Return Levels from GEV Models

From Annual Exceedance Probability to a Return Level

A return level translates a probability from a fitted extreme-value model into the scale of the quantity being measured. For example, a flood level described as a “100-year level” is not a level guaranteed to occur only once in a century. It is a quantile whose probability of being exceeded in any one year is one percent, under the model.

Let \(M\) denote an annual maximum, such as the largest daily rainfall in a year. Suppose its distribution is modeled by a generalized extreme-value (GEV) distribution with location \(\mu\), scale \(\sigma>0\), and shape \(\xi\). For a specified annual exceedance probability \(p\), the return level is the value \(z_p\) satisfying \(P(M>z_p)=p\). Since the GEV distribution is continuous, this is the quantile at probability \(1-p\).

The return period is a convenient way to report an annual probability, not a schedule for when an event will occur. The calculation below concerns a single year’s maximum. Interpreting a multi-year probability additionally requires assumptions about how different years behave.

Inverting the GEV Distribution

For \(\xi\ne0\), the GEV cumulative distribution function is

The condition in the formula is the support condition. When \(\xi>0\), the support has lower endpoint \(\mu-\sigma/\xi\) and extends without bound above it. When \(\xi<0\), it extends without bound below the finite upper endpoint \(\mu-\sigma/\xi\). In the limiting case \(\xi=0\), the GEV distribution is Gumbel, with

To derive the formula for \(\xi\ne0\), set \(F(z_p)=1-p\). Taking logarithms gives

Here \(t_p>0\), so raising both sides of the second equation to the power \(-\xi\) is valid. Rearranging then yields

For \(\xi=0\), the equation \(F(z_p)=1-p\) instead gives \(\exp[-\exp(-(z_p-\mu)/\sigma)]=1-p\). Taking logarithms twice gives \(\exp(-(z_p-\mu)/\sigma)=t_p\), hence \(z_p=\mu-\sigma\log(t_p)\). These solutions are unique: each GEV distribution function is strictly increasing on its support, so exactly one support value has cumulative probability \(1-p\). \(\square\)

Worked Return-Level Calculations

For all three examples, take \(\mu=100\), \(\sigma=20\), and an annual exceedance probability \(p=0.01\). This corresponds to \(T=1/0.01=100\) years. First calculate the quantity shared by all three cases:

Shape, Return Period, and Interpretation

The three calculations show why the shape parameter matters. With positive shape, the modeled upper tail is unbounded and the return level can rise sharply as the target exceedance probability decreases. With negative shape, return levels approach a finite upper endpoint as \(p\) decreases. With zero shape, the Gumbel formula applies. In each case, the location shifts the level and the scale controls its magnitude, but the tail behavior is governed by the shape.

The formula also has the correct limiting form as \(\xi\) approaches zero. Write \(t_p^{-\xi}=\exp(-\xi\log t_p)\). Since \(\lim_{a\to0}(e^a-1)/a=1\),

Adding \(\mu\) recovers the Gumbel return-level formula. This confirms that the zero-shape expression is the continuous limit of the nonzero-shape expression, not a separate arbitrary convention. \(\square\)

For a fixed annual exceedance probability \(p\), independence across years gives a useful interpretation over several years. If annual maxima are independent and have the same distribution, each year has exceedance probability \(p\), and the probability of no exceedance in \(T\) years is \((1-p)^T\). Therefore the probability of at least one exceedance in those years is

For a \(T\)-year return level, \(p=1/T\), so this probability is \(1-(1-1/T)^T\), not one. For example, over 100 independent years, the chance of at least one exceedance of the 100-year level is \(1-0.99^{100}\approx0.634\). Also, the expected waiting time until the first exceedance is \(1/p\): the probability that the waiting time exceeds \(k\) years is \((1-p)^k\), and summing these tail probabilities for \(k=0,1,\ldots\) gives \(\sum_{k=0}^{\infty}(1-p)^k=1/p\).

Common Mistakes and a Practical Check

A frequent error is to substitute the annual exceedance probability \(p\) directly as the CDF probability. The return level must satisfy \(F(z_p)=1-p\), not \(F(z_p)=p\). For a rare high event, \(p\) is small and the relevant quantile is near the upper end of the distribution, so reversing these probabilities can produce a level with the opposite interpretation.

Another mistake is to read “100-year level” as “one event every 100 years.” The return period is the reciprocal of the annual exceedance probability; it neither fixes the timing of events nor guarantees one exceedance in a period of that length. The multi-year probability calculation above also depends on independence and a constant annual exceedance probability. If those assumptions do not hold, that calculation does not follow just from the return-period label.

Before reporting a return level, check the probability convention, the shape-specific formula, and the support. For \(\xi<0\), explicitly compare the result with \(\mu-\sigma/\xi\). Finally, distinguish the model-based return level from an observed maximum: the level is a quantile implied by a fitted distribution, not a prediction that a particular future year will attain it.