The forecast already exists. The warning doesn't.
For most of the world's catchments the rain is predicted days ahead and nobody is told. That last step is the whole problem, and it is the one we build.
Weather forecasting is global. Flood warning is not.
Numerical weather prediction is one of the quiet successes of the last fifty years. The DWD's ICON ensemble, and the models like it, resolve the atmosphere over every river basin on earth several times a day, free at the point of use. The rain that destroyed Derna in September 2023 was in those forecasts for days.
Flood warning is a different thing, and it is not global at all. Turning a rainfall forecast into "this catchment, this threshold, this hour, tell these people" needs a hydrological model, calibrated thresholds and something to run the whole chain on a schedule. That capability follows national wealth, and it stops where the money does.
40%
of countries still report no multi-hazard early warning system, as of late 2025.WMO, Global Status of MHEWS 2025
a few %
of the world's watersheds have a stream gauge, and gauge density tracks national GDP.Nearing et al., Nature 2024
1.8bn
people are exposed to flood risk. Almost 90% of them live in low- and middle-income countries.Nearing et al., Nature 2024
No gauge, no record, no warning.
A stream gauge is an instrument in the river that records how much water goes past it. The conventional way to build a flood warning starts there, and it cannot start anywhere else. Decades of those measurements are what let you fit a rating curve, work out what a hundred-year flood looks like on that particular river, and calibrate a model an official will act on.
Take the gauge away and every step after it fails in order. No record, so nothing to calibrate against. No calibration, so no defensible thresholds. No thresholds, so nothing for a forecast to be compared with, and no warning anyone can issue. The method does not get less accurate in an ungauged catchment. It stops.
No gauge
Most of the world's catchments have never had one. Nothing is measuring the river, and nothing ever has.
So no record
Calibration needs decades of measured discharge. No amount of modelling invents the years that were never recorded.
So no thresholds
Return periods and warning levels are derived from that record. Without it there is no defensible number to say the river has crossed.
So no warning
The forecast still arrives, on time, for that catchment like any other. It just has nothing to be compared against, so nobody is told.
That is the gap this page is about, and the precision matters: not catchments that are watched badly, but catchments that are not watched at all. Building the missing gauge network is decades of capital spending, so waiting for one is not a plan for the people living there now.
What the gap costs, in numbers that are not ours.
The value of closing it is one of the better-evidenced findings in the field, and none of the evidence comes from us.
43%
reduction in flood fatalities attributable to early warning, with economic losses down 35 to 50%.Nearing et al., Nature 2024
23,000
lives a year the World Bank estimates would be saved by raising developing-country warning systems to developed-country standards.World Bank, via Nearing et al.
6x
lower disaster mortality in countries with comprehensive early warning coverage than in those with limited capability.WMO, Global Status of MHEWS 2025
The rate of flood-related disasters has more than doubled since 2000, and the UN's Early Warnings for All initiative targets universal coverage by 2027. The deadline is closer than the gap is small.
How RA-FEWS closes it.
We start from the forecast that already exists and build only the missing part. The DWD ICON-EPS ensemble gives forty rainfall trajectories for any catchment on earth, every cycle. We route them through a catchment model, compare every member against the thresholds you configure rather than ones a national service chose, and issue a per-zone warning with the probability attached.
Two consequences matter. A catchment with no gauge history is not disqualified, because thresholds are a configuration input: a local engineer's judgement, a design standard or a reconstructed estimate can stand in until a record exists. And because forty members disagree, what you get is a probability rather than a single line, which is the honest shape of the answer when observations are thin.
Every number is also traceable. Each metric on screen opens to its own methodology page with the formula and the references behind it. Where data is sparse, being able to show your working is what makes a warning defensible.
Derna is the whole argument in one event.
The signal was there. The warning is the product.
What we are not claiming.
Stating the limits is the point, particularly when the whole argument is about places where evidence is scarce.
- Our case studies are hindcasts. They replay real events through the system after the fact. They are not warnings we issued at the time, and they benefit from knowing which catchment to look at.
- A forecast is only as good as the thresholds it is compared against. Where those are estimated rather than measured, the warning inherits that uncertainty, and the interface says so.
- An ensemble spread describes the model's own uncertainty, not the full uncertainty of the world. Rare convective events remain hard everywhere, including here.
- We forecast to a 120 hour horizon. Beyond it we have nothing to say, and we would rather say nothing than extrapolate.
Sources
- World Meteorological Organization, Global Status of Multi-Hazard Early Warning Systems 2025, published 12 November 2025. wmo.int
- Nearing, G. et al., Global prediction of extreme floods in ungauged watersheds, Nature, 2024. nature.com
- United Nations and WMO, Early Warnings for All, targeting universal early-warning coverage by 2027. un.org