ECMWF SEAS5 · May initialization · CHIRPS v2.0 · 0.25°

Calibrated seasonal rainfall outlooks for Ethiopia

Bias-corrected, probability-calibrated June–September rainfall outlooks for Ethiopia, built from the ECMWF SEAS5 ensemble and evaluated against CHIRPS observations over 1993–2025, with a frozen 2026 forecast verified as observations arrive.

Research reconstruction. These are retrospective research products, not official forecasts of the Ethiopian Meteorology Institute (EMI) or ICPAC.

2026 JJAS outlook

50% below normal

All Ethiopia, anomaly -66 mm · JJAS R1+R2 domain 55% below, -101 mm

Skill 2017–2025 (JJAS)

RPSS +0.050

Ranked probability skill vs climatology

2026 verified so far

Jun +0.19 / Jul +0.25 / Aug +0.28

RPSS against CHIRPS; Sep & JJAS pending (Sep CHIRPS: unavailable)

Workflow

Every stage is a script in scripts/ with its own checks; documents 01–37 in docs/ record each step and decision.

1
Download
ECMWF SEAS5 (system 51) daily accumulated precipitation, May initialization, 183 days, 25 hindcast / 51 forecast members, via the Copernicus CDS.
download_seasonal_forecasts_daily_c3s.py
2
Inspect
Inventory every year: initialization, lead times, member counts, accumulation increments (GRIB packing tolerance 0.2 mm).
inspect_inputs.py
3
Prepare
De-accumulate to daily totals, label days by interval start, sum JJAS (122 days) and each month; CHIRPS on the same dates.
prepare_seasonal.py · run_monthly.py
4
Regrid
Each 1° model cell copied into its 4×4 block of 0.25° CHIRPS cells; edges and area conservation verified.
regrid_seasonal.py
5
Calibrate
Per-cell mean–variance amount correction, smoothed tercile counts, one climatology-blend weight per target (leave-one-year-out).
final_shared_blend.py
6
Evaluate
Nested cross-validation 1993–2016, fixed-fit evaluation 2017–2025, whole-year significance gates.
local_blend.py · decision_gates.py
7
Deliver
Maps, bulletin and offline gallery; forecasts frozen with SHA-256 hashes before observations are used.
finalize_forecast_delivery.py · run_operational.py
8
Verify
Official CHIRPS v2.0 p25 monthly files, overlap-checked against the archive, scored per month and season.
prepare_verification_2026.py · verify_frozen_2026.py

Data and methods

Data

Source	Details
ECMWF SEAS5	CDS `seasonal-original-single-levels`, system 51, total precipitation, 1 May initialization, 1993–2026; 25 members (1993–2016), 51 (2017–2026); 1° grid, 3–15°N, 33–48°E
CHIRPS v2.0	0.25° daily; archive 1993–2025; 2026 from official p25 monthly files, overlap-checked against the archive
Targets	JJAS (122 days) and the months Jun, Jul, Aug, Sep
Masks	Ethiopia (1,484 cells, cell-centre rule); descriptive rainfall regimes R0–R3 fixed on 1993–2025

Quality control

All 34 forecast files inventoried; re-downloaded years were bit-identical, confirming SEAS5 system 51.
Small negative daily increments (≤0.15 mm) are GRIB packing rounding (2⁻¹³ m); tolerated up to 0.2 mm, clipped, and the seasonal sum is checked to telescope.
CHIRPS sea cells (223) are permanently missing; no partial seasons enter any fit.

Calibration (final method)

Amount correction per cell, equal weight per year:
x̂ = max(0, μ_obs + r·(x − μ_model)), r = clip(σ_obs/σ_model, 0.5, 2)
Tercile probabilities from corrected members against CHIRPS terciles, smoothed: (n + 0.5)/(M + 1.5)
Climatology blend: p = (1 − λ)·p_model + λ·p_clim, one λ per target fitted on leave-one-year-out RPS (λ = JJAS 0.50, Jun 0.74, Jul 0.60, Aug 0.48, Sep 0.79)

Evaluated and not adopted: regularized Dirichlet recalibration, per-cell and regime-specific blend weights, an Ethiopia-only blend domain and removing the blend. Uncalibrated ECMWF probabilities score RPSS -0.41 for JJAS over 2017–2025.

