KAI-Hydro · parametric triggers · now on GloFAS 5.0
Can a global flood model set a flood trigger where there is no gauge?
We took every large river gauge in Spain's Guadalquivir basin, set flood triggers from the GloFAS global flood model as if no gauge existed, and scored them against what the gauges measured, year by year. Now on GloFAS 5.0, side by side with 4.0. Every chart below is interactive.
Real data: daily river flows from 14 CEDEX gauges (Anuario de Aforos 2021-2022, Spain's Ministry for the Ecological Transition) and the GloFAS v4.0 reanalysis from the Copernicus Emergency Management Service. Only the fitted flood levels and triggers are modelled.
River gauges tested14500 to 16,200 km² · water years 1985–2022
Trigger as the model's 1-in-533 of 60flood years paid · 34 false payouts · 4.0: 30 · 27
Trigger in m³/s26 of 60flood years paid · 30 false payouts · 4.0: 6 · 10
Same 286 gauge-years (1998–2022) for both model versions, so the difference is the model. GloFAS 5.0 over all its years (1985–2022): 53 of 89 flood years paid as a return period, 41 in m³/s.
The same 14 gauges and the same 286 gauge-years (1998 to 2022), with GloFAS 4.0 and the new GloFAS 5.0, so the difference is the model, not the years. A 1-in-5 trigger, pooled over all gauges.
GloFAS 4.0
GloFAS 5.0
Model's 10-year flood ÷ gauge's, median
0.22
1.05
Gauges with the 10-year flood within ×2
4 of 14
10 of 14
Do the big years line up? (rank correlation, median)
0.42
0.55
Written as the model's own return period: flood years paid · false payouts
30 of 60 · 27
33 of 60 · 34
Written in m³/s from a gauge study: flood years paid · false payouts
6 of 60 · 10
26 of 60 · 30
The low bias is gone. A trigger read off a gauge study in m³/s now works, where with 4.0 it almost never paid.
The model's own return period is still the safer way to write it (Peirce skill 0.40 against 0.30), though on these years the gap is no longer statistically clear.
Basis risk has not gone: about half of all payouts still fall in years without a flood. KAI screens each site against local records before a trigger enters a contract.
Over every year both cover (1985 to 2022, 89 flood years) the return-period trigger paid 53 flood years with 44 false payouts, and the m³/s trigger 41 with 31. Data and script: kai-case-studies (paper/compare_v4_v5.py). Contains modified Copernicus Emergency Management Service information 2026 (GloFAS v5.0).
The map · switch between map, light and satellite
Pick a gauge, or pin your own forecast point
Each dot is a river gauge, coloured by how well the global model ranks its big flood years; the orange square is the 5 km GloFAS river cell matched to it. Tap a gauge to load its flood curves and trigger history below, or tap anywhere on a river to pin a point of your own and ask for a trigger study there.
Pick a forecast point
Tap a gauge to load its analysis below. Tap anywhere else to pin your own point: KAI will set the trigger there.
Gauge: big years match (0.6+)
0.3 to 0.6
below 0.3
GloFAS river cell (5 km)
Sub-catchments draining to the chosen gauge
Your forecast point
Basemaps: OpenFreeMap (OpenStreetMap) and Esri World Imagery. Rivers, lakes and boundaries: OpenStreetMap. Sub-catchments: HydroBASINS level 8 (WWF, Lehner and Grill 2013).
1 · How big is a 10-year flood?
GloFAS 5.0 sees these floods at about their real size
GloFAS 4.0 sees these floods at about a fifth of their size
Each dot is a gauge: the 10-year flood measured at the river against the same flood in GloFAS at the best-matching river cell. With GloFAS 4.0 most dots sit far below the 1 : 1 line; with 5.0 they gather round it. A search up to 25 km round the largest gauge found no bigger river cell, so this is the model, not the map: the Guadalquivir is dammed and irrigated, and much of its flood water is managed in ways a global model cannot know.
2 · One gauge at a time
Flood-frequency curves, with their uncertainty
Each year's largest daily flow is plotted at its return period; the line is the fitted curve (GEV or Log-Pearson III, by L-moments, whichever fits better) and the shading its 90 % bootstrap band. Where the two curves run parallel, a trigger written as a return period still works; where they cross, it does not.
3 · Year by year
Would the policy have paid in the flood years?
Trigger level
Trigger written as
–
flood years at the gauge
–
caught: the policy paid
–
missed: flood, no payout
–
false payouts: paid, no flood
Loading the reference case…
4 · All 14 gauges
Return period or m³/s: the share of flood years paid
The share of flood years each way of writing the trigger would have caught, pooled over all gauges, for the version chosen above. With GloFAS 4.0 a fixed flow almost never fires, because its floods are too small; with 5.0 it works, but the model's own return period still catches more.
5 · Basis risk
Does the model get the big years right?
The rank correlation of each gauge's annual floods with the model's: 1 means the biggest years line up exactly. Below about 0.6, a policy will pay in the wrong years too often to sell without local calibration.
Work package · KAI-Hydro Professional
This study, for your sites
Everything in this reference case, run on GloFAS 5.0 for up to five sites of your choice, gauged or not, and kept current when the model is upgraded.
Site and riverEach site moved onto the river by catchment size; small streams flagged for a local study.
Flood history40+ years of daily flow from GloFAS 5.0 at each site, plus any gauge or records you hold.
Return levels1-in-2 to 1-in-200-year floods with 90 % bands; beyond the record flagged.
Trigger designWritten four ways, with attachment and exhaustion levels: a payout band, not a single line.
Back-testEvery year: flood years paid, missed, and payouts without a flood.
Basis-risk scoreHow well the model ranks your real floods, checked against your local evidence.
Trigger sheetOne page per site, ready for the term sheet, with a method note and the data files.
Kept currentEach trigger re-tested when GloFAS is upgraded, as here from 4.0 to 5.0.
Professional planUp to 5 sites · design floods for engineers from the same study · programmes and many basins on the Anticipatory Action plan
Real: daily mean flows at 14 gauges with catchments of 500 km² or more and at least 25 years of record, water years 1985 to 2022 (CEDEX).
Real: daily river discharge from the GloFAS reanalysis (LISFLOOD, consolidated), at the 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.
Modelled: the flood-frequency fits, their bands, and the 1-in-T triggers.
Not redistributed: the CEDEX files are fetched from CEDEX at run time and checked by SHA-256; this page shows only annual maxima and fitted levels.
Real: GloFAS 5.0 (consolidated, Copernicus Early Warning Data Store, 1984 to 2022) and GloFAS 4.0 (Open-Meteo), each at its own best-matching cell; switch between them at the top.
How it works
Select every gauge with a large enough catchment and a long enough record, and find its GloFAS river cell.
Take each year's largest daily flow (October to September) in the years both sources cover.
Fit GEV and Log-Pearson III by L-moments, keep the better fit, and bootstrap a 90 % band for each return level.
Call a year a flood year when the gauge passed its own 1-in-T level; then count when a trigger set from the model would have paid.
Hosking and Wallis (1997), Regional Frequency Analysis, Cambridge · Harrigan et al. (2020), GloFAS-ERA5 reanalysis, Earth Syst. Sci. Data 12, 2043–2060.
Fuente: CEDEX, Anuario de Aforos 2021-2022 (Ministerio para la Transicion Ecologica y el Reto Demografico). Contains modified Copernicus Emergency Management Service information 2026 (GloFAS v4.0) via Open-Meteo (CC BY 4.0).
KAI-Hydro
Set a trigger for your site
We pull the flood history for your location, design the trigger four ways and show you its hits, misses and false payouts before you write the policy.