On the day a team runs two market models on the same portfolio, it discovers something written nowhere in the brochures: the results differ, and they differ a great deal. Gaps of twenty to forty percent on a two-hundred-year loss are not a pathology, they are the ordinary state of the subject. Knowing what to do with that gap is a skill in itself, and it starts by refusing the two reflexes that present themselves first, picking the lower one because it suits, or averaging because that feels neutral.
The first thing to do is to locate the gap rather than note it. A global gap cannot be discussed; a decomposed gap discusses very well. One looks at whether it is present everywhere or concentrated on a region, on a construction class, on a band of return periods. A gap concentrated in the tail and absent elsewhere points to extrapolation assumptions; a gap uniform across all return periods points rather to vulnerability functions or to the treatment of exposure. Decomposition turns a quarrel about numbers into a technical question.
The second is to check that the same thing is being compared, which fails more often than one would think. Did both models receive the same exposure, with the same geocoding and the same values. Do they return the same thing, direct damage alone or damage and business interruption. Do they apply the same event definition. Are these two OEPs, or an OEP and an AEP. An appreciable share of spectacular gaps disappears at this stage, and it disappears in half a day's work rather than in three meetings.
The third is to find where the difference of view comes from, when it survives the first two. Models do not diverge at random: they diverge because they made different choices on documented points. A shorter or longer observation period. A different extrapolation law. A secondary phenomenon taken into account or not, runoff, site effect, failure of a protective structure. A different treatment of buildings whose class is not recorded, which are often half of a householders portfolio. Those choices are written down, and asking for them is not a mark of distrust.
Then comes the awkward question, that of the figure to retain. The professional answer consists neither in picking nor in silently averaging, but in holding an own view and writing it down. An own view may be one model retained as the reference with a justification, a weighting between two models with the reason for the weighting, or an adjustment applied to one output on a precise segment. What makes it defensible is not its value, it is that it was stated before the results, applied without exception, and revisable on a criterion rather than on mood.
One must name what this discipline protects against, because it is a real and rarely conscious temptation. When the view is chosen after the figures have been seen, it is always chosen in the direction of comfort: the low model when the market is tight and business must be written, the high model when a protection is being negotiated. A house arbitrating that way twice a year does not hold a view of risk, it holds an instrument of justification, and that instrument turns against it the day a loss requires explaining why the low curve was being used.
The last point concerns version change, which is the same problem in another form. When a vendor delivers a new version, modeled loss moves without any risk having changed. The rule of conduct is to run the old and the new version on the same portfolio before adopting, to measure the gap, to document it, and to make it known to those who take decisions on those figures. An unreported gap resurfaces twelve months later as a drift attributed to climate, and nobody can tell them apart any more.
A company runs two models on its commercial portfolio exposed to flood, 4.2 billion euros of insured values. Model A returns 118 million at the two-hundred-year return period, model B returns 167 million. Three facts emerge on examination: model B includes urban runoff, model A does not handle it; fourteen percent of the portfolio's sites have no construction class recorded, and the two models apply different default values to them; finally, the extraction submitted to A was dated January and the one submitted to B June, a period during which a large account of 380 million in values came in. The finance department asks which figure to retain for the reinsurance plan. What should be answered?
Neither figure is yet comparable to the other, and the first task is to make them comparable before speaking of a choice. The third fact settles part of the gap with no technical discussion at all: two different extractions are not two views of the same portfolio, and the arrival of a 380 million account between January and June is enough to explain a substantial fraction of the 49 million gap. Rerunning A on the June extraction is the prior step, and it costs a day. The second fact is handled the same way: fourteen percent of sites with no construction class means both models filled in themselves what the portfolio did not say, and the resulting gap measures not a difference of view on risk but a data shortfall only the company can close. Documenting those sites is an internal task, and it will bring the two figures closer without any vendor having changed anything. That leaves the first fact, and it is the only genuine difference of view: one model handles urban runoff, the other does not. That one is not resolved by a calculation, because it raises a substantive question, whether the portfolio contains sites whose damage would come from water rushing down rather than water overflowing. The answer is established on the portfolio itself, by examining sites at the foot of slopes and in sealed areas. What has to be returned to the finance department is therefore not one of the two numbers but a written view: the model retained as the reference, the reason for that choice, the adjustment applied to the segment concerned if there is one, and the revision rule. And that view must be settled before the renewal, failing which it will be settled according to what the renewal turned out to cost.
- 01A twenty to forty percent gap between two models on a two-hundred-year loss is the ordinary state of the subject, not an anomaly.
- 02A gap is located before it is discussed: concentrated in the tail it points to extrapolation, uniform it points to vulnerability or exposure.
- 03An appreciable share of spectacular gaps comes from not comparing the same thing: same extraction, same perimeter, same event definition.
- 04An own view is stated before the results and applied without exception; chosen afterwards, it always runs in the direction of the moment's comfort.
- 05A version change is measured on the same portfolio before adoption, failing which it resurfaces twelve months later as drift attributed to climate.