Historical skill

Ranked probability skill score (RPSS) of the final method against climatology, Ethiopia cells, one score per year. 1993–2016 uses nested leave-one-year-out fits; 2017–2025 uses fits on 1993–2016 (exploratory, years seen during method selection). Bold: one-sided whole-year permutation p < 0.05.

Target	RPSS 1993–2016 (nested CV)	Years better	RPSS 2017–2025	Years better	Probability coverage
JJAS	+0.046 p=0.03	19/24	+0.050 p=0.06	6/9	98%
Jun	+0.010 p=0.29	14/24	+0.024 p=0.10	6/9	79%
Jul	+0.017 p=0.09	17/24	+0.036 p=0.16	5/9	79%
Aug	+0.021 p=0.09	15/24	+0.066 p=0.01	7/9	81%
Sep	+0.004 p=0.32	12/24	+0.010 p=0.25	4/9	96%

After adjusting for five targets, no gain over climatology is statistically significant; JJAS is significant on its own (p = 0.03). Monthly skill is small.

Where the forecast is skilful

Region (nested 1993–2016)	JJAS	Jun	Jul	Aug	Sep
R0 arid / marginal	+0.108 p=0.03	-0.007 p=0.56	+0.006 p=0.44	+0.073 p=0.01	+0.008 p=0.37
R1 western unimodal	+0.055 p=0.04	+0.008 p=0.33	+0.031 p=0.01	+0.004 p=0.38	+0.005 p=0.35
R2 Belg–Kiremt	+0.101 p=0.01	+0.028 p=0.21	+0.036 p=0.04	+0.050 p=0.01	+0.007 p=0.30
R3 Gu–Deyr	-0.008 p=0.64	-0.007 p=0.62	-0.030 p=0.81	-0.004 p=0.57	-0.000 p=0.51
R1+R2 JJAS domain	+0.078 p=0.01	+0.018 p=0.23	+0.034 p=0.01	+0.027 p=0.04	+0.006 p=0.29

R3 (Gu–Deyr, south and south-east) shows no clean-period skill for any target; June and September show none in any region.

Decisions tested with significance gates

A change is adopted only if it improves nested 1993–2016 scores (Holm-adjusted p < 0.05) and does not worsen 2017–2025.

Candidate change	Δ score 1993–2016	Δ score 2017–2025	Decision
Fit blend weight on Ethiopia only	+0.0003 to +0.0013	-0.0038 to -0.0001	Not adopted
Square-root amount correction (probabilities)	-0.0007 to +0.0003	-0.0033 to -0.0006	Not adopted
Remove the climatology blend	+0.0106 to +0.0316	-0.0137 to +0.0211	Not adopted
Plain climatology	+0.0017 to +0.0210	+0.0044 to +0.0303	Not adopted
Square-root amount correction (rainfall amounts, CRPS)	-0.13 to -0.10 mm	better in all targets	Passes for all targets (next cycle)

Ensemble diagnostics

Target	RPS, 51 members	RPS, 25-member subsets	Spread / error (hindcast → operational)
JJAS	0.4284	0.4307	0.72 → 0.80
Jun	0.4590	0.4600	0.68 → 0.64
Jul	0.4506	0.4522	0.74 → 0.74
Aug	0.4264	0.4284	0.78 → 0.76
Sep	0.4548	0.4557	0.78 → 0.73

51-member forecasts score at least as well as 25-member hindcast-sized ensembles, so the transfer is safe. The corrected ensemble is under-dispersive (spread/error below 1), which the climatology blend partly compensates.

2026 forecast

May-initialized, 51 members, fitted on 1993–2025 and frozen before any 2026 observation was used. Results are shown for all of Ethiopia and for the JJAS R1+R2 rainfall domain, the same two views as the operational gallery.

