Releases and change log
Each release is the site as published at one commit. Archived releases open the exact page published then.
Release 2026-10-19.0 · 2026-10-19
How statistical downscaling is implemented
- Display or explanatory-text change: The statistical downscaling page explains how canonical correlation analysis and principal component regression are implemented (fields, EOFs, the regression and canonical steps, the choice of modes, the forecast distribution and the evaluation) and shows each comparison under the benchmark (Holm) and under the selection rule (Benjamini-Hochberg). docs/53 section 2 gives the full description with the function behind each step.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
This is the current release.
Release 2026-10-17.3 · 2026-10-17
Method selection with Benjamini-Hochberg
- Change to the scientific method: The method selection, the bias-correction suggestions and the summary across seasons adjust for testing many methods and targets with Benjamini-Hochberg (false discovery rate) instead of Holm, within each benchmark family, recomputed from the p-values each benchmark stored (decision of 17 October 2026, after the sensitivity analyses). Eight target-product combinations now have a candidate instead of six: JJAS June amounts (gamma quantile mapping) and JJAS season amounts (principal component regression) are new, and for JJAS August amounts principal component regression replaces Variance Scaling. None is recommended yet: that needs three validation seasons. The published forecasts are unchanged. docs/50.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-17.2 · 2026-10-17
Real-time issue from 2027, statistical downscaling, more statistical power, dry-season maps
- Change to the scientific method: From 2027 each forecast is issued in real time: frozen and published as a GitHub release before its target period (scripts/issue_forecast.py), with every statement about its status read from its issue record; targets already under way when issued (February and FMAM of the February-initialized FMAM cycle) are labelled in season. From 2027 the forecasts use the first 25 SEAS5 members, as the hindcasts. 2026 forecasts remain research reconstructions. docs/52.
- New evidence or diagnostics (method unchanged): Statistical downscaling by model output statistics (canonical correlation analysis and principal component regression, the CPT approach of ICPAC and IRI) tested against the published pipeline for the four cycles: confirmed only for JJAS September amounts; worse than the published approach when pooled over all targets, though better in the 2026 JJAS season. docs/53.
- New evidence or diagnostics (method unchanged): More statistical power on the Research overview: every method tested over all 33 out-of-sample years and pooled over its targets and seasons, with Holm and Benjamini-Hochberg corrections. For amounts, Variance Scaling, square-root MVS, gamma QM and quantile regression are significantly better when pooled; for probabilities no method is. The selection rule is unchanged.
- Display or explanatory-text change: Tercile maps show cells without terciles as "Dry season: no tercile forecast" in their own colour.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-17.1 · 2026-10-17
The platform is now called the Seasonal Rainfall Outlook Platform
- Display or explanatory-text change: New name, Seasonal Rainfall Outlook Platform, in the header, page titles, home page, classic page, PDF bulletins and the README of every download package (previously Seasonal Calibrated Rainfall Outlook, SCRO). Content and forecasts are unchanged.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-16.7 · 2026-10-16
Method comparison by category
- Display or explanatory-text change: Research > Which methods are better and Method selection: 27 methods in four categories (12 bias corrections, 8 probability calibrations, 6 rainfall-distribution models, 1 two-stage calibration); the Leung–Dirichlet calibration counts as one method (the whole calibration for probabilities, its corrected amounts for amounts). Findings unchanged; the published forecasts are unchanged.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-16.6 · 2026-10-16
Larger site title
- Display or explanatory-text change: The header shows Seasonal Calibrated Rainfall Outlook as a larger title with an SCRO tag (on smaller screens the title alone, over two lines).
