Ethiopia rainfall forecast verification — JJAS 2026, May initialization
May 2026 initialization | JJAS 2026 season and its months | ECMWF seasonal forecasts | CHIRPS v3.0 verification | Reference period 1993–2025
Verified targets: Jun, Jul, Aug. Pending in this report: Sep, JJAS. No full-JJAS conclusion is inferred from monthly results.
This is a retrospective assessment of the reconstructed forecasts. It does not establish an actual May 2026 issuance, independent prospective validation, or an official EMI/ICPAC forecast. All results below concern the submitted verification evidence.
Decision and principal findings
Retain frozen shared-blend forecasts. Regional annotations are post-event verification findings, not forecast-time confidence labels or new masks. No automatic model switch.
Jun: corrected rainfall CRPSS +40.4%; final probability RPSS +12.5%; amount-mean error +10.93 mm. Probability results cover 99.9% of country grid area.
Jul: corrected rainfall CRPSS +22.7%; final probability RPSS +19.8%; amount-mean error +22.86 mm. Probability results cover 86.1% of country grid area.
Aug: corrected rainfall CRPSS +43.0%; final probability RPSS +27.7%; amount-mean error +15.49 mm. Probability results cover 89.6% of country grid area.
Positive skill means a lower score than the stated climatological reference on the same support. It is not percentage forecast accuracy. Positive mean error means the forecast was wetter than observed, which can occur even when the forecast anomaly is below normal.
Country results
| Target | Raw CRPS mm | Corrected CRPS mm | Climatology CRPS mm | Corrected CRPSS | Final RPSS |
|---|---|---|---|---|---|
| Jun | 28.45 | 14.01 | 23.48 | +40.4% | +12.5% |
| Jul | 39.61 | 25.99 | 33.62 | +22.7% | +19.8% |
| Aug | 27.71 | 15.46 | 27.14 | +43.0% | +27.7% |
Main JJAS rainfall domain: JJAS R1+R2 rainfall domain
This fixed pre-2026 domain contains 833 cells and 55.9% of country grid area. Definition: cleaned R1/R2 classes with the previously defined rainfall criteria. Monthly rows evaluate each month inside that domain; they are not a complete-JJAS assessment.
| Target | Amount CRPSS | Smoothed RPSS | Final RPSS | Mean error mm | Probability coverage |
|---|---|---|---|---|---|
| Jun | +41.0% | +55.9% | +14.4% | 19.95 | 100.0% |
| Jul | +20.5% | +27.4% | +17.7% | 40.94 | 100.0% |
| Aug | +41.1% | +51.6% | +28.9% | 23.58 | 100.0% |
Regional performance notes
Jun 2026
| Regime | Amount CRPSS | Smoothed RPSS | Final RPSS | Probability coverage | Blend effect |
|---|---|---|---|---|---|
| R0: arid / marginal | -16.0% | -11.5% | +4.3% | 100.0% | Helped |
| R1: western unimodal | +33.0% | +43.3% | +12.0% | 100.0% | Weakened |
| R2: Belg–Kiremt rule | +53.4% | +67.8% | +16.6% | 100.0% | Weakened |
| R3: Gu–Deyr rule | +36.7% | +28.1% | +10.3% | 99.8% | Weakened |
Jul 2026
| Regime | Amount CRPSS | Smoothed RPSS | Final RPSS | Probability coverage | Blend effect |
|---|---|---|---|---|---|
| R0: arid / marginal | +27.1% | +30.9% | +24.0% | 100.0% | Weakened |
| R1: western unimodal | +2.5% | +5.1% | +8.3% | 100.0% | Helped |
| R2: Belg–Kiremt rule | +31.7% | +47.5% | +26.2% | 100.0% | Weakened |
| R3: Gu–Deyr rule | +53.5% | +26.6% | +24.0% | 64.4% | Weakened |
Aug 2026
| Regime | Amount CRPSS | Smoothed RPSS | Final RPSS | Probability coverage | Blend effect |
|---|---|---|---|---|---|
| R0: arid / marginal | +64.1% | +86.6% | +46.3% | 100.0% | Weakened |
| R1: western unimodal | +36.7% | +47.6% | +25.3% | 99.8% | Weakened |
| R2: Belg–Kiremt rule | +44.7% | +55.6% | +32.4% | 100.0% | Weakened |
| R3: Gu–Deyr rule | +49.8% | +18.3% | +21.8% | 73.4% | Helped |
June R0: raw, corrected and climatological CRPS are 3.36, 2.08 and 1.79 mm. Corrected mean rainfall is 3.26 mm versus 5.34 mm observed. Final RPSS is +4.3%. Small climatological rainfall/error scales make percentage skill particularly sensitive here; report absolute scores alongside skill.
July R2: corrected mean error is +60.75 mm. Observed and forecast mean anomalies are -94.11 and -33.36 mm. This documents a remaining error in the magnitude of the rainfall anomaly.
August R3: smoothed RPSS is +18.3%, versus +21.8% after blending. The final above-normal-category Brier Skill Score is +17.4%. Overall RPS improvement therefore does not mean every category improved over climatology. Probability coverage is 73.4% of R3 area.
These labels describe the observed performance of these targets. They do not justify hiding weak regions, declaring confidence at issuance, applying a new mask, or replacing forecast values after observing outcomes.
Historical context
Historical review was not supplied to this build. No historical ranking is inferred.
Methods and interpretation
Rainfall amounts: equal-year, cell-specific mean–variance bias correction. All 51 operational members are retained. The ensemble and climatological reference are evaluated using empirical CRPS; their finite ensemble sizes are 51 and 33, respectively, without an iid/fair-ensemble adjustment.
Probabilities: observed training-period terciles, member counts, alpha = 0.5 additive smoothing, then the selected shared climatology blend. The final product does not use the experimental Dirichlet, local or regime-specific probability mappings. Probability blending does not change the corrected rainfall members or amount CRPS.
RPS is the sum of the first two cumulative-category squared errors. Brier scores are category-specific; log loss uses natural logarithms and the existing probability floor. Country and domain scores use grid-cell area weights on identical support across methods within each metric family.
Amount and probability masks differ; unscored cells are not near-normal conditions. R0–R3 are the reconciled GitHub-derived climatological rules, not administrative boundaries or independently validated official EMI zones. No onset-detection gate is used.
One verification year cannot establish multi-year reliability or statistical significance. Spatial cells and overlapping months/JJAS are not independent evaluation samples. No pooled overall score is produced. The 2026 outcomes are now inspected evidence and must not be reused as an untouched test set.
Verification maps
Completion and future work
Retain the frozen forecasts and archive this report as a verification addendum. A future calibration revision needs a separate historical evaluation with all preprocessing and model selection inside year-based validation. This report does not fit or adopt a revised model.
Complete the remaining months and JJAS only after the official CHIRPS v3.0 monthly observations pass the existing completeness and overlap checks. Then regenerate this report with all targets. Until then, no full-JJAS verification result is available in this report.
Evidence and reproducibility
Built on the project computer. Frozen forecast hashes checked before and after report construction; Step 30 report, field and observation hashes matched to Step 31. Summary consistency and disjoint-regime reconstruction checked. No new calibration or scoring performed.