Historical skill · Sep 2026
How forecasts made with the same method scored in past years against climatology (the observed frequency of each tercile in the training years, about one-third each). The main number is the cross-validated 1993–2016 score: each year is forecast with parameters fitted without that year. Later years are exploratory: they were seen while the method was chosen.
Takeaway — Sep 2026
Historical skill is uncertain: cross-validated 1993–2016 RPSS is +0.006, and its interval (−0.012 to +0.024) includes no improvement over climatology. The exploratory 2017–2025 result (+0.006) is encouraging but those years were inspected during method selection. The 2026 verification (+0.086) is one season and does not override the historical uncertainty.
Stronger historical evidence (interval above zero): JJAS +0.044, Jul +0.031.
| Target | Cross-validated 1993–2016 (main) | In words | Exploratory 2017–2025 |
|---|---|---|---|
| Sep | +0.006 (−0.012 to +0.024) | Small estimated improvement; skill uncertain | +0.006 (−0.018 to +0.030) |
| JJAS | +0.044 (+0.002 to +0.084) | Modest improvement over climatology (interval above zero) | +0.032 (+0.010 to +0.052) |
| Jun | −0.001 (−0.035 to +0.027) | No demonstrated improvement over climatology | +0.015 (−0.007 to +0.038) |
| Jul | +0.031 (+0.004 to +0.058) | Modest improvement over climatology (interval above zero) | +0.027 (−0.021 to +0.079) |
| Aug | +0.011 (−0.006 to +0.027) | Small estimated improvement; skill uncertain | +0.031 (+0.012 to +0.054) |
RPSS +0.05 means a 5% lower ranked probability score than climatology. An interval that includes zero means the gain is not established.
Technical detail: years better, p-values and blend weights
| Target | Years better (cross-validated) | p (one target / Holm) | Years better (exploratory) | Blend weight λ |
|---|---|---|---|---|
| JJAS | 17 of 24 | 0.029 / 0.115 | 7 of 9 | 0.59 |
| Jun | 15 of 24 | 0.510 / 0.537 | 6 of 9 | 0.83 |
| Jul | 19 of 24 | 0.019 / 0.094 | 5 of 9 | 0.60 |
| Aug | 16 of 24 | 0.103 / 0.308 | 7 of 9 | 0.58 |
| Sep | 14 of 24 | 0.268 / 0.537 | 4 of 9 | 0.81 |
Each interval and one-target p-value describes that target alone. Because 5 targets are examined, the Holm-adjusted p is the stricter test: no target remains significant after adjustment. λ is the weight given to climatology in the blend (higher = closer to climatology).
Historical verification explorer
The evidence behind the scores above, for Sep 2026 in this area. Choose the evaluation period; all diagnostics use the blended probabilities against climatology.
Performance by year
Probability skill (RPSS) of each year; better than climatology in 14 of 24 years. Weakest years: 1993 (−0.067), 2015 (−0.057), 2014 (−0.054).
Reliability
Points on the dashed diagonal are reliable: events forecast with probability p happen p of the time. Points flatter than the diagonal mean over-confident probabilities. Open points have fewer than 100 cell-years (0.5% of the sample, at least 50) and are not joined. Hover a point for its count.
Reliability table (forecast bins, observed frequency, counts)
| Category | Bin | Mean forecast | Observed frequency | Cell-years |
|---|---|---|---|---|
| below normal | 20%–30% | 29% | 26% | 2,141 |
| below normal | 30%–40% | 34% | 35% | 15,323 |
| below normal | 40%–50% | 42% | 40% | 2,528 |
| near normal | 20%–30% | 28% | 26% | 9,453 |
| near normal | 30%–40% | 33% | 34% | 10,485 |
| near normal | 40%–50% | 41% | 72% | 54 |
| above normal | 20%–30% | 29% | 37% | 2,536 |
| above normal | 30%–40% | 35% | 33% | 15,942 |
| above normal | 40%–50% | 42% | 46% | 1,514 |
How strong are the signals?
Share of cell-years by the leading tercile probability.
| Category | Forecast | Climatology | Brier skill |
|---|---|---|---|
| below normal | 0.2246 | 0.2264 | +0.008 |
| near normal | 0.2101 | 0.2123 | +0.011 |
| above normal | 0.2259 | 0.2268 | +0.004 |
Rainfall amounts
Historical rainfall-amount diagnostics (CRPS, CRPSS and amount errors) have not been generated for Sep 2026 in this area and period; this needs a separate cross-validated amount evaluation. Single-season amount scores are in the Verification section.