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-16.5 · 2026-10-16
The platform is now called SCRO: Seasonal Calibrated Rainfall Outlook
- Display or explanatory-text change: The site is named Seasonal Calibrated Rainfall Outlook (SCRO): header, page titles, home page, classic page, PDF bulletins and the README of every download package. Content and forecasts are unchanged.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-16.4 · 2026-10-16
Research overview: which methods are better, and which to use in each season
- Display or explanatory-text change: Research (first entry of the Research menu): how a method is judged for tercile probabilities and rainfall amounts, the method to use in each season (the published approach everywhere for now, with the candidates and their validation progress), findings, and comparison tables of every method. The outlook-map switch now reads Grid cells | Smoothed. The published forecasts are unchanged.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-16.3 · 2026-10-16
Outlook map: switch between the smoothed map and the grid cells
- Display or explanatory-text change: The outlook map at the top of every outlook page has a Smoothed / Grid cells (0.25°) switch. The choice is remembered in the browser and shared with the map viewer further down the page; ?style=cells opens a page on the grid cells. The published forecasts are unchanged.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-16.2 · 2026-10-16
Leung–Dirichlet calibration: the name in the code, files and addresses
- Display or explanatory-text change: The Leung–Dirichlet calibration now uses its name throughout: scripts, configurations, output folders, site data, figure and map files, download names and the release log. The benchmarks were rerun under the new names with identical results. The authors of process_calib.py stay credited; the previous page address forwards to the new one. The published forecasts are unchanged.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-16.1 · 2026-10-16
Leung–Dirichlet calibration: the new name on every page
- Display or explanatory-text change: The calibration from process_calib.py (modified-Leung member-wise affine correction + identity-anchored Dirichlet) is now called the Leung–Dirichlet calibration on every page, figure, map and table: Research > Leung–Dirichlet calibration (the previous address forwards there), method selection and the summary of all methods. The source is credited. Results unchanged; the published forecasts are unchanged.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-15.6 · 2026-10-15
Leung–Dirichlet for FMAM and JJAS; every method against the published approach
- New verification observations: JJAS 2026 complete: September and the JJAS season verified against CHIRPS (scored by the daily check on 2026-10-15). On the JJAS R1+R2 rainfall domain the published probabilities score RPSS +0.09 (Sep) and +0.45 (JJAS), the corrected amounts CRPSS +0.16 and +0.54.
- New evidence or diagnostics (method unchanged): The Leung–Dirichlet calibration tested for FMAM (January initialization) and JJAS (May initialization) as for ONDJ: nothing is confirmed better than the published approach for either season (FMAM February, Dirichlet on MVS, is promising but not significant after the Holm adjustment). The 2026 FMAM and JJAS seasons are scored as independent validation seasons for every Leung–Dirichlet pipeline.
- New evidence or diagnostics (method unchanged): Research > Method selection: a table of every bias-correction, probability-stage, distribution and Leung–Dirichlet method against the published approach over all seasons and targets (confirmed better, better on average 1993-2016 and 2017 onward, better in the 2026 validation seasons), refreshed by the daily scoring. The published forecasts are unchanged.
- Display or explanatory-text change: Research > Leung–Dirichlet calibration: FMAM and JJAS sections with grid-cell maps on each season's rainfall domains and All Ethiopia, and the 2026 validation table.
- Display or explanatory-text change: A verification-only run (the daily scoring) rebuilt the cycle gallery without its forecast views, so the site build could not find them; the gallery now keeps the rendered forecast views. Forecasts are unchanged.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-15.5 · 2026-10-15
Leung–Dirichlet maps on the grid cells, with the seasonal rainfall domain
- Display or explanatory-text change: Research > Leung–Dirichlet calibration: the tercile outlook and the below-, near- and above-normal probabilities of every pipeline for ONDJ 2026/27 and its months are drawn on the 0.25-degree grid cells (no smoothing), first on the ONDJ R3 (Deyr) rainfall domain and then on All Ethiopia, next to the published grid-cell maps of the same view. They replace the smoothed maps of release 2026-10-15.4. Not published as a forecast; the published forecasts are unchanged.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-15.4 · 2026-10-15
Leung–Dirichlet: tercile outlook maps, ONDJ 2026/27
- Display or explanatory-text change: Research > Leung–Dirichlet calibration: the rainfall tercile outlook of each Leung–Dirichlet pipeline for Oct, Nov, Dec, Jan and ONDJ 2026/27, drawn exactly as the published outlook map and shown next to it (not published as a forecast). The published forecasts are unchanged.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-15.3 · 2026-10-15
Leung–Dirichlet: tercile-probability maps and reliability diagrams
- Display or explanatory-text change: Research > Leung–Dirichlet calibration: maps of the below-, near- and above-normal probabilities of every pipeline for each ONDJ 2026 target next to the published ones (not published as a forecast), and reliability diagrams of every pipeline from the 1993-2016 cross-validated hindcasts and the 2017-2024 test years. The published forecasts are unchanged.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-15.2 · 2026-10-15
Leung–Dirichlet calibration (modified-Leung affine + Dirichlet) tested for ONDJ
- New evidence or diagnostics (method unchanged): The Leung–Dirichlet calibration (process_calib.py, Munoz and Demissie, ILRI) implemented as separate methods: the modified-Leung member-wise affine bias correction and the Identity-Anchored Dirichlet probability recalibration, tested as uploaded and stage by stage against the published pipeline for ONDJ (September initialization), 1993-2016 cross-validated and 2017-2024 test years. No Leung–Dirichlet method improves the published tercile probabilities; the raw member counts are overconfident and the Dirichlet mostly flattens them. The Leung affine is confirmed better than the MVS ensemble for January rainfall amounts and enters the method selection as an alternative (square-root MVS stays selected). The 2026 forecast with the Leung–Dirichlet calibration is shown for comparison and is not published. The published forecasts are unchanged.
