 and look at the sign of ξ to name the type: Fréchet, Gumbel or Weibull.
4Decide what is asked: a probability (use F(x)), a quantile (invert F), or a VaR for the original losses.
5For a quantile, set F(x) = p and solve with the right formula. Use the Gumbel form if ξ = 0.
6For a VaR on single-period losses, convert the confidence level: p = α^n. Then find x_p and read it as the VaR.
7Interpret the result in loss units and state the assumptions: iid losses, enough blocks, stable parameters.
8Sanity check: a larger ξ or σ should push the quantile higher, and a higher p should give a larger loss.

Quickest way: Sign of ξ and Gumbel shortcut

When to use it: Use for conceptual MCQs and for quantile questions where parameters are given and time is short.

Look at ξ first. Positive means Fréchet and heavy tail. Zero means Gumbel. Negative means Weibull and bounded.
If ξ = 0, use x_p = μ − σ ln(−ln p). It needs only two natural logs.
If ξ ≠ 0, compute (−ln p) first, raise it to −ξ, then plug in.
For VaR, check whether the question gives p for the maximum or α for single observations. Convert with p = α^n if needed.
Eliminate options that put the quantile below μ when p > about 0.37, because the Gumbel/GEV median of maxima sits above μ in typical cases.

Common mistakes in Generalized Extreme Value (GEV) Distribution

Treating the location and scale as the mean and standard deviation.
The names sound like normal distribution parameters.
Fix: Say location is a centring parameter and scale is a spread parameter for the maxima. They are not the mean or standard deviation, except by coincidence.
Mixing up the sign of ξ: calling ξ > 0 Weibull or thin-tailed.
Students memorise the names but not the signs.
Fix: Remember: positive ξ = fat tail = Fréchet. Zero = Gumbel. Negative = Weibull with a finite upper limit.
Using the GEV quantile directly as the VaR of daily losses.
The fitted distribution is for block maxima, not single-day losses.
Fix: Convert the confidence level with p = α^n before inverting, or state that the quantile is a block-maximum quantile.
Using the ξ ≠ 0 formula when ξ = 0 and getting a division by zero.
The general formula has ξ in the denominator.
Fix: Switch to the Gumbel formula x_p = μ − σ ln(−ln p) whenever ξ = 0.
Thinking larger blocks always improve the estimate.
Larger blocks match the theory better.
Fix: Larger blocks cut bias but leave fewer maxima, so variance rises. It is a bias-variance trade-off.
Forgetting that the method assumes independent, identically distributed observations.
Real returns show volatility clustering.
Fix: Mention that clustering breaks the assumption and may need filtering or de-clustering before applying the method.

Worked examples

Example 1

Monthly maximum daily losses for a portfolio are modelled as Gumbel with μ = 2.0 and σ = 0.6 (in % of portfolio value). Find the monthly-maximum loss quantile at p = 0.95. Use ln(−ln 0.95) = ln(0.051293) = −2.970.

Show the solution

ξ = 0, so use the Gumbel quantile: x_p = μ − σ × ln(−ln p).
−ln 0.95 = 0.051293.
ln(0.051293) = −2.970.
x_p = 2.0 − 0.6 × (−2.970) = 2.0 + 1.782 = 3.782.

Answer: The 95% quantile of the monthly maximum daily loss is about 3.78% of portfolio value.

Example 2

Weekly (5-day) maxima of daily losses follow a GEV with μ = 1.5, σ = 0.5 and ξ = 0.2 (in %). Find the 99% VaR of the daily loss. Use ln(−ln p) steps below and assume iid daily losses. Data: p = 0.99^5 = 0.95099, −ln p = 0.050150.

Show the solution

For 5 iid daily losses, P(max ≤ x) = F_daily(x)^5. We want F_daily(x) = 0.99, so p = 0.99^5 = 0.95099.
−ln p = 0.050150.
Raise to −ξ = −0.2: 0.050150^(−0.2) = exp(−0.2 × ln 0.050150). ln 0.050150 = −2.9926. So exponent = 0.59852 and exp(0.59852) = 1.8195.
x_p = μ − (σ ÷ ξ) × [1 − 1.8195] = 1.5 − (0.5 ÷ 0.2) × (−0.8195) = 1.5 + 2.5 × 0.8195.
2.5 × 0.8195 = 2.0488, so x_p = 3.549.

Answer: The 99% daily VaR is about 3.55% of portfolio value. Because ξ > 0 the distribution is Fréchet, so the tail is heavy.

Exam tips

Expect questions that give ξ and ask you to name the type. Learn the signs cold: positive Fréchet, zero Gumbel, negative Weibull.
Read whether a quantile is for the block maximum or the single-period loss. Examiners test the p = α^n conversion.
Know that the block size trade-off is bias versus variance, and that block maxima use data less efficiently than peaks over threshold.
Be ready for the interpretation: a higher ξ means fatter tails and a higher extreme VaR, and the normal belongs to the Gumbel domain.

Practice questions from Parametric Approaches (II): Extreme Value

Generalized Extreme Value (GEV) Distribution in other exams

The same ground in other exams, if you are preparing for more than one or want another angle on it.

Actuarial Core Principles (IAI)Introduction to Extreme Value Theory

Generalized Extreme Value (GEV) Distribution: frequently asked questions

What is the difference between Gumbel, Fréchet and Weibull?

They are the three cases of the GEV, set by the tail index ξ. Fréchet has ξ > 0 and a heavy tail. Gumbel has ξ = 0 and an exponential tail. Weibull has ξ < 0 and a finite upper bound.

Which GEV type fits financial losses?

Fréchet is the usual fit because returns have fat tails, like the Student t. Gumbel is a thinner-tailed approximation. Weibull is rare because losses rarely have a hard upper limit.

How do you estimate GEV parameters for VaR?

Split the losses into blocks, take each block maximum, and fit μ, σ and ξ, usually by maximum likelihood. Then convert the VaR confidence level to p = α^n and invert the GEV to get the quantile.

How do you choose the block size?

There is a trade-off. Bigger blocks fit the GEV theory better but give fewer maxima and noisier estimates. Smaller blocks give more data but more bias.