Return Periods and Exceedance Probabilities

Annual Exceedance Probability and Return Period

An exceedance curve reports the probability that annual loss is greater than a selected threshold. The reciprocal of that annual probability is commonly expressed as a return period. This conversion is simple, but its interpretation matters: a return period summarizes a long-run probability; it does not predict a calendar schedule for events.

For an occurrence exceedance probability (OEP), the annual loss is the largest event loss in the year. For an aggregate exceedance probability (AEP), it is the total loss from all events in the year. The return-period conversion is the same for either curve, but the annual loss quantity and therefore the probability at a given threshold can differ.

Definition: At threshold $x$, let $p(x)$ be the annual exceedance probability: the probability that the annual loss measure exceeds $x$. The return period at $x$ is $T(x)=1/p(x)$, measured in years, provided $p(x)>0$. Conversely, an annual exceedance probability associated with return period $T$ is $p=1/T$. State whether $p(x)$ is an OEP or AEP probability.

For example, an annual exceedance probability of $0.04$ corresponds to a return period of $1/0.04=25$ years. It means that the loss exceeds the threshold in about 4% of years under the modeled annual probability. It does not mean that an exceedance is due exactly once every 25 years, or that the next exceedance will occur 25 years from now.

From an Annual Probability to a Multi-Year Probability

A return period alone does not determine the chance of an exceedance over several years unless the dependence between annual outcomes is specified. In the standard calculation, assume the annual exceedance probability $p$ is constant from year to year, and that exceedance outcomes in different years are independent. Under these assumptions, each year is a Bernoulli trial: exceedance occurs with probability $p$, and no exceedance occurs with probability $1-p$.

P(\text{no exceedance in }n\text{ years})=(1-p)^n

The event “at least one exceedance” is the complement of “no exceedances.” Therefore,

P(\text{at least one exceedance in }n\text{ years}) =1-(1-p)^n =1-\left(1-\frac{1}{T}\right)^n

This is not generally equal to $n/T$. The expression $n/T=np$ is the expected number of exceedance years over $n$ independent years, not the probability of one or more exceedances. It can even exceed 1, which a probability cannot do.

Assumption: Use $1-(1-p)^n$ when annual exceedance outcomes are independent and have the same probability $p$ in every year. If annual probabilities differ but outcomes remain independent, the probability of at least one exceedance is $1-\prod_{j=1}^{n}(1-p_j)$. If outcomes are dependent, the annual probability by itself is not enough to determine the multi-year probability.

Why the Formula Holds

The multi-year formula follows from the complement rule and independence. Let $E_j$ denote an exceedance in year $j$, so $P(E_j)=p$. The probability of no exceedance in one year is $P(E_j^c)=1-p$. Independence means the probability of no exceedance in all $n$ years is the product of the annual no-exceedance probabilities.

P\left(\bigcap_{j=1}^{n} E_j^c\right) =\prod_{j=1}^{n}P(E_j^c) =(1-p)^n

Taking the complement gives $P(\text{at least one exceedance})=1-(1-p)^n$. This derivation also shows exactly where the independence and constant-probability assumptions enter; without them, the product need not be valid.

Worked Example: Comparing OEP and AEP Return Periods

A hypothetical catastrophe model reports, at a loss threshold of $40 million, an annual OEP exceedance probability of $0.10$ and an annual AEP exceedance probability of $0.05$. For each curve, calculate the return period and the probability of at least one annual exceedance in the next five years. Assume annual outcomes are independent and the relevant annual probability is constant. No interest rate or mortality assumption is relevant; the annual exceedance indicator has a Bernoulli distribution with the stated probability.

Method. Convert each annual probability using $T=1/p$, then apply the multi-year complement formula separately to each curve.

OEP return period. The annual OEP probability is $p_{\mathrm{OEP}}=0.10$, so

T_{\mathrm{OEP}}=\frac{1}{0.10}=10\text{ years}.

The probability of no OEP exceedance in five years is $0.90^5=0.59049$. Thus, the probability of at least one is $1-0.59049=0.40951$, or $0.4095$ to four decimal places.

AEP return period. The annual AEP probability is $p_{\mathrm{AEP}}=0.05$, so

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

The probability of no AEP exceedance in five years is $0.95^5=0.7737809375$. Therefore, the probability of at least one is $1-0.7737809375=0.2262190625$, or $0.2262$ to four decimal places.

Check. Each five-year probability is greater than its corresponding one-year probability, but less than 1. The OEP and AEP return periods differ because their annual exceedance probabilities at the threshold differ. The common mistake is to apply one curve’s probability to the other curve or to call both probabilities “the chance of a $40 million event” without identifying the annual loss measure.

Return Period Is Not an Event Schedule

Under a constant annual probability and independent years, the number of years until the next exceedance has a geometric distribution. If $W$ is the number of years up to and including the first exceedance, then $P(W=k)=(1-p)^{k-1}p$, for $k=1,2,\ldots$. Its expected value is $E[W]=1/p=T$. Thus, in this particular model, the return period is also the mean waiting time until an exceedance.