- Display or explanatory-text change: Research > Leung–Dirichlet calibration: what the uploaded code does, the comparison, the fitted Dirichlet and the 2026 maps.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-15.1 · 2026-10-15
Method selection per season, target and product; menu fix
- New evidence or diagnostics (method unchanged): Method selection: for each season, target and product (tercile probabilities, rainfall amounts) the platform selects the approach the evidence supports from every benchmark together (amount corrections, probability-stage methods, distribution models): a method must be confirmed better than the published approach over 1993-2016 and not worse over 2017 onward, then becomes recommended after three independent validation seasons that are not worse. One change at a time. Six combinations currently have a candidate; none is recommended yet. Every distribution-model candidate now has its own shadow forecast. The published forecasts are unchanged.
- Display or explanatory-text change: The Research and About menus open again (they were hidden by the menu bar); Research > Method selection and an approach box on every outlook page.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-14.8 · 2026-10-14
The new site at the main address; search, glossary, methods, domains and packages
- Display or explanatory-text change: The new site is now at the main address; the previous single page is kept unchanged at /classic/, its links (anchors and cycle views) open the matching new page, and the /next/ preview addresses forward to the new ones.
- Display or explanatory-text change: New: site search; a glossary, with the first use of each term on a page explained on hover; verification-status badges; the methods and data page; the rainfall climatology and domain review per season.
- Display or explanatory-text change: Each cycle's download package is now a ZIP built with the site (every map, the verification reports, the official-outlook comparison files and the forecast NetCDF files), with the bulletin, summary, metadata and README next to it. This fixes the package button of the previous page, which had failed since the downloads moved to release assets (release 2026-10-14.3).
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-14.7 · 2026-10-14
Research pages: summaries and sub-pages
- Display or explanatory-text change: On the new pages, each cycle's bias-correction and probability-calibration benchmark opens on a short summary (what the evidence supports and the suggested method; what each phase tested and what was confirmed better), with the detail on sub-pages: interpretation, data and methods, screening (one region at a time), confirmation, ratio limits, shadow and validation, figures and downloads; Phase 1, 2 and 3, the amount and range shadow and downloads. A sub-menu keeps the sub-page when the cycle changes, long pages list their headings beside the text, and previous/next links lead through the sub-pages. Every score is unchanged (rechecked against the current page).
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-14.6 · 2026-10-14
New site layout and navigation
- Display or explanatory-text change: The new pages (/next/) put the forecast first: the home page opens on the current outlook (headline, probabilities, map, amount, signal and skill) with every target as a small map; the menu is Outlook, Verification, Skill, Research, About and Downloads. Forecast pages carry a context bar with breadcrumbs, the cycle, area and target, and the sections Outlook and maps, Official outlooks, Verification and Historical skill, which keeps the selection when the section changes. The official-outlook comparison has its own page, verification and skill have one page per target, and the amount and range shadow of every cycle is on one research page. Content and numbers are unchanged and were rechecked against the current page for every cycle, area and target.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-14.5 · 2026-10-14
Outlooks, maps and verification on the new pages; daily scoring of new observations
- Display or explanatory-text change: The forecast outlooks, maps, verification and historical performance are on the new pages (/next/): one page per cycle, area and target with the map viewer and the official-outlook comparison, and per cycle and area the verification and the historical verification explorer. Their text was checked against the current page for every cycle, area and target.
