Exceedance Probabilities and Return Periods

Introduction

A loss threshold is often more useful for a risk decision than a description of the entire loss distribution. For example, a decision-maker may ask how likely an annual loss is to exceed a specified amount, or how often an event of that size might be expected. These questions require care: the probability is tied to a defined time period, and a return period is an average waiting time, not a timetable.

This tutorial uses \(X\) for the loss in one year and \(x\) for a specified threshold. The annual exceedance probability is \(P(X>x)\). Under a stationary model with independent annual losses, its reciprocal is the mean waiting time, in years, between exceedances of that threshold. The calculations are aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus.

Definition: For an annual loss \(X\) and threshold \(x\), the annual exceedance probability (AEP) is \(p_x=P(X>x)\). If \(F_X(x)=P(X\leq x)\), then \(p_x=1-F_X(x)\). When annual exceedance events are independent and have the same probability \(p_x>0\) each year, the return period is \(T_x=1/p_x\) years: the expected waiting time until the next annual exceedance.

Calculating an Exceedance Probability

The basic calculation depends on the loss distribution and on whether the threshold is included in the event. For a continuous loss distribution, the probability of a loss exactly equal to \(x\) is zero, so \(P(X>x)=P(X\geq x)\). For a discrete distribution, the loss can have positive probability at \(x\), and these two events need not have the same probability.

When the distribution is given by a probability table, add the probabilities of all loss values strictly greater than \(x\). When it is given by a cumulative distribution function, subtract \(F_X(x)\) from one. The latter works for strict exceedance because \(F_X(x)\) includes all outcomes equal to or less than \(x\).

Worked Example: A Discrete Annual Loss Threshold

A hypothetical insurer models its annual loss \(X\), in thousands of dollars, as follows. The model assigns probabilities \(0.50, 0.25, 0.20,\) and \(0.05\) to losses of \(0, 40, 100,\) and \(250\), respectively. Find the annual probability of a loss exceeding \(100\) thousand dollars and the associated return period. The probabilities sum to \(1.00\). No discounting or mortality assumptions apply; the model concerns one year's loss.

Method. For a discrete distribution, sum the probabilities attached to values strictly above the threshold. The only loss value greater than \(100\) is \(250\), with probability \(0.05\). Thus \(p_{100}=0.05\).

The return period is the reciprocal of this annual probability:

T_{100}=\frac{1}{0.05}=20\text{ years}.

Check. An annual exceedance probability of \(0.05\) means a 5% chance in each year, or one expected exceedance per 20 years under the stated model. The loss of exactly \(100\) is not counted because the event is “exceeding \(100\),” not “at least \(100\).” The common mistake is to include its \(0.20\) probability and report \(0.25\); that would calculate \(P(X\geq100)\), not \(P(X>100)\).

For a continuous distribution with density \(f_X\), the same calculation can be written as the area to the right of the threshold:

P(X>x)=\int_x^\infty f_X(t)\,dt=1-F_X(x).

This expression assumes the density describes the full distribution of the annual loss. If the loss model has a point mass as well as a continuous component, use the full cumulative distribution function, including any mass at the threshold.

Worked Example: Exceedance Under an Exponential Loss Model

For a hypothetical portfolio, annual loss \(X\) follows an exponential distribution with rate parameter \(\lambda=1/200\) per thousand dollars. Equivalently, the scale parameter is \(200\) thousand dollars. Find the probability that annual loss exceeds \(300\) thousand dollars and the corresponding return period. This is a continuous annual-loss model; no discounting or mortality assumptions apply.

Method. An exponential distribution with rate \(\lambda\) has survival function \(P(X>x)=e^{-\lambda x}\) for \(x\geq0\). The units agree: \(\lambda x=(1/200)(300)=1.5\), a dimensionless quantity.

Therefore,

p_{300}=e^{-(1/200)(300)}=e^{-1.5}=0.223130\ldots

To four decimal places, the annual exceedance probability is \(0.2231\). Its reciprocal is

T_{300}=\frac{1}{0.223130\ldots}=4.4817\text{ years}.

Check. The probability is between zero and one, and the return period is greater than one year, as expected for an event with annual probability below one. A common error is to use \(1-e^{-\lambda x}\), which is the probability \(P(X\leq x)\), not the probability of exceeding \(x\).

Why the Reciprocal Is a Return Period

A return period is derived from the waiting time to the first exceedance. Suppose the annual exceedance probability is a fixed \(p\), with \(0<p\leq1\), and each year's event is independent of the others. Let \(W\) be the number of years until the first year in which the loss exceeds \(x\). The probability that the first exceedance occurs in year \(k\) is \((1-p)^{k-1}p\): the first \(k-1\) years do not exceed, and year \(k\) does.

Thus \(W\) has a geometric distribution. A direct derivation of its mean uses the tail-sum formula for a positive integer-valued waiting time: \(E[W]=\sum_{k=0}^{\infty}P(W>k)\). An exceedance has not occurred after \(k\) years precisely when all \(k\) years are non-exceedance years, which has probability \((1-p)^k\). Therefore,

E[W]=\sum_{k=0}^{\infty}(1-p)^k =\frac{1}{1-(1-p)} =\frac{1}{p}.

