Skip to content

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

Peaks Over Threshold and the Generalized Pareto Distribution

Updated 11 October 2026 · Fact-checked

The peaks over threshold (POT) approach picks a high threshold u and models only the losses that exceed it. The excesses (X − u) are fitted with a Generalized Pareto Distribution (GPD) using a shape ξ and scale β. The key choice is u: too low adds bias, too high leaves too few data points.

Understand Peaks Over Threshold and Generalized Pareto Distribution

Extreme value theory studies the tail, not the centre, of a loss distribution. There are two main ways to do it. The block maxima method splits data into blocks, such as months, and fits the Generalized Extreme Value (GEV) distribution to each block's maximum. The peaks over threshold (POT) method does something different.

POT chooses a high threshold u and looks at every observation above it. The amount by which each one exceeds u is called an excess or exceedance: Y = X − u. A key theorem (Pickands-Balkema-de Haan) says that for a high enough threshold, the distribution of these excesses is approximately a Generalized Pareto Distribution (GPD). This holds for a wide class of underlying loss distributions.

The GPD has two parameters. The shape (tail index) ξ controls how heavy the tail is. ξ > 0 gives a heavy, fat tail that decays like a power law, which is typical for financial losses. ξ = 0 gives an exponential tail. ξ < 0 gives a tail with a finite upper limit. The scale β (must be positive) sets the size of the excesses.

POT uses data more efficiently than block maxima. Block maxima keeps one point per block and discards other large losses. POT keeps every extreme loss above u. That is why POT is usually preferred for risk measures such as VaR and Expected Shortfall.

The central practical issue is the threshold trade-off. If u is too low, the GPD approximation is poor and the estimates are biased, because non-tail data enters the fit. If u is too high, very few exceedances remain and the parameter estimates have large variance. You look for a threshold that is high enough for the theory to hold but low enough to leave enough data. Tools include the mean excess plot, which should look roughly linear above a good u, and checking that the estimates of ξ are stable as u changes. The Hill estimator is a related estimate of the tail index for heavy tails, and it faces the same choice of how many top observations k to use.

Key formulas to remember

Excess over threshold
Y = X − u, for X > u
POT models only these excesses, not the full loss.
GPD cumulative distribution
G(y) = 1 − (1 + ξy ÷ β)^(−1/ξ), for ξ ≠ 0, y ≥ 0, β > 0
For ξ < 0, y must also satisfy y ≤ −β ÷ ξ.
GPD with ξ = 0
G(y) = 1 − e^(−y ÷ β)
Exponential tail, the limit as ξ approaches 0.
Tail probability estimate
P(X > x) ≈ (Nu ÷ n) × (1 + ξ(x − u) ÷ β)^(−1/ξ), for x > u
Nu is the number of exceedances and n is the total sample size. Nu ÷ n estimates P(X > u).
Hill estimator
ξ̂ = (1 ÷ k) × Σ ln(X(i) ÷ X(k+1)), i = 1 to k
X(1) ≥ X(2) ≥ ... are ordered losses. It applies to heavy tails (ξ > 0). The choice of k is the threshold problem again.
Tail index and shape
α = 1 ÷ ξ, for ξ > 0
A smaller α means a heavier tail. The k-th moment exists only if α > k.

How to solve Peaks Over Threshold and Generalized Pareto Distribution questions

Use this order for any POT or GPD question. Most questions test one step, so find which one.

  1. 1Identify the method. If the data are block maxima, the model is GEV. If the data are exceedances above a threshold, the model is GPD.
  2. 2Note the threshold u, total sample n and number of exceedances Nu. Estimate P(X > u) as Nu ÷ n.
  3. 3Read the shape ξ and scale β. Interpret ξ: positive means a fat tail, zero means exponential, negative means a bounded tail.
  4. 4If asked about the threshold, apply the trade-off: lower u means more data and lower variance but more bias; higher u means less bias but higher variance.
  5. 5For a tail probability, use P(X > x) ≈ (Nu ÷ n) × (1 + ξ(x − u) ÷ β)^(−1/ξ). Compute the bracket first, then the power.
  6. 6Check the result is sensible: it must be below Nu ÷ n for any x above u, and fall as x rises.
  7. 7State the conclusion in risk terms, such as the probability of a loss beyond a given level.

Quickest way: Threshold-trade-off shortcut

When to use it: Use this on conceptual questions asking what happens when the threshold is raised or lowered, or which method suits the data.

  1. Raise u: fewer exceedances, higher variance, lower bias.
  2. Lower u: more exceedances, lower variance, higher bias.
  3. Excesses over a threshold means POT with GPD. Block maxima means GEV.
  4. Positive ξ means fat tail and a power-law decay.
  5. Mean excess plot roughly linear and upward sloping above u supports a GPD with ξ > 0.

