How to design a parametric flood trigger, and how to measure its basis risk

A parametric flood trigger is designed in five steps: choose the index and the point on the river, build a long flood history at that point, write the trigger level in the index's own terms, set where payouts start and stop, and back-test every year to count hits, misses and false payouts. The last step is the one that measures basis risk, and it is the one most often skipped. This guide walks through each step, with the numbers from an open test on 14 river gauges in Spain.

What a parametric flood insurance index has to do

Parametric cover pays an agreed sum when a measured index crosses an agreed level, with no loss assessment. For river floods the natural index is the flow at a point on the river. The policy is only as good as that index: if it crosses the trigger in years when nobody was flooded, or stays below it in a year when they were, the buyer carries the difference. That difference is basis risk. It runs both ways: a payment larger than the loss, or cover that falls short of the loss [1].

Basis risk is not a detail to tidy up after pricing. Economic theory shows that it changes how much index cover a rational buyer should want at all, and it offers one explanation for why take-up of index products is often low [2]. So the design work is mostly about knowing, before anyone signs, how often the index and the flood disagree.

Step 1: choose the index and the forecast point

Where a river gauge with a long record exists, use it. Most places at risk have none, or a record that is short, broken or not public. The usual fallback is a global flood reanalysis: a hydrological model run over the whole land surface, driven by past weather, giving daily river flow at every river cell. The GloFAS reanalysis is the most widely used. Its ERA5-driven version was found skilful in most of the catchments it was checked against [3], and version 5 runs at 0.05° (about 5 km) from 1980 to 2025 and includes 1,486 reservoirs [4].

The point that matters is not the town but the river cell that carries its flood. A site placed on the map is moved onto the river cell whose upstream catchment matches the real one in size. Small streams below the model's resolution should be flagged rather than forced onto a large neighbouring river.

Step 2: build the flood history and its return levels

Take the largest daily flow in each water year, then fit a flood-frequency curve. Generalised extreme value (GEV) and log-Pearson III distributions, fitted by L-moments, are standard choices; keep whichever fits better. Bootstrap the fit to put an uncertainty band on each return level. A 1-in-5-year flood is well constrained by 40 years of record; a 1-in-100-year flood is an extrapolation, and its band should say so.

Step 3: write the trigger level, four ways

There are four common ways to write the level at which the index triggers:

• Fixed value. Pays above a set flow, for example 450 m³/s. Easy to read and settle.
• Percentile. Pays above the index's own 95th or 99th percentile flow.
• Return period. Pays above the index's own 1-in-T-year flood, with T agreed in the policy. This is the flood return period trigger.
• Share of a reference flood. Pays above a share of a named past flood in the index, for example 80 % of the 2010 peak.

The last three are written in the index's own terms. That matters because a global model can be biased in size and still rank the years correctly. In KAI's test on 14 gauges in Spain's Guadalquivir basin, GloFAS 4.0 saw floods at about a fifth of their measured size. A 1-in-5 trigger written in m³/s from a gauge study paid in only 6 of 60 flood years; the same trigger written as the model's own 1-in-5 level paid in 30 of 60. The newer GloFAS 5.0 removed most of the bias, and the m³/s trigger rose to 26 of 60, but the return-period version still caught more (33 of 60). The full comparison is in the Guadalquivir flood trigger case.

A second lesson from the same test: set each site's level from its own record. One absolute level shared by all 14 gauges paid in anywhere from 0 % to 78 % of years from one gauge to the next. Once the level is set per site, it can still be written into the contract in m³/s for settlement.

Step 4: attachment and exhaustion

A single trigger line makes a cliff: one cubic metre per second either side of it decides between no payout and a full one. Most designs instead use a payout band. The attachment level is where payout starts; the exhaustion level is where it reaches the full sum; in between, payout rises with the index. A band softens near-misses, which is one of the standard ways of reducing basis risk [1].

Where to put the two levels is a pricing and appetite question for the insurer and the buyer. The design study's job is to tell both sides how often each level would have been reached, and how often in the right years.

Step 5: back-test and measure basis risk

Back-testing means replaying the trigger over every past year with an independent record of what happened: a gauge, satellite flood maps, flood marks, or insurance claims. Each year falls into one cell of a contingency table [5]:

• Hit: a flood year, and the trigger paid.
• Miss: a flood year, and it did not pay. This is the buyer's basis risk.
• False payout: no flood, but it paid. This is the insurer's basis risk.
• Correct no-payout: no flood, no payment.

From the table come the measures worth putting in a term sheet [5]:

• Hit rate (probability of detection): hits ÷ flood years. Here 33 ÷ 60 = 0.55.
• False-alarm rate: false payouts ÷ years without a flood. Here 34 ÷ 226 = 0.15.
• Share of payouts that were false (false-alarm ratio): false payouts ÷ all payouts. Here 34 ÷ 67, about half.
• Peirce skill (hit rate minus false-alarm rate): 0 means no better than chance, 1 is perfect. Here 0.40.

Report all four. A trigger can look good on hit rate alone simply by paying often, and a low false-alarm rate can hide the fact that, because floods are rare, a large share of payouts still fall in dry years. In the Guadalquivir test, a trigger with clear skill still paid without a flood in about half of its payouts. That is a statistically real trigger and still too much basis risk to sell on the global model alone.

A screen you can run before the back-test

Basis risk shows up most simply as rank correlation: do the model's biggest years match the river's biggest years? In the GloFAS 4.0 test the share of flood years caught at each gauge followed that correlation closely. Gauges where it was 0.6 or above caught half or more of their flood years; gauges where it was near zero caught a fifth to two-fifths. With GloFAS 5.0 the link was weaker on these years, so treat the correlation as a screen, not a verdict. A few years of local evidence are enough to compute it.

Flood trigger calibration and re-testing

Where a gauge, satellite flood maps or local flood records exist, calibrate the trigger on them before it enters a contract. A global model alone carries too much basis risk to sell. Then re-test it whenever the model is upgraded: a level written in the model's own terms is re-derived from the new record, while a fixed flow can suddenly pay far more or far less often, as the move from GloFAS 4.0 to 5.0 showed.

The same method applies to flood index insurance in Africa and other regions where gauges are sparse. Gauge density does not change the steps; it changes how much local evidence (flood marks, satellite flood extents, claims records) has to be gathered for the screen and the calibration.

What KAI offers

KAI runs trigger design and back-testing as a KAI-Hydro study. It moves each site onto the river by catchment size, pulls the GloFAS v5 reanalysis from 1980 to today, fits GEV and log-Pearson III return levels with 90 % bands, writes the trigger the four ways above with attachment and exhaustion levels, and back-tests each year by year: flood years paid, missed, and payouts without a flood. It scores basis risk against your local evidence and recommends calibration on gauges, satellite maps or local records, because a global model alone carries too much basis risk. The output is a trigger sheet per site with a method note and the data files. Details are on the parametric flood triggers page and the KAI-Hydro product page.