- Display or explanatory-text change: New observations are checked daily and each season is scored as soon as its CHIRPS month is published (verification, official-outlook comparison, bias-correction validation, the amount and range shadow and the method suggestions); publishing the new scores stays a release step.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-14.4 · 2026-10-14
Probability calibration, Phase 3
- New evidence or diagnostics (method unchanged): Phase 3 of the probability and rainfall-distribution calibration, for all four cycles: Bayesian joint probability, BMA for precipitation and quantile regression forests, scored for terciles, amounts (CRPS), quantiles and interval coverage. No model is confirmed better than the published probabilities or the MVS-corrected ensemble. BMA is the strongest (CRPSS above the MVS ensemble in 13 of 20 targets, none significant); BJP and BMA give near-calibrated 10-90% ranges (78-86%), the forest too narrow ones (69-76%). Across Phases 1-3, the incumbent climatology blend stays; EMOS remains the model of the amount and range shadow. The published forecasts are unchanged.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-14.3 · 2026-10-14
Amount and range shadow; downloads as release assets; new site pages
- New evidence or diagnostics (method unchanged): Shadow rainfall-amount and range forecast (not published): every cycle also runs the EMOS distribution model and records each target's median and 10-90% range next to the MVS-corrected ensemble, scored once CHIRPS is available. 2026 so far (13 targets observed): the EMOS range is closer to calibrated coverage in 9 and has the higher CRPSS in 5; the MVS ensemble is clearly better in CRPS for JJAS June-August. The published forecasts are unchanged.
- New evidence or diagnostics (method unchanged): Probability calibration Phase 3 (Bayesian joint probability, BMA for precipitation, quantile regression forests) is implemented; its benchmark is running and the results follow in the next release.
- Display or explanatory-text change: Large download files (NetCDF, JSON, PDF, CSV) are served from the GitHub release 'site-files', named by their content so links never change; the site is half its former size. Every build also writes the site's data as JSON (site_data/), from which new pages are built: overview, probability calibration, bias correction, downloads and releases, under /next/ (docs/49).
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-14.2 · 2026-10-14
Probability calibration, Phase 2
- New evidence or diagnostics (method unchanged): Phase 2 of the probability and rainfall-distribution calibration, for all four cycles: one-vs-rest binary calibrators (Platt, beta, isotonic) against the incumbent blend, and rainfall-distribution models from the raw ensemble (EMOS with a censored shifted gamma, extended logistic regression, quantile regression) scored for terciles, amounts (CRPS), quantiles and interval coverage. No method is confirmed better than the published probabilities. EMOS (January-initialized FMAM season; JJAS September) and quantile regression (JJAS September) are confirmed better than the MVS-corrected ensemble in CRPS, and the distribution models' 10-90% intervals are close to calibrated where the MVS ensemble's are too narrow. The published forecasts are unchanged.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-14.1 · 2026-10-14
Probability calibration, Phase 1
- New evidence or diagnostics (method unchanged): New section 'Probability and rainfall-distribution calibration': Phase 1 benchmark of six tercile-probability methods (pseudocount smoothing, the incumbent climatology blend, temperature scaling, regularized Dirichlet calibration, multinomial and ordinal logistic regression) with the amount correction fixed, nested year-based validation and a selection-aware procedure, for all four cycles. No method is confirmed better than the incumbent blend; the published forecasts are unchanged.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-13.2 · 2026-10-13
Shadow forecast, suggested methods and sortable tables
- New evidence or diagnostics (method unchanged): Ongoing shadow forecast: square-root MVS for February of the February-initialized FMAM cycle runs automatically next to the published forecast in every such cycle and is scored once observed (never published as a forecast).
- New evidence or diagnostics (method unchanged): Suggested method for every cycle, target and product, from the historical CHIRPS confirmation and the independent validation seasons, by a rule fixed in advance. After 2026: candidates for February FMAM probabilities (square-root MVS), JJAS July-August amounts (VS) and ONDJ January amounts (square-root MVS); keep MVS elsewhere. The published forecasts are unchanged.