The infinite geometric series converges because \(0<p\leq1\). This shows why the reciprocal is the mean waiting time; it does not say an exceedance will occur exactly every \(1/p\) years. Several exceedances can occur close together, or a long period can pass with none.

Define the annual event.
Specify the loss measure, its one-year period, the threshold, and whether the event is \(X>x\) or \(X\geq x\).

Calculate its annual probability.
Use the probability table or distribution function, taking account of any probability mass at the threshold.

Interpret the reciprocal conditionally.
If annual exceedance probabilities are constant and events are independent across years, report \(1/p\) as the expected waiting time in years.

Several-Year Probabilities and Return-Period Interpretation

The return period alone does not give the probability of at least one exceedance over a specified number of years. Under the same independent, stationary annual model, the probability of no exceedance in any of \(n\) years is \((1-p)^n\). Taking the complement gives:

P(\text{at least one exceedance in }n\text{ years})=1-(1-p)^n.

This is not generally equal to \(n p\). The expression \(np\) is the expected number of exceedance years in \(n\) years under the model; it can exceed one and is not a probability of at least one event. The complement formula accounts for the possibility of multiple exceedance years.

Worked Example: Chance of an Exceedance Within Five Years

A hypothetical risk model assigns a \(0.08\) probability in each year that annual loss exceeds a stated threshold. The annual exceedance events are assumed independent, and the probability remains unchanged from year to year. Find (1) the return period and (2) the probability of at least one exceedance during the next five years. No discounting or mortality assumptions apply.

Method. First take the reciprocal of the annual probability:

T=\frac{1}{0.08}=12.5\text{ years}.

For the five-year probability, the chance of no exceedance in one year is \(1-0.08=0.92\). Independence means the probability of no exceedance in all five years is

$$ (0.92)^5=0.6590815232. $$

Consequently, the probability of at least one exceedance is

1-(0.92)^5=1-0.6590815232=0.3409184768\approx0.3409.

Check. The result is a probability between zero and one, and it is greater than the one-year probability of \(0.08\). Reporting \(5(0.08)=0.40\) as the probability of at least one exceedance is incorrect; that product is the expected number of exceedance years over five years, not the probability of at least one.

Assumptions, Uses, and Common Traps

A return period is meaningful only after defining what is being repeated. The threshold must refer to a specific loss measure and consistent units, and the period must be stated. A return period of 20 years for an annual loss threshold does not describe the return period for a monthly loss or for a different loss measure.

The reciprocal interpretation also relies on assumptions about time. A constant annual exceedance probability is a stationarity assumption; independence says that an exceedance in one year does not change the probability in another. If annual risks trend, cluster, or are otherwise dependent, \(1/p\) may still be the reciprocal of a one-year marginal probability, but it need not equal the expected waiting time. A multi-year calculation must then use a model for the joint behavior across years rather than automatically applying \(1-(1-p)^n\).

Confusing strict and inclusive thresholds. For a discrete loss, \(P(X>x)=1-F_X(x)\), whereas \(P(X\geq x)\) includes any probability mass at \(x\).
Calling a return period a schedule. A 20-year return period does not guarantee one exceedance in every block of 20 years.
Ignoring the annual model. A probability for one year's loss cannot be interpreted as a return period without specifying the annual event and, for the waiting-time interpretation, the assumptions on years.
Using a linear approximation as an exact probability. For independent years, use \(1-(1-p)^n\) for at least one exceedance, not \(np\).

Key takeaway: Calculate \(P(X>x)\) using the correct tail and threshold convention. Its reciprocal is the expected waiting time only under a constant annual exceedance probability and independent years; it is not a promise about when an exceedance will occur.

Check Your Understanding

For each problem, assume losses and probabilities are hypothetical. Give probability answers to four decimal places where needed.

An annual loss \(X\), in thousands of dollars, equals \(0, 75, 150,\) or \(400\) with probabilities \(0.40, 0.30, 0.20,\) and \(0.10\), respectively. Find \(P(X>150)\) and its return period under independent, stationary years. (Answer: \(0.1000\); 10 years.)
An annual loss has a uniform distribution from \(0\) to \(500\) thousand dollars. Find the annual probability of exceeding \(350\) thousand dollars and the return period. (Answer: \(0.3000\); 3.3333 years.)
The annual probability of exceeding a threshold is \(0.04\), independently from year to year. Find the probability of at least one exceedance in six years. (Answer: \(1-(0.96)^6=0.2172\).)
An annual loss equals \(0, 200,\) or \(600\) thousand dollars with probabilities \(0.70, 0.20,\) and \(0.10\). Find both \(P(X>200)\) and \(P(X\geq200)\). (Answer: \(0.1000\) and \(0.3000\), respectively.)
A threshold has a 25-year return period under an independent, stationary annual model. Find the probability of at least one exceedance in the next 12 years. (Answer: \(1-(0.96)^{12}=0.3873\).)

What You'll Learn