A mean is not a promise about timing. Some exceedances may occur in consecutive years; other intervals may be much longer than the mean. Even after a long period with no exceedance, the probability in the next year remains $p$ under the independent, constant-probability assumption. The prior quiet years do not make an exceedance “due.”

A related but different quantity is the median waiting time: the smallest whole number of years by which the probability of at least one exceedance reaches at least 50%. It need not equal the return period. The multi-year formula, rather than the reciprocal alone, determines this probability.

Worked Example: A 25-Year Return Period Over a 10-Year Window

A hypothetical flood portfolio has an annual AEP exceedance probability of $0.04$ at a specified aggregate-loss threshold. Calculate the probability of no exceedances, exactly one exceedance, and at least two exceedance years in a 10-year period. Assume independent years with a constant annual probability. The corresponding annual indicator is Bernoulli with $p=0.04$; the 10-year count has a binomial distribution with $n=10$ and $p=0.04$. No interest rate or mortality assumption is relevant.

Method. First convert the probability to a return period. Then use the binomial probabilities for the number of exceedance years.

Return period. $T=1/0.04=25$ years.

No exceedances. The probability is

P(N=0)=(1-0.04)^{10}=0.96^{10}=0.66483263599\approx 0.6648.

Exactly one exceedance year. There are 10 choices for which year has the exceedance, so

P(N=1)=\binom{10}{1}(0.04)(0.96)^9 =10(0.04)(0.6925339957) =0.2770135983\approx 0.2770.

At least two exceedance years. Subtract the probabilities of zero and one from 1:

P(N\ge 2)=1-0.66483263599-0.2770135983 =0.0581537657\approx 0.0582.

Check. The three mutually exclusive probabilities sum to approximately $0.6648+0.2770+0.0582=1.0000$. The common mistake is to calculate $10/25=0.40$ and report it as the probability of at least one exceedance. Here $0.40$ is the expected number of exceedance years; the probability of at least one is $1-P(N=0)=1-0.66483263599=0.3352$.

Worked Example: Mean Versus Median Waiting Time

For a hypothetical location, the annual OEP exceedance probability at a loss threshold is $0.10$ in every year, with independent annual outcomes. Find the return period, the probability of at least one exceedance in six and seven years, and the smallest whole number of years by which that probability reaches 50%. The annual exceedance indicator is Bernoulli with $p=0.10$; no interest rate or mortality assumption is relevant.

Method. The return period is the reciprocal of the annual probability. For each horizon, calculate the complement of no exceedance.

Return period and mean waiting time.

T=\frac{1}{0.10}=10\text{ years}.

Under the independent, constant-probability model, the waiting time is geometric, so its mean is $E[W]=1/p=10$ years.

Six-year probability.

P(W\le 6)=1-0.90^6=1-0.531441=0.468559\approx 0.4686.

Seven-year probability.

P(W\le 7)=1-0.90^7=1-0.4782969=0.5217031\approx 0.5217.

The probability is below 50% at six years and above 50% at seven years, so the median waiting time is seven years. This differs from the 10-year mean waiting time. The common mistake is to treat “10-year return period” as a guarantee of one exceedance by year 10; the actual probability by then is $1-0.90^{10}=0.6513$, not 1.

Practical Interpretation and Exam Traps

When reading or reporting a return period, keep the threshold, annual loss measure, and probability convention attached to the number. A “25-year return period” is incomplete unless it is clear which loss threshold is being exceeded and whether the underlying curve is OEP or AEP. It is also important to distinguish an annual exceedance probability from a multi-year probability.

Do not read a return period as a timetable. A 25-year return period does not mean an event occurs once in each fixed 25-year block.
Do not multiply $p$ by the number of years to get a probability of at least one. Use $1-(1-p)^n$ under the stated independence and constant-probability assumptions.
Do not mix OEP and AEP. Apply the reciprocal to the annual probability from the specific curve being interpreted.
Check assumptions before using a multi-year formula. A stable annual probability and independent annual outcomes are needed for the standard calculation.

Return periods provide a convenient way to communicate annual exceedance probabilities, while multi-year probabilities answer questions about a stated time horizon. Keeping those two uses separate avoids interpreting a long-run average as a schedule or mistaking an expected count for a probability.

Check Your Understanding

For multi-year calculations, assume independent years and constant annual exceedance probability unless a question says otherwise. Round probabilities to four decimal places.

A hypothetical AEP curve gives an annual exceedance probability of $0.025$ at a specified threshold. Find the return period and the probability of at least one exceedance in eight years. (Answer: 40 years; 0.1834.)
A loss threshold has a 20-year return period. What is the probability of no exceedance in the next three years? (Answer: 0.8574.)
A modeled annual OEP exceedance probability is $0.08$. Calculate the probability of at least one exceedance in four years. (Answer: 0.2836.)
A hypothetical annual exceedance probability is $0.03$. What is the smallest whole number of years for the probability of at least one exceedance to reach or exceed 50%? (Answer: 23 years.)
At one threshold, a hypothetical model reports OEP probability $0.12$ and AEP probability $0.04$. Give each return period and explain why the values differ. (Answer: OEP 8.33 years; AEP 25 years; the annual probabilities differ because OEP and AEP refer to different annual loss measures.)

What You'll Learn