- Display or explanatory-text change: Every table on the site can be sorted by clicking a column header (numbers by value, text alphabetically; a third click restores the original order).
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-13.1 · 2026-10-13
Bias correction: ratio-limit study and a shadow pilot
- New evidence or diagnostics (method unchanged): MVS ratio-limit study for all cycles: wider spread-ratio limits do not change the published probabilities (at most 0.002 RPSS) and help rainfall amounts only in some months (JJAS July-September, ONDJ January, FMAM February-March for the January initialization), while lowering the ONDJ season's amount skill.
- New evidence or diagnostics (method unchanged): Shadow pilot of square-root MVS for February of the February-initialized FMAM cycle: reproduces the published February 2026 forecast with MVS, and scores the alternative against February 2026 (one season, an illustration). Not a published forecast; the published forecasts are unchanged.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-12.2 · 2026-10-12
Bias correction: all four cycles
- New evidence or diagnostics (method unchanged): The bias-correction screening and confirmation now cover all cycles (January and February FMAM, May JJAS, September ONDJ), with a cycle picker in the Bias correction section. Published probabilities: MVS is not significantly beaten except for February in the February-initialized cycle, where square-root MVS is confirmed better. Amount gains are month-specific (ONDJ January, FMAM February-March, JJAS June-September). The published forecasts are unchanged.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-12.1 · 2026-10-12
Bias correction: review revision 3 (downloads, reproducibility, edge cases)
- Display or explanatory-text change: Bias correction downloads: the per-cell scores and corrected forecast members (NetCDF) are now published; the site build checks that every advertised download and asset is deployable.
- New evidence or diagnostics (method unchanged): Reproducible reruns: one dependency fingerprint (configurations, masks, inputs, calculation code) validates checkpoints and is recorded as provenance. Fixes: configured smoothing in the confirmation, dry rule after EQM/DQM extrapolation, SDM rank order; results unchanged at the reported precision.
- New evidence or diagnostics (method unchanged): January amounts: the gains of square-root MVS and Gamma QM come almost entirely from cells that are wet in every training year; no cell of Ethiopia has 0 mm in every training year, so an all-zero guard for MVS (evaluated as a variant) never acts. The published forecasts are unchanged.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-11.6 · 2026-10-11
Bias correction: review revision (screening, confirmation, method fixes)
- New evidence or diagnostics (method unchanged): Bias correction section revised after an external review: screening marks distinguish within-method and family-wide Holm adjustment ('screen_pass', not 'adopt'); a new selection-aware confirmation runs every method through the full probability pipeline with the blend weight fitted inside the folds; the test-period shortlist rule was removed.
- New evidence or diagnostics (method unchanged): Method fixes: Scaled Distribution Mapping follows the published steps (checked against an independent implementation); a documented dry-total and tie policy for the quantile-mapping methods; Variance Scaling uses the same fallback as MVS; DQM and QDM descriptions corrected; activation of limits and fallbacks reported.
- Display or explanatory-text change: Ordinary CRPS is the primary amount score (fair CRPS as a diagnostic); coverage, dry shares and run provenance added; the region switch is labelled as applying to the result tables, and the summary, interpretation and figures as national. The published forecasts are unchanged.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-11.5 · 2026-10-11
Bias correction section
- New evidence or diagnostics (method unchanged): New 'Bias correction' section: benchmark of 13 rainfall-amount bias-correction methods against the operational mean-variance scaling for the September-initialized ONDJ cycle (data, method definitions, steps, results by target and region, gate statistics, figures, interpretation, downloads and references; docs/47). The published forecasts are unchanged.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-11.4 · 2026-10-11
Method and references
- Display or explanatory-text change: Methods section: new 'Method and references' panel naming the bias-correction method (per-cell mean–variance scaling), the probability and blending steps, the scores, and the scientific references with DOIs (checked on Crossref). Full text in docs/46. Forecasts and scores are unchanged.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-11.3 · 2026-10-11
Grid-cell map style (0.25°)
- Display or explanatory-text change: Forecast maps can be shown as Smoothed (default) or Grid cells (0.25°): every cell drawn exactly as computed, without interpolation or smoothing, with the same colours, classes and legend. Both styles are in the download package. Forecast values and statistics are unchanged.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-11.2 · 2026-10-11
Smoother maps (display only)
- Display or explanatory-text change: Map display smoothing raised from 0.6 to 1.5 native cells (about 0.38 degrees), which softens the 1-degree steps the forecasts inherit from the ECMWF model grid; rainfall-domain edges are smoothed within the country instead of following the 0.25-degree mask cells. Forecast values, statistics, verification and exported files are unchanged (native 0.25-degree grid).