Common mistakes in Peaks Over Threshold and Generalized Pareto Distribution

  • Reversing the threshold trade-off, saying a higher threshold lowers variance.

    Students link 'higher' with 'more accurate' because a higher threshold fits the theory better.

    Fix: A higher threshold means fewer data points, so variance rises. It cuts bias, not variance. Lower threshold: low variance, high bias.

  • Mixing up GEV and GPD.

    Both are extreme value distributions and the names are similar.

    Fix: GEV models block maxima. GPD models excesses over a threshold. Ask what the data are: maxima per block or exceedances.

  • Applying the GPD to the loss X instead of the excess X − u.

    The tail formula looks like a distribution of losses.

    Fix: Subtract u first. The GPD describes y = x − u. Then multiply by Nu ÷ n to get an unconditional tail probability.

  • Forgetting the factor Nu ÷ n in the tail probability.

    The GPD gives a probability conditional on exceeding u, and students treat it as unconditional.

    Fix: Always multiply the GPD tail by Nu ÷ n, the estimated P(X > u).

  • Reading ξ < 0 as a fat tail.

    Students recall that the sign matters but not which way.

    Fix: ξ > 0 is fat and heavy. ξ = 0 is exponential. ξ < 0 has a finite upper bound. Financial losses are usually ξ > 0.

  • Treating the Hill estimator as valid for any tail.

    It is taught as a general tail index tool.

    Fix: Hill applies to heavy tails with ξ > 0. It is also sensitive to the number of top observations k, which is the same bias-variance choice.

Worked examples

Example 1

A bank has 1,000 daily loss observations. With a threshold u = 2 (USD million), 50 losses exceed u. The GPD fitted to the excesses has ξ = 0.2 and β = 0.8. Estimate the probability that a daily loss exceeds USD 4 million.

Show the solution
  1. Nu ÷ n = 50 ÷ 1,000 = 0.05.
  2. Excess: x − u = 4 − 2 = 2.
  3. Compute ξ(x − u) ÷ β = 0.2 × 2 ÷ 0.8 = 0.5.
  4. Bracket: 1 + 0.5 = 1.5.
  5. Exponent: −1 ÷ ξ = −5.
  6. 1.5^(−5) = 1 ÷ 7.59375 = 0.13169.
  7. P(X > 4) ≈ 0.05 × 0.13169 = 0.006584.

Answer: About 0.66%, or roughly 0.0066.

Example 2

An analyst raises the threshold in a POT analysis from the 90th percentile to the 99th percentile of losses on a sample of 2,000 observations. Describe the effect on the number of exceedances and on the bias and variance of the GPD estimates.

Show the solution
  1. At the 90th percentile, about 10% of 2,000 = 200 observations exceed u.
  2. At the 99th percentile, about 1% of 2,000 = 20 observations exceed u.
  3. Fewer exceedances means the parameters ξ and β are estimated from less data, so variance rises.
  4. A higher threshold is closer to the true tail, so the GPD approximation holds better and bias falls.

Answer: Exceedances fall from about 200 to about 20. Bias falls but variance of the GPD estimates rises.

Exam tips

  • Expect conceptual questions on the threshold trade-off. Memorise: higher u, lower bias, higher variance.
  • Always identify whether the data are block maxima (GEV) or exceedances (GPD) before doing anything else.
  • In tail probability calculations, do not skip Nu ÷ n and use x − u, not x.
  • Know the sign of ξ and what each case means for tail thickness.
  • If the question mentions a mean excess plot, a linear upward slope above u points to a GPD with a positive shape.

Practice questions from Parametric Approaches (II): Extreme Value

Peaks Over Threshold and Generalized Pareto Distribution in other exams

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

Peaks Over Threshold and Generalized Pareto Distribution: frequently asked questions

What is the difference between GEV and POT?

GEV models the maximum of each block of data, such as the worst loss each month. POT models all losses above a high threshold using the GPD. POT uses more of the extreme data, so it is often more efficient.

How do you choose the threshold in extreme value theory?

You balance bias and variance. A low threshold gives many data points but a poor GPD fit. A high threshold gives a better fit but few observations. Practitioners use the mean excess plot and check that the estimates are stable across thresholds.

What does the shape parameter ξ tell you?

It measures tail heaviness. ξ > 0 means a fat, power-law tail, ξ = 0 an exponential tail and ξ < 0 a bounded tail. Financial losses usually have ξ > 0.

What is the Hill estimator?

It estimates the tail index of a heavy-tailed distribution from the k largest observations. It is the average of the log ratios of those losses to the (k+1)-th largest loss. Like the POT threshold, the choice of k trades off bias and variance.