JJAS R1+R2 rainfall domain. Fixed 1993-2025 descriptive domain: cleaned GitHub-refined R1/R2; JJAS climatological rainfall >=120 mm and >=20% of annual rainfall. The same domain is used for Jun, Jul, Aug, Sep and JJAS. No onset gate. A presentation and summary domain, not a separate calibration, physical land mask, monthly rainfall-regime classification, or independently validated EMI zone.

Target	View	Below / near / above	Mean rainfall / reference	Mean anomaly	Share of country area
JJAS	All Ethiopia	50% / 25% / 24%	375 / 441 mm	-66.0 mm	100%
	JJAS R1+R2 rainfall domain	55% / 25% / 20%	614 / 715 mm	-100.6 mm	56%
Jun	All Ethiopia	43% / 30% / 27%	66 / 81 mm	-15.1 mm	100%
	JJAS R1+R2 rainfall domain	43% / 30% / 27%	109 / 131 mm	-22.4 mm	56%
Jul	All Ethiopia	45% / 29% / 26%	113 / 130 mm	-16.9 mm	100%
	JJAS R1+R2 rainfall domain	43% / 30% / 27%	194 / 217 mm	-24.0 mm	56%
Aug	All Ethiopia	51% / 26% / 23%	117 / 136 mm	-18.9 mm	100%
	JJAS R1+R2 rainfall domain	48% / 27% / 25%	198 / 225 mm	-26.8 mm	56%
Sep	All Ethiopia	36% / 31% / 33%	83 / 94 mm	-11.7 mm	100%
	JJAS R1+R2 rainfall domain	39% / 31% / 30%	117 / 141 mm	-24.0 mm	56%

Probabilities are area means of local grid-cell probabilities, not probabilities of the domain-total rainfall. Monthly and seasonal outlooks are calibrated separately: over all Ethiopia the corrected monthly means sum to 378 mm against 375 mm for JJAS (+3.3 mm).

2026 verification

Frozen forecasts scored against official CHIRPS v2.0 observations, for all of Ethiopia and the JJAS R1+R2 rainfall domain. Single-season results, not evidence of multi-year reliability. Positive skill means improvement over climatology on that support.

Target	View	Final RPSS	Corrected CRPSS	Corrected bias	Observed below / near / above	Observed / forecast anomaly
Jun	All Ethiopia	+0.193	+0.388	+5.4 mm	71% / 19% / 10%	-20.6 / -15.1 mm
	JJAS R1+R2 rainfall domain	+0.217	+0.399	+9.8 mm	80% / 15% / 5%	-32.1 / -22.4 mm
Jul	All Ethiopia	+0.254	+0.279	+19.1 mm	84% / 11% / 5%	-36.0 / -16.9 mm
	JJAS R1+R2 rainfall domain	+0.196	+0.253	+33.6 mm	84% / 11% / 5%	-57.6 / -24.0 mm
Aug	All Ethiopia	+0.279	+0.441	+12.9 mm	79% / 11% / 10%	-31.9 / -18.9 mm
	JJAS R1+R2 rainfall domain	+0.356	+0.440	+20.6 mm	88% / 10% / 2%	-47.4 / -26.8 mm
Sep	both views	pending	—	—	—	—
JJAS	both views	pending	—	—	—	—

January initialization · FMAM 2026 forecast

The same pipeline applied to another season: ECMWF SEAS5 initialized on 1 January 2026, 51 members, calibrated against CHIRPS for FMAM seasons 1993–2025 (33 seasons; 25-member hindcasts to 2016, 51 members after). The season runs from February to May 2026. Verification needs CHIRPS through May 2026.

FMAM R2 (Belg) rainfall domain. Fixed 1993-2025 descriptive domain (Belg): rainfall regime R2 (GitHub-refined classification of the CHIRPS daily climatology) with FMAM climatological CHIRPS rainfall >=80 mm; patches smaller than 3 cells removed. Scientific masking walkthrough logic; no onset gate. The same domain is used for the season and each of its months. A presentation and summary domain, not a separate calibration, physical land mask, monthly rainfall-regime classification, or independently validated EMI zone.

