James presents the Meridian model run results at the Monday pricing meeting. He puts a single line on the screen: "1-in-100 year loss: USD 38M."
The active underwriter nods and says: "So roughly once a century we'd expect a USD 38M loss on this account. We can work with that."
The reinsurance actuary shakes her head. "That's not quite what it means," she says. "It means there's a 1% chance of exceeding USD 38M in any given year. We could see it again next year. We could go 200 years without seeing it. The century framing is misleading."
Both of them are looking at the same number. One of them understands what it actually says.
A return period is a way of expressing the probability of a loss exceeding a specific threshold in any given year. A 1-in-100 year loss is a loss level that has a 1% probability of being exceeded in any single year. That is all it is, a probability statement about a threshold, expressed as a reciprocal to make large numbers easier to read.
The phrase "1-in-100 year" does not mean the event occurs once per century. It does not mean the event is scheduled. It does not mean that if it occurred last year, it is unlikely to occur this year. It says nothing about when the event will happen, only about how probable it is in any given year. A 1% annual probability is consistent with two occurrences in the same decade, or no occurrence in three centuries. The probability is the same every year, regardless of what happened before.
The return period is the reciprocal of the annual exceedance probability. A 1% annual probability corresponds to a 1-in-100 year return period. A 0.5% probability corresponds to 1-in-200. A 2% probability corresponds to 1-in-50. The relationship is simply: return period = 1 ÷ annual exceedance probability. The return period is a convenience notation, not a calendar prediction.
This implies a regular schedule, that the event occurs on a roughly 100-year cycle, and if it happened last year, you are now "safe" for another 99 years. It suggests that two occurrences in a decade are evidence the model is wrong, and that a long quiet period means a big event is "due." None of these implications are correct.
Each year is an independent draw. The probability is 1% regardless of what happened in previous years. Over a 100-year horizon, the probability of at least one exceedance is approximately 63%, not 100%. Two exceedances in a decade are unusual but entirely consistent with a 1% annual probability. No event is ever "due."
An underwriter who believes the 1-in-100 year loss is a 100-year scheduled event will underprice risk after a major loss, assuming the clock has "reset" and the next large event is far away. A portfolio manager who believes the 1-in-200 year loss cannot occur again for 200 years after it occurs will underhold capital. These are not hypothetical errors, they reflect a genuine and widespread misunderstanding of what return periods say. Getting this right is one of the most important conceptual shifts a new cat modelling practitioner makes.
A single return period figure, the 1-in-100 year loss, is one point on a curve. The Exceedance Probability (EP) curve is that complete curve: a mapping from every possible loss level to the annual probability of exceeding it. It shows not just what might happen at the 1-in-100 year level, but the full shape of the annual loss distribution, from frequent, moderate losses at low return periods to rare, extreme losses at high return periods.
The EP curve is what the cat model actually produces. The stochastic catalogue, tens of thousands of synthetic events each with an annual rate, generates a loss estimate for the portfolio for every event. Sorting those losses from smallest to largest and assigning cumulative probabilities produces the EP curve. A return period figure like "1-in-100 year loss: USD 38M" is simply a single coordinate read from that curve.
The shape of the EP curve is itself informative. A curve that rises steeply at low return periods and then flattens, like the one above, indicates a portfolio dominated by frequent moderate losses with a relatively constrained tail. A curve that remains flat at low return periods and then rises sharply suggests a portfolio where most of the risk is concentrated in rare, extreme events with very little attritional contribution. The shape tells you something about the character of the risk before you read a single number off the axis.
The EP curve is not a prediction of what will happen in any specific year. It is a probability distribution, a map of all possible outcomes and their relative likelihoods. In any given year, the actual loss will land somewhere on the horizontal axis. Across many years, the frequency with which losses exceed each threshold should correspond to the curve's probabilities. But in any single year, any outcome is possible.
The EP curve can be constructed in two fundamentally different ways, depending on what question you are trying to answer. Both use the same stochastic catalogue and the same loss estimates, but they aggregate those losses differently before building the curve.
Asks: what is the probability that the single largest event in a year causes a loss exceeding a given threshold? Each simulated year contributes only its worst event to the curve, all smaller events in that year are ignored.
OEP describes the tail risk from individual catastrophic events. A 1-in-100 year OEP loss is the level exceeded by the single worst event in the year with 1% annual probability.
Asks: what is the probability that the sum of all losses in a year exceeds a given threshold? Every event in a simulated year contributes to the annual aggregate, ten moderate events are summed together.
AEP describes the total annual loss risk from all events combined. A 1-in-100 year AEP loss is the total annual loss level exceeded with 1% probability, accounting for the possibility of multiple events in a year.
OEP is always equal to or lower than AEP at any given return period. This must be true mathematically: the single worst event in a year cannot exceed the sum of all events in that year. For perils dominated by a single large event, a major earthquake or a Category 5 hurricane, OEP and AEP are often close together, because one event typically drives most of the annual loss. For perils with many moderate events per year, severe convective storms, flood, AEP can be substantially higher than OEP at the same return period, because a bad year can be the accumulation of many events rather than one catastrophic one.
Meridian's portfolio includes locations across North America, Europe, Asia-Pacific, and the Middle East. In any given year, multiple peril events could each cause moderate losses across different regions simultaneously, a European windstorm in January, a US tornado outbreak in May, a Japanese typhoon in September. The AEP curve captures the possibility that all three hit in the same year and their combined loss exceeds a threshold that no single event would reach alone. The OEP curve only captures the worst single event. For a geographically diversified portfolio like Meridian's, the difference between OEP and AEP at the same return period can be significant.
Each year is an independent probability draw. The occurrence of a 1-in-100 year event does not reduce the probability of another one next year. Seismic stress, atmospheric conditions, and hydrological systems do not "reset" after a major event. The probability is 1% in every year, unconditionally.
Twice the return period means half the annual probability, 0.5% vs 1%. But twice as rare in annual probability terms does not mean twice as unlikely to occur in a 10-year or 20-year window. Over a 10-year horizon, the probability of at least one exceedance of a 1-in-100 year loss is about 10%; for a 1-in-200 year loss it is about 5%. The ratio between those probabilities is still roughly 2:1, but neither is "unlikely" over a professional career.
They are constructed differently and answer different questions. OEP is relevant when the risk is driven by single large events, the worst thing that could happen in a year. AEP is relevant when the risk accumulates across multiple events, the total cost of a bad year. For many portfolios, OEP and AEP diverge materially, and using one where the other is appropriate produces incorrect answers to the question being asked.
The EP curve is a probability distribution, not a forecast. It does not predict that a 1-in-100 year loss will occur in 100 years, it says that in any given year, there is a 1% chance of exceeding that loss level. Over many years, actual losses should track the distribution on average. In any specific year, any outcome is possible. The EP curve tells you about the space of possible outcomes and their relative likelihoods, not about which specific outcome will occur.
The reciprocal of the annual exceedance probability. A 1-in-100 year return period means a 1% annual probability of exceedance. A statement about probability, not about scheduling. Does not imply the event occurs once per return period or that occurrences are evenly spaced.
The probability that a loss exceeds a given threshold in any given year. Equal to 1 ÷ return period. A 1-in-100 year return period corresponds to a 1% annual exceedance probability. The fundamental metric from which return periods are derived.
Exceedance Probability Curve. A complete mapping from every possible loss level to its annual exceedance probability. The primary output of a cat model. Return period figures are individual points read from this curve.
Occurrence Exceedance Probability. The EP curve built from the single largest event in each simulated year. Describes the tail risk from individual catastrophic events. Used for Cat XL pricing and RDS comparison.
Aggregate Exceedance Probability. The EP curve built from the sum of all events in each simulated year. Describes total annual loss risk including accumulation across multiple events. Used for capital modelling and aggregate reinsurance structures.
The synthetic database of simulated events, each with an assigned annual rate, from which the EP curve is constructed. Introduced in Lesson 2.1. The EP curve is built by running the full portfolio through every event in the catalogue and aggregating the results.
Five questions · 4 of 5 correct (80%) to pass