"Hundred-year flood" is the most used phrase in the field and the most misunderstood. It does not say that such an event occurs once a century, nor that it cannot occur two years running. It says that at a given point, for a given quantity, the probability of reaching or exceeding that level is one percent in any given year. It is an annual probability disguised as a duration, and the disguise alone produces half the misunderstandings with clients.
The first consequence is arithmetic and it always surprises. If each year carries one chance in a hundred, the probability of seeing at least one such event over thirty years is not thirty percent but about twenty-six percent, and over a hundred years it comes to around sixty-three percent, not a hundred. Conversely, nothing forbids two occurrences in ten years: the probability is low, it is not zero, and across a country's thirty-five thousand or so communes, the event that is improbable somewhere is a certain event everywhere.
The second consequence is that a return period attaches to ONE POINT and ONE QUANTITY, never to an event and never to a portfolio. One storm can be a thirty-year event at one station and a three-hundred-year event twenty kilometers away. One discharge can be a hundred-year event upstream and a fifty-year event downstream, because a tributary did not respond the same way. Speaking of "the hundred-year storm" without saying where and on what is therefore not loose language, it is a sentence to which no truth value can be assigned.
The third is the most costly in practice: a HAZARD return period is not a LOSS return period. A hundred-year discharge occurring over a lightly built area produces a ten-year loss; a thirty-year discharge occurring where a logistics platform has since been built produces a loss far rarer than thirty years. The two scales merge in conversation and separate in the accounts. It is loss that concerns an insurer, and it is hazard that public sources document.
The fourth concerns how these numbers are obtained, and it governs the caution owed to them. One rarely holds more than fifty to a hundred years of homogeneous measurement at a point. A return period of two hundred or five hundred years is therefore never observed: it is extrapolated, by fitting a statistical law to the tail of the series. The choice of that law is an expert's choice, and two equally defensible choices can put thirty percent between the levels assigned to the two-hundred-year event. Uncertainty grows with rarity, exactly where reinsurance treaties go to find their attachment points.
It must be added that the observation series is not always homogeneous, and that a non-homogeneous series produces a wrong number with the same confidence as a right one. A catchment where a dam has been built, thirty percent of the surface sealed or a channel re-profiled no longer produces the same discharges for the same rainfall. Fitting on fifty years of measurement whose first twenty describe a catchment that no longer exists is not a calculation error, it is an error of material, and it does not show in the result.
The practical conclusion is a working vocabulary, not a warning. One does not say "hundred-year flood" but "discharge with a hundred-year return period at such a station". One does not compare two return periods without knowing whether they bear on hazard or on loss. And when a client states that he has seen two hundred-year events in ten years and that the model is therefore wrong, the three possible answers are now available: the two events were not hundred-year events at the same point, neither was a hundred-year event in loss terms, or the series the number was fitted on no longer describes today's catchment. All three can be checked.
A retail group operates a logistics platform built in 2016 on a riverside business park. It is hit by a flood in January 2023, then again in November 2025. The chief executive refuses the rate increase proposed at renewal, with an argument he believes decisive: both floods were reported in the local press as hundred-year events, two hundred-year events in three years are impossible, therefore the modeling the insurer relies on is wrong and the rate is indefensible. The technical file shows that the reference gauging station is eleven kilometers upstream, that two business parks have been built in the catchment since 2009, and that the platform was built on former farmland raised by 60 centimeters of fill. How should this be answered?
The argument is wrong, but none of the three possible answers consists in telling the client he misread his newspaper, and one of them partly proves him right. The first is that the hundred-year label, as it circulated, bears on a discharge measured at a station eleven kilometers away, which establishes nothing about the return period at the platform itself: two hundred-year events at the station are not two hundred-year events here, and nothing in the file says what they were worth at this spot. The second is that the return period of a discharge is not that of a loss: on a site raised by 60 centimeters of fill, the threshold at which water enters is not that of the natural ground, so the site's loss distribution is shifted relative to the discharge distribution, in either direction depending on level. The third deserves to be examined for its own sake: the statistical fit that produced the number rests on an observation series, and two business parks built since 2009 mean that the catchment of the last twenty years no longer produces the same discharges for the same rainfall as that of the previous fifty. If the series has not been corrected for that non-stationarity, the return period shown is too long, that is, the event is more frequent than stated. The conclusion is therefore the opposite of the chief executive's: the weakness he points to, if it exists, works against him, and it is settled neither by a rate nor by an argument, but by asking on which series and at which station the number was fitted.
- 01A return period is an annual probability disguised as a duration: one percent a year, not once a century.
- 02It attaches to a point and a quantity: the same storm is a thirty-year event here and a three-hundred-year event twenty kilometers away.
- 03A hazard return period is not a loss return period, and it is loss that concerns an insurer.
- 04Beyond the length of the measured series the number is extrapolated: two equally defensible fits can differ by thirty percent.
- 05An urbanized or re-profiled catchment makes the series non-homogeneous, and a non-homogeneous series produces a wrong number with the confidence of a right one.