Target	View	Below / near / above	Mean rainfall / reference	Mean anomaly	Share of country area	Probability coverage
Feb 2026	All Ethiopia	28% / 39% / 33%	14 / 13 mm	+0.8 mm	100%	74%
	FMAM R2 (Belg) rainfall domain	28% / 37% / 35%	18 / 16 mm	+2.0 mm	28%	95%
Mar 2026	All Ethiopia	31% / 35% / 34%	48 / 42 mm	+5.9 mm	100%	87%
	FMAM R2 (Belg) rainfall domain	31% / 35% / 34%	55 / 50 mm	+5.5 mm	28%	100%
Apr 2026	All Ethiopia	35% / 32% / 33%	91 / 94 mm	-2.7 mm	100%	100%
	FMAM R2 (Belg) rainfall domain	37% / 31% / 32%	73 / 77 mm	-3.2 mm	28%	100%
May 2026	All Ethiopia	36% / 34% / 30%	95 / 99 mm	-4.8 mm	100%	100%
	FMAM R2 (Belg) rainfall domain	37% / 33% / 30%	66 / 70 mm	-3.9 mm	28%	100%
FMAM 2026	All Ethiopia	34% / 32% / 34%	248 / 249 mm	-1.0 mm	100%	100%
	FMAM R2 (Belg) rainfall domain	34% / 33% / 33%	212 / 213 mm	-0.3 mm	28%	100%

Outside the rainfall domain much of Ethiopia is outside its main rainy season; cells with negligible climatological rainfall have no tercile probabilities, which lowers national probability coverage.

Verification against CHIRPS

Target	View	Final RPSS	Corrected CRPSS	Corrected bias	Observed below / near / above
Feb 2026	both views	pending	—	—	—
Mar 2026	All Ethiopia	+0.037	+0.086	-53.7 mm	0% / 4% / 96%
	FMAM R2 (Belg) rainfall domain	+0.032	+0.079	-38.3 mm	0% / 8% / 92%
Apr 2026	All Ethiopia	+0.025	+0.038	+3.7 mm	36% / 31% / 33%
	FMAM R2 (Belg) rainfall domain	+0.036	+0.080	+5.7 mm	42% / 27% / 30%
May 2026	All Ethiopia	+0.017	-0.012	+0.0 mm	33% / 50% / 16%
	FMAM R2 (Belg) rainfall domain	+0.044	+0.089	+4.4 mm	28% / 63% / 9%
FMAM 2026	both views	pending	—	—	—

Single-season scores against official CHIRPS v2.0 observations; positive skill beats climatology on that support. Maps are in the map explorer (product: Verification).

Historical skill for this season

Target	RPSS 1993–2016 (nested CV)	RPSS 2017–2025	Climatology weight λ
Feb	+0.032 p=0.03 16/24 yrs	-0.006 p=0.58 5/9 yrs	0.70
Mar	+0.001 p=0.42 16/24 yrs	+0.003 p=0.33 6/9 yrs	0.81
Apr	+0.002 p=0.43 15/24 yrs	+0.007 p=0.32 6/9 yrs	0.71
May	+0.026 p=0.05 17/24 yrs	-0.000 p=0.50 6/9 yrs	0.57
FMAM	+0.027 p=0.08 18/24 yrs	-0.004 p=0.57 4/9 yrs	0.43

RPSS of the final method against climatology, Ethiopia cells; nested leave-one-year-out fits for the development years and fits on those years for the later evaluation years. Bold: one-sided whole-year permutation p < 0.05. Maps for every target and both views are in the map explorer.

September initialization · ONDJ 2026/27 forecast

The same pipeline applied to another season: ECMWF SEAS5 initialized on 1 September 2026, 51 members, calibrated against CHIRPS for ONDJ seasons 1993–2024 (32 seasons; 25-member hindcasts to 2016, 51 members after). The season crosses the year boundary (October 2026 to January 2027). Verification needs CHIRPS through January 2027.

ONDJ R3 (Deyr) rainfall domain. Fixed 1993-2025 descriptive domain (Deyr / Hagaya): rainfall regime R3 (GitHub-refined classification of the CHIRPS daily climatology) with ONDJ climatological CHIRPS rainfall >=30 mm; patches smaller than 3 cells removed. Scientific masking walkthrough logic; no onset gate. The same domain is used for the season and each of its months. A presentation and summary domain, not a separate calibration, physical land mask, monthly rainfall-regime classification, or independently validated EMI zone.

