Can GloFAS set a flood trigger where there is no river gauge?

Yes, but only if the trigger is written in the model's own terms and its basis risk is checked against local evidence first. KAI tested this on 14 river gauges in Spain's Guadalquivir basin. With GloFAS 4.0, a trigger read off a gauge study in m³/s would have paid in 6 of 60 flood years. With GloFAS 5.0, which removes most of that bias, it paid in 26 of 60. Written as the model's own return period it did better with both versions, yet about half of all payouts still fell in years without a flood.

Why the question matters on ungauged rivers

A parametric flood policy pays when a river index crosses an agreed level. Most places that need cover have no gauge, or a record too short to fit a flood curve. The Global Flood Awareness System (GloFAS) fills that gap with a daily river-flow reanalysis for every river cell on Earth, driven by past weather. The ERA5-driven version was found skilful in 86 % of the catchments it was evaluated in [1]. Version 5 runs at 0.05°, roughly 5 km, from 1980 to 2025 [2].

So a trigger for any reach can be read from GloFAS in minutes. The hard question is whether it pays in the years the river really floods. "Skilful on average" does not answer that for one reach, and an insurer pricing a contract, or a buyer relying on it, needs that answer per site.

The test: score GloFAS as if no gauge existed

The Guadalquivir is a hard case for a global model: semi-arid, heavily dammed and irrigated. That makes it a useful test, because many ungauged rivers that need cover share those traits. The method, published openly with its data and code:

• Every gauge in the basin with a catchment of at least 500 km² and at least 25 years of record in water years 1985–2022: 14 gauges, from 583 to 16,166 km², daily flows from CEDEX for Spain's Ministry for the Ecological Transition.
• For each gauge, the GloFAS river cell within 0.1° whose mean flow is closest to the gauge's. That choice uses the gauge, so it is the model's best case; at a truly ungauged site the cell is chosen by catchment area.
• The largest daily flow of each water year, at the gauge and in the model, fitted with GEV and log-Pearson III curves by L-moments, with 90 % bootstrap bands.
• A flood year is a year the gauge passed its own 1-in-T level. A trigger set from the model "pays" when the model passes its level. Every year is scored as a hit, a miss or a false payout.

Both model versions were scored on the same 286 gauge-years (1998–2022), so any difference is the model, not the years. At 1-in-5 there were 60 flood years in that sample.

What GloFAS 4.0 got wrong: the size of the flood

With GloFAS 4.0, the model's 10-year flood was 22 % of the gauged one at the median gauge, and within a factor of two at only 4 of the 14. At Mengibar, the largest catchment, the gauge's 10-year flood was about 355 m³/s and the model's about 70 m³/s. A search up to 25 km round that gauge found no bigger river cell, so the problem was the model, not the map.

A trigger copied from a gauge study in m³/s therefore almost never fired: 6 of 60 flood years at 1-in-5, Peirce skill 0.06, no better than chance. The same trigger written as the model's own 1-in-5 flood caught 30 of 60 (Peirce 0.38). The model was wrong about size but still carried useful information about which years were big. Writing the level in its own terms kept the information and dropped the bias.

What GloFAS 5.0 changed

GloFAS 5.0 was downloaded for the whole basin and scored on the same gauge-years. Its 10-year flood came out at 1.05 of the gauged one at the median gauge, within a factor of two at 10 of 14 gauges, and it ranked the big years better (median rank correlation 0.55 against 0.42). Version 5 represents 1,486 reservoirs, held in their early-2026 configuration across the whole record [2], and it was recalibrated at 5,379 gauging stations with revised lake and reservoir inputs [3]. The Guadalquivir test does not isolate which of these changes removed most of the bias. A reservoir in the model usually lowers flood peaks rather than raising them, so the larger reservoir set is unlikely on its own to explain the larger floods; what the static set does mean is that early years are routed through today's dams, which suits a trigger written for today's river [2].

With the bias largely gone, the m³/s trigger rose from 6 to 26 of 60 flood years. The model's own return period still caught more, 33 of 60 (Peirce 0.40 against 0.30), but the gap was no longer statistically clear on these years: of the flood years caught by one method and not the other, 9 went to the return period and 2 to the fixed flow (McNemar p = 0.065). Over every year GloFAS 5.0 and the gauges both cover, 1985–2022 with 89 flood years, the return-period trigger paid 53 flood years with 44 false payouts, and the m³/s trigger 41 with 31.

The model-version lesson

The same rivers, the same years and the same rules gave a fixed-flow trigger that paid in 6 flood years under one version of the model and 26 under the next. For anyone holding a contract written in m³/s from a global model, that is a warning: an upgrade can change how often the policy pays, overnight, in either direction.

A trigger written as the model's own return period, percentile or share of a reference flood is re-derived from the new record when the model changes, so its payout frequency stays close to what was agreed. That is the practical reason to prefer it, beyond the extra flood years it caught. Either way, every trigger should be re-tested when the model it reads is upgraded.

Basis risk did not go away

A better model narrowed the gap between methods; it did not cut basis risk. With GloFAS 5.0, the return-period trigger made 34 false payouts against 33 hits, and the m³/s trigger 30 against 26. About half of all payouts fell in years without a flood, as with version 4.0. In verification terms, the false-alarm ratio stayed near 0.5 even though the false-alarm rate, measured over the many dry years, was low [4]. In a regulated basin that is expected: reservoirs cut and delay flood peaks in some years and not in others, and the model does not know how they were operated.

Basis risk matters to a buyer as much as price. Theory shows that it limits how much index cover it is rational to hold [5]. A trigger that pays in the wrong years half the time is a hard product to sell and a hard one to keep.

Can you tell in advance where it will work?

Partly. With GloFAS 4.0, the share of flood years a gauge's trigger caught followed the rank correlation between the gauge's annual floods and the model's (Spearman 0.80). Gauges where the model ranked the big years well, such as Mengibar, Castro del Río, Central de Cala and Aznalcázar Puente, caught half or more of their flood years. With GloFAS 5.0 the link was weaker on these years (0.41, not significant). A short local record, even a few years of flood marks, satellite flood extents or claims, still gives that correlation and is the best early screen available, but it is a screen, not a guarantee.

A second index helps

The same study, on GloFAS 4.0, also tried rainfall from NASA POWER. Storm rainfall (the wettest three days) caught 23 of 60 flood years, the October–April rainfall total 31, and river flow 30. Requiring both river flow and season rainfall to pass cut false payouts from 47 % to 29 % of payouts, at the cost of 6 flood years. In this basin floods follow wet winters, so the season total works; elsewhere the right second index may differ.

Limits of the evidence

One basin, regulated and semi-arid, so the size of the bias will differ elsewhere. Daily mean flows, so instantaneous peaks are higher. Between 17 and 24 years per gauge in the matched sample, so levels beyond 1-in-25 are extrapolations. Flood years are defined from gauge flows, not losses; a contract also has to consider where and how losses happen.

What KAI offers

KAI-Hydro runs this study for your sites, gauged or not: each site moved onto the river by catchment size, the GloFAS v5 reanalysis from 1980 to today, return levels with 90 % bands, the trigger written four ways (fixed flow, percentile, return period, share of a reference flood) with attachment and exhaustion, and a back-test of hits, misses and false payouts. KAI recommends calibrating each trigger on a gauge, satellite flood maps or local records before it enters a contract, because a global model alone carries too much basis risk. Each trigger is re-tested when GloFAS is upgraded, as here from 4.0 to 5.0. The method itself is set out in the guide to parametric flood trigger design, and the commercial study on the parametric flood triggers page.