Where RA-FEWS is strong, and where it needs support.

RA-FEWS turns global ensemble weather forecasts into local, probabilistic river-flood warnings. Each forecast run, a 40 member ensemble, is fed through a catchment hydrology model, rainfall to runoff to routed river discharge, and expressed as the probability that flow will exceed defined flood levels, days ahead, per reach. What follows is where that chain is strong and where it needs support. The boundaries below are not defects we are hiding. They define the regimes where RA-FEWS is the right tool and the regimes where it works best alongside a complementary one.

What RA-FEWS is built for

RA-FEWS is strongest on large, rainfall-driven river floods forced by organised, large-scale weather systems: monsoon surges, El Nino seasonal rains, tropical and Mediterranean storm systems. These are well resolved and predictable by global ensembles days in advance, and they produce basin-scale flood waves that catchment hydrology routes faithfully.

That is not a narrow niche. It is where most of the world's high-fatality riverine flooding occurs. In reconstructions of real events the system delivered clear, escalating, high-confidence warnings several days ahead of catastrophic river floods driven by exactly these mechanisms.

How to read the system's edges

In the order we would want a hydrologist to read them.

1. 01

   Isolated convective flash floods

   The hardest case for this approach and the most important limit to understand. Short-lived thunderstorm clusters that dump extreme rain on a few square kilometres in hours, often in arid regions, are effectively invisible to a global medium-range ensemble. Simulating deep convection rather than statistically approximating it needs grid spacing near 4 km; global ensembles run at tens of kilometres, so the cell is smaller than one grid box. The exact position and timing is chaotic beyond a few hours, which is a limit of physics rather than of any particular model. Convection over dry terrain is the hardest variant: high-instability air produces pop-up storms that flood one district while the next stays dry, and global models tend to under-forecast them. This is a property of the single-cell trigger, not of the region: where the rain over a dry catchment comes from an organised system, as at Derna, the forecast holds days ahead. Capturing these operationally needs a different data tier, a convection-permitting regional model and radar or satellite nowcasting, which delivers warnings hours rather than days ahead. That is a complementary system, not a setting we can enable. One partial exception is real and useful: a coarse ensemble can often forecast the ingredients for convection, high instability, deep moisture, a triggering disturbance, without placing the cell. That supports a probabilistic elevated flash-flood risk flag for a region and a day, which is actionable even when the routed hydrograph for a specific basin reads zero.

2. 02

   Catchment and event scale mismatch

   RA-FEWS forecasts discharge at a defined catchment outlet, driven by the average rainfall over that catchment. When the disaster concentrates in small tributaries far below the modelled outlet, the basin-average signal can be modest even when the forecast rainfall is accurate, because an intense local downpour averaged over a large basin is a moderate mean. In one reconstruction the forecast rainfall matched observations closely and the modelled flood at the outlet still stayed low, because the deadly flooding happened in small upstream torrents whose scale sits below both the catchment and the model grid. The catchment has to be delineated to match the scale of the risk being warned for. That is a setup decision we make with you: we size the outlet to the risk and, on large basins, model sub-catchment outlets as well, so a localised extreme is reported where it happens rather than averaged into a moderate basin mean.

3. 03

   Small, fast-responding catchments

   Separately from whether the weather is forecastable, the hydrological model has a lower bound on catchment size and response speed below which it cannot reproduce the flood peak, even with perfect rainfall. A small, steep catchment concentrates its runoff in a very short time and its lag can be an hour or less. Reproducing that sharp, brief peak needs rainfall resolved at sub-hourly intensity, because peak runoff is governed by the most intense burst over a window roughly equal to the time of concentration, and a model timestep short relative to that response. When the response time approaches the forcing and computational timestep, the peak is flattened: the model spreads a short, violent flood over too long a window and under-estimates the crest. For very small, flashy headwater catchments, fast mountain torrents and urban wadis, the modelled peak should be read as a lower bound. This limit is a property of the hydrology and is independent of forecast skill.

4. 04

   Snowmelt and rain-on-snow

   RA-FEWS models rainfall-driven runoff. It has no snowpack or snowmelt component today, so floods dominated by snowmelt or by rapid rain-on-snow release are not represented in the current physics. A snow module is on our development roadmap. In the meantime, rain-driven flooding in the same basin is modelled normally, and we will tell you if snowmelt is the dominant mechanism in your catchment.

5. 05

   Regulated rivers and reservoirs

   The hydrology assumes natural runoff. Dams, reservoirs, barrages and their operating rules, storage, controlled release and gate operations, are not modelled. On a heavily regulated river the absolute discharge and the timing will be wrong under normal operations. During an extreme event, once reservoirs are full and passing their inflow, natural-runoff timing becomes a reasonable approximation, but the routed discharge remains indicative rather than a regulated-flow simulation.

6. 06

   Satellite rainfall under-catch

   Where satellite precipitation such as GPM IMERG is used to check forecasts or to estimate antecedent soil moisture, note that satellite retrievals systematically under-catch orographic and convective rainfall extremes. Observed rainfall figures derived this way should be read as a lower bound, and verification against them can understate a forecast's accuracy for intense events. This affects evaluation and initial conditions, not the forward forecast.

7. 07

   Thresholds require per-catchment calibration

   Flood severity levels, the 2, 10, 20 and 50 year floods, are only as good as their calibration. Where a catchment has not yet been fitted to gauged flood-frequency records, its thresholds are provisional: area-scaled, illustrative levels chosen for readability, not calibrated design floods. In that state the probabilities should be read as rising flood risk, not as calibrated exceedance of a certified design flood. Calibration to local gauge or reference data removes this caveat for that catchment. This is the caveat that applies most often in practice, and it is the one behind the illustrative label on our published case-study figures.

8. 08

   The probability is the product

   Two things a user has to internalise. First, read the ensemble as a probability rather than a single line: with 40 parallel forecasts the informative quantity is how many cross a flood level, the exceedance probability, not whether the middle member does. Days ahead the members disagree and the median can sit below threshold while a meaningful minority signal a flood; judged on the median alone that looks like a miss. Operational services such as GloFAS and EFAS verify on exceedance probability for exactly this reason. Second, the value is the early, high-confidence warning and not the exact crest hour. Crest timing is less certain than crest risk, reconstructed crests can fall a day or two from the observed peak, and the median can soften right as an event arrives. On large, slow-responding basins the payoff is a long, confident warning horizon rather than a precise arrival time.

9. 09

   Lead time and the predictability horizon

   Forecast skill degrades with lead time. RA-FEWS is designed for multi-day warning on synoptic-scale systems, not for sub-daily nowcasting. Its 120 hour horizon is a strength for slow-building river floods and a poor fit for events whose entire life cycle is shorter than the model's useful range. Beyond that horizon we have nothing to say, and we would rather say nothing than extrapolate.

A note on the published reconstructions