Target	View	Below / near / above	Mean rainfall / reference	Mean anomaly	Share of country area	Probability coverage
Oct 2026	All Ethiopia	33% / 32% / 35%	87 / 85 mm	+2.5 mm	100%	99%
	ONDJ R3 (Deyr) rainfall domain	27% / 30% / 43%	126 / 106 mm	+20.5 mm	39%	100%
Nov 2026	All Ethiopia	26% / 26% / 48%	53 / 36 mm	+16.7 mm	100%	98%
	ONDJ R3 (Deyr) rainfall domain	21% / 22% / 57%	83 / 51 mm	+31.4 mm	39%	100%
Dec 2026	All Ethiopia	33% / 33% / 34%	13 / 10 mm	+2.3 mm	100%	62%
	ONDJ R3 (Deyr) rainfall domain	32% / 31% / 37%	16 / 12 mm	+4.1 mm	39%	54%
Jan 2027	All Ethiopia	31% / 35% / 34%	12 / 10 mm	+1.8 mm	100%	64%
	ONDJ R3 (Deyr) rainfall domain	29% / 37% / 34%	9 / 7 mm	+1.5 mm	39%	43%
ONDJ 2026/27	All Ethiopia	28% / 27% / 45%	167 / 141 mm	+26.5 mm	100%	100%
	ONDJ R3 (Deyr) rainfall domain	19% / 24% / 57%	240 / 176 mm	+64.3 mm	39%	100%

Outside the rainfall domain much of Ethiopia is outside its main rainy season; cells with negligible climatological rainfall have no tercile probabilities, which lowers national probability coverage.

Historical skill for this season

Target	RPSS 1993–2016 (nested CV)	RPSS 2017–2024	Climatology weight λ
Oct	-0.001 p=0.52 13/24 yrs	+0.024 p=0.25 6/8 yrs	0.44
Nov	+0.046 p=0.15 15/24 yrs	+0.005 p=0.47 4/8 yrs	0.23
Dec	+0.011 p=0.26 15/24 yrs	+0.027 p=0.18 6/8 yrs	0.64
Jan	+0.002 p=0.37 15/24 yrs	+0.004 p=0.35 5/8 yrs	0.82
ONDJ	+0.020 p=0.31 14/24 yrs	+0.063 p=0.25 4/8 yrs	0.28

Map explorer

Choose a target season or month, a product and a rainfall domain. Statistics and maps are the operational gallery's presentation layers; the rainfall-domain view is for display and summaries only.

Target season or monthProductView (rainfall domain)

Limitations

Nine fixed-fit evaluation years and one verification season give wide uncertainty; regional and monthly results are noisy.
The 2017–2025 period was inspected during method development; clean evidence comes from nested 1993–2016 scores.
The 0.25° maps repeat 1° model information; they do not add dynamical detail.
Not an official EMI or ICPAC product.

Reproduce

git clone https://github.com/YonSci/calibrated_seasonal_rainfall.git
python -m venv .venv && .venv\Scripts\activate && pip install -r requirements.txt
python scripts\prepare_seasonal.py --config config\project.json --all-years
python scripts\final_shared_blend.py --config config\project.json --region-mask data\masks\ethiopia_common.nc
python scripts\run_operational.py --workflow all

Raw ECMWF and CHIRPS data are not stored in the repository; download instructions are in docs/37. Status and decisions: docs/36.

Calibrated seasonal rainfall outlooks for Ethiopia

Workflow

Download

Inspect

Prepare

Regrid

Calibrate

Evaluate

Deliver

Verify

Data and methods

Data

Quality control

Calibration (final method)

Historical skill

Where the forecast is skilful

Decisions tested with significance gates

Ensemble diagnostics

2026 forecast

2026 verification

January initialization · FMAM 2026 forecast

Verification against CHIRPS

Historical skill for this season

September initialization · ONDJ 2026/27 forecast

Historical skill for this season

Map explorer

Limitations

Reproduce