Performance by year
Probability skill (RPSS) of each year; better than climatology in 4 of 9 years. Weakest years: 2019 (−0.052), 2023 (−0.030), 2024 (−0.014).
Reliability
Points on the dashed diagonal are reliable: events forecast with probability p happen p of the time. Points flatter than the diagonal mean over-confident probabilities. Open points have fewer than 50 cell-years (0.5% of the sample, at least 50) and are not joined. Hover a point for its count.
Reliability table (forecast bins, observed frequency, counts)
| Category | Bin | Mean forecast | Observed frequency | Cell-years |
|---|---|---|---|---|
| below normal | 20%–30% | 29% | 12% | 911 |
| below normal | 30%–40% | 34% | 19% | 6,466 |
| below normal | 40%–50% | 41% | 11% | 120 |
| near normal | 20%–30% | 29% | 20% | 1,008 |
| near normal | 30%–40% | 33% | 30% | 6,486 |
| near normal | 40%–50% | 41% | 33% | 3 |
| above normal | 20%–30% | 29% | 54% | 781 |
| above normal | 30%–40% | 34% | 53% | 6,538 |
| above normal | 40%–50% | 41% | 58% | 178 |
How strong are the signals?
Share of cell-years by the leading tercile probability.
| Category | Forecast | Climatology | Brier skill |
|---|---|---|---|
| below normal | 0.1719 | 0.1721 | +0.001 |
| near normal | 0.2027 | 0.2053 | +0.013 |
| above normal | 0.2866 | 0.2893 | +0.009 |
Rainfall amounts
Historical rainfall-amount diagnostics (CRPS, CRPSS and amount errors) have not been generated for Sep 2026 in this area and period; this needs a separate cross-validated amount evaluation. Single-season amount scores are in the Verification section.
Historical performance by rainfall region
Probability skill by rainfall region (full regional matrix)
| Region | JJAS | Jun | Jul | Aug | Sep |
|---|---|---|---|---|---|
| All Ethiopia | +0.027 (−0.006 to +0.058) 18/24 yrs | 0.000 (−0.025 to +0.019) 15/24 yrs | +0.014 (−0.016 to +0.042) 17/24 yrs | +0.012 (−0.006 to +0.030) 12/24 yrs | +0.003 (−0.012 to +0.018) 14/24 yrs |
| R0 arid / marginal | +0.066 (−0.003 to +0.133) 15/24 yrs | −0.001 (−0.034 to +0.027) 13/24 yrs | +0.004 (−0.070 to +0.078) 14/24 yrs | +0.042 (+0.002 to +0.081) 17/24 yrs | +0.003 (−0.036 to +0.037) 14/24 yrs |
| R1 western unimodal | +0.031 (−0.004 to +0.066) 15/24 yrs | −0.007 (−0.038 to +0.020) 14/24 yrs | +0.030 (+0.004 to +0.057) 16/24 yrs | −0.002 (−0.017 to +0.012) 15/24 yrs | +0.006 (−0.014 to +0.026) 14/24 yrs |
| R2 Belg–Kiremt | +0.055 (−0.006 to +0.115) 17/24 yrs | +0.006 (−0.036 to +0.037) 14/24 yrs | +0.033 (−0.008 to +0.074) 17/24 yrs | +0.024 (+0.002 to +0.046) 14/24 yrs | +0.006 (−0.015 to +0.029) 14/24 yrs |
| R3 Gu–Deyr | −0.001 (−0.037 to +0.034) 14/24 yrs | 0.000 (−0.020 to +0.017) 14/24 yrs | −0.028 (−0.090 to +0.027) 12/24 yrs | +0.007 (−0.020 to +0.035) 13/24 yrs | −0.001 (−0.028 to +0.023) 13/24 yrs |
| JJAS R1+R2 rainfall domain | +0.044 (+0.002 to +0.084) 17/24 yrs | −0.001 (−0.035 to +0.027) 15/24 yrs | +0.031 (+0.004 to +0.058) 19/24 yrs | +0.011 (−0.006 to +0.027) 16/24 yrs | +0.006 (−0.012 to +0.024) 14/24 yrs |
Positive values mean lower RPS than climatology. Intervals resample whole years; regions are the fixed 1993–2025 rainfall regimes, and the last rows are this cycle's rainfall domains. Regional results are noisier than national ones.