Can AI-calibrated forecasts give East Africa earlier warning of extreme rain?

The evaluation spans several years, thresholds and reference systems

The authors used six-hourly physical forecasts on a 0.1-degree grid and converted them into 24-hour rainfall totals. The AI post-processing produced daily forecasts at one- to five-day lead times on the same grid. Satellite-derived IMERG version 7 rainfall was the main reference. Because satellite estimates can contain their own errors, the team also compared forecasts with 24-hour totals from six gauges around Nairobi: Jomo Kenyatta International Airport, Wilson Airport, Eastleigh, Dagoretti, Kabete and Ngong.

For the calibration analysis, 2018, 2019 and 2020 were fitting years and 2021, 2023 and 2024 were held out for testing, although only June–December 2023 and March–June 2024 were available in parts of the analysis. The underlying generative model also used 2021 as a validation year. That separation is stronger than evaluating on the same periods used to fit the system, but it is not a prospective trial of warnings issued in future seasons. The study is a retrospective forecast evaluation using historical observations.

The regional analysis fitted probabilities for daily rainfall thresholds from 20 to 60 millimetres in 10-millimetre steps, chosen to span medium to severe conditions described by local meteorologists and World Meteorological Organization guidance. Additional regional plots examined thresholds such as 15, 20, 25 and 40 millimetres a day. The researchers assessed reliability, resolution, reductions in cross-entropy loss and Relative Economic Value across different cost-loss ratios. That last measure asks whether using a forecast could be preferable to always acting or never acting under a stylised decision cost—not whether an agency actually saved money or prevented harm.[1][2]

The gains were geographically uneven—and that is the most useful finding

The calibrated AI-assisted combination produced probabilities that more closely matched observed frequencies in several higher-threshold and longer-lead settings. At five-day lead times, the paper reports stronger forecast utility for arid parts of northern Kenya, Somalia and northern Sudan, and for some warm semi-arid areas in southern Kenya, Ethiopia and southern North Sudan. In an Ethiopia case-study region, the AI-assisted reliability slope was close to one at four and five days, while the calibrated physical forecast approached a no-added-skill climatological reference.

But the pattern was not uniform. In parts of the Ethiopian highlands at one-day lead time, the calibrated physical forecast was rated useful while the AI-assisted version was only marginally useful. At three days, the two systems were more similar. Across many regions and thresholds, both remained in the study's marginally useful category. The authors therefore do not establish that one system should replace the other. They provide a map of conditions in which the added post-processing may help and conditions in which it may not.

That local variation is exactly what an early-warning service needs to see before adoption. A headline average could hide poorer performance in a particular climate zone or at the threshold used to trigger action. The paper's decomposition by rainfall level, probability and region gives forecasters a way to match evidence to a decision. It also exposes where a conventional forecast remains preferable. The authors say the hybrid system has been used routinely in East African operational centres since 2023, but this study does not report how many warnings used it, which decisions followed, or whether communities received more useful notice.[1][2]

What this could mean for people—and what remains unproven

For meteorological services with constrained computing budgets, generating a larger probabilistic ensemble through post-processing could make uncertainty information more available. If a well-calibrated five-day forecast identifies a meaningful chance of extreme rain, agencies may have more time to pre-position supplies, alert health facilities, protect livestock or review drainage risks. Those are plausible uses, not measured outcomes from this paper. Whether extra lead time helps depends on communication, trust, transport, local authority and the resources available to act.

A separate peer-reviewed study published in the Bulletin of the American Meteorological Society on 29 October illustrates that last-mile gap in southern Africa. It synthesised community fieldwork and stakeholder engagement and reported household surveys at two sites in each of Malawi, Mozambique and Madagascar: 838, 1,103 and 760 households respectively, or 2,701 in total. Warning receipt was reported by 80% of respondents in Malawi, 77% in Mozambique and 90% in Madagascar, yet fewer than half evacuated after an alert. The authors identify late delivery, scepticism, network failures, transport and financial constraints, responsibility for land or animals, disability-related barriers and weak coordination among the reasons a received warning may not become protective action.

Those figures should not be treated as an outcome test of the East African AI-assisted forecast. The BAMS paper covers different countries, hazards, operational arrangements and survey projects, and it does not compare AI with conventional forecasting. It materially changes the practical interpretation by showing, with a stated household denominator, why a better-calibrated probability is only one part of an early-warning system. Forecast providers also need to measure reach, trust, accessibility, available transport, community participation and whether people can afford to act.

False alarms also matter. Repeated warnings that do not materialise can consume scarce budgets and weaken public response. A technically reliable probability does not prescribe the right action threshold because the cost of moving supplies differs from the cost of evacuating a settlement. The study's cost-loss framework helps users explore that trade-off, but the authors acknowledge that forecasters do not know every end user's costs. The analysis also fits individual rainfall thresholds separately, so probabilities across thresholds may not always be mutually consistent.

The work was co-developed with forecasting teams in Kenya, Ethiopia, Uganda and Rwanda and includes authors from regional meteorological services, the Intergovernmental Authority on Development Climate Prediction and Applications Centre, the World Food Programme, Oxford, Cardiff and ECMWF. Google.org funded the research through the World Food Programme; the authors declared no competing interests. That collaboration strengthens operational relevance, but funding and deployment context should remain visible when future benefits are assessed.

The next evidence should be prospective: pre-specified seasons, locally chosen thresholds, records of every alert, user decisions and consequences. Researchers should report missed events, false alarms, time gained, who received warnings and whether the information changed action. Comparisons should include updated physical and AI systems because both evolve. Until then, the paper supports cautious use of calibrated AI post-processing as an additional forecasting tool in selected settings—not a claim that AI has solved extreme-rainfall warning across East Africa.[1][2][3]