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-11.1 · 2026-10-11
CHIRPS v3.0 adopted: every cycle recalibrated and re-verified (2026 forecasts retrospective)
- Change to the scientific method: Observations changed from CHIRPS v2.0 (daily p25) to CHIRPS v3.0 (official monthly 0.05 degree product averaged onto the 0.25 degree grid; Funk et al. 2026). The whole chain was refitted with the same method: amount correction, terciles, blend weights, historical evaluation, development runs and decision gates (which give the same outcome as with v2.0).
- New verification observations: The 2026 forecasts of all four cycles were recalibrated with CHIRPS v3.0 after their outcomes were known and are verified against CHIRPS v3.0 (FMAM Feb-May and season, JJAS Jun-Aug). They are retrospective reconstructions; the CHIRPS v2.0 versions are release 2026-10-10.2.
- Rainfall-domain presentation update: The FMAM rainfall contribution domain keeps its rule (FMAM >= 100 mm and >= 20% of annual) on the CHIRPS v3.0 climatology: 75.6% of Ethiopia (72.6% with v2.0). The JJAS R1+R2 and ONDJ R3 regime domains are unchanged; the regimes come from the CHIRPS v2.0 daily climatology.
- New evidence or diagnostics (method unchanged): The v2.0 vs v3.0 comparison (climatology about 15% wetter with v3.0, seasonal shares and year rankings closely agreeing; calibration skill similar overall) is in Methods and docs/45.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-10.2 · 2026-10-10
Climatology and domain reviews for JJAS and ONDJ; historical rainfall-amount skill for FMAM and ONDJ
- New evidence or diagnostics (method unchanged): Rainfall climatology and domain reviews for JJAS (with EMI's Kiremt 2026 long-term maps 2(e) and 3(e), digitization in draft) and ONDJ (no EMI long-term Bega maps published yet). Evidence only: the JJAS R1+R2 and ONDJ R3 domains are unchanged.
- New evidence or diagnostics (method unchanged): Calibration development runs (fit 1993-2016, evaluate 2017 onwards) for both FMAM initializations and ONDJ, so Historical performance now shows rainfall-amount skill for every cycle (all of Ethiopia): FMAM CRPSS -0.001 (January) and +0.053 (February), ONDJ +0.18 (2017-2024), JJAS +0.017.
- Display or explanatory-text change: The Rainfall climatology and domains section shows the selected season's review; comparison notes and map captions name each cycle's own periods and seasons.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization), JJAS 2026, ONDJ 2026/27
Release 2026-10-10.1 · 2026-10-10
FMAM 2026 verified; February initialization; FMAM rainfall contribution domain; outlooks scored against CHIRPS
- New verification observations: February 2026 and the full FMAM 2026 season verified for both initializations. CHIRPS never published the February by_month file; February is taken from the official annual p25 file after checking it is identical to the published by_month files of January and March 2026.
- Rainfall-domain presentation update: Main FMAM area: the FMAM rainfall contribution domain (FMAM >= 100 mm and >= 20% of annual rainfall, CHIRPS 1993-2025), replacing the earlier R2-based domain; the FMAM main-season domain (>= 40% of annual) is a second view. The earlier R2-based verification report is archived as superseded.
- New evidence or diagnostics (method unchanged): The platform, EMI and ICPAC are scored against CHIRPS for each one's own period (single season): favoured-category matches with observation, outlook and favoured-category coverage, and RPSS for the platform and EMI. Observed CHIRPS terciles are shown alongside the three outlooks. Comparison results are computed separately for every domain view.
- Display or explanatory-text change: Verification reports state the cycle's own initialization and season; one current report per cycle. Comparison maps draw the national border once. Historical performance opens with a takeaway; technical detail is expandable.
Cycles: FMAM 2026 (January initialization), FMAM 2026 (February initialization)
Release 2026-10-09.11 · 2026-10-09
Comparison pipeline ready for other seasons
- Display or explanatory-text change: Map titles and table notes use each season's own labels and months; zone tables hide arrow columns for seasons whose EMI zones have no arrows. ONDJ results unchanged.
Cycles: ONDJ 2026/27
Release 2026-10-09.10 · 2026-10-09
Reviewed EMI rainfall regions published
- New evidence or diagnostics (method unchanged): The digitized EMI homogeneous rainfall regions were reviewed by the project team; the site now shows the redrawn region map with EMI's 2026/27 values beside the platform map, region lines on the platform map, and platform means over each whole EMI zone (zone III opposing, VI-VIII agreeing on above normal).
Cycles: ONDJ 2026/27
Release 2026-10-09.9 · 2026-10-09
EMI zones as homogeneous rainfall regions (digitized, pending review)
- New evidence or diagnostics (method unchanged): EMI homogeneous rainfall regions I-VIII digitized from Korecha and Sorteberg (2013, doi:10.1002/2013WR013760) as a reviewable reference layer; once reviewed, the comparison adds platform means over each whole EMI zone and a redrawn region map with EMI's 2026/27 values beside the platform map. Region boundaries drawn on the platform map.
Cycles: ONDJ 2026/27
Release 2026-10-09.8 · 2026-10-09
EMI figure beside the platform map; ICPAC comparison table
- Display or explanatory-text change: The platform map with EMI zone values is shown next to EMI's official ONDJ 2026/27 figure; a new ICPAC table gives ICPAC's printed favoured category and interval against the platform's below/near/above probabilities at the sampled locations and for each area.
Cycles: ONDJ 2026/27
Release 2026-10-09.7 · 2026-10-09
Comparison reports and scope details corrected
- Display or explanatory-text change: Public and packaged comparison reports generated from the published (reviewed, current) data only; report figures resolve inside the ZIP; EMI samples labelled by their overlap with the domain; evidence maps follow each finding's scope; publication check includes the favoured-category threshold and the saved source manifest.
Cycles: ONDJ 2026/27
Release 2026-10-09.6 · 2026-10-09
Clearer, scoped comparison with official outlooks; review and publication safeguards
- Display or explanatory-text change: Comparison follows the selected area (season scope stated for monthly targets); key findings, maps and compact tables first; agreement shares always with their compared area; clickable evidence; comparison downloads and package files; 'Compare outlooks' in the navigation; archived comparison pages repaired.
- New evidence or diagnostics (method unchanged): Bounded sensitivity checks (EMI neighbourhood size and arrow position, ICPAC registration by one pixel); tie-aware favoured category; reviews bound to the extracted contents; publication withheld when the saved comparison no longer matches its inputs.
Cycles: ONDJ 2026/27
Release 2026-10-09.5 · 2026-10-09
Reviewed comparison with official outlooks published
- New evidence or diagnostics (method unchanged): EMI zone values and the digitized ICPAC map were reviewed by the project team; the comparison findings, numbers and maps are now published (agreement in the south and south-east, disagreement in EMI zone III).
Cycles: ONDJ 2026/27
Release 2026-10-09.4 · 2026-10-09
Comparison with official outlooks (ICPAC OND 2026 update, EMI Bega 2026/27)
- New evidence or diagnostics (method unchanged): New pipeline: ICPAC and EMI products discovered, downloaded and versioned by content hash; standardized records; comparison with the native ONDJ forecast; rule-based interpretation. Comparison numbers are published only after a person reviews the extracted values.
- Display or explanatory-text change: Outlooks section shows the official figures, the providers' own text, source records and the status of each comparison; ONDJ gallery links the review report.
Cycles: ONDJ 2026/27
Release 2026-10-08.3 · 2026-10-08
Historical verification explorer, forecast-and-evidence package, map comparison and release archive
- New evidence or diagnostics (method unchanged): Historical diagnostics per cycle, target and area: performance by year, reliability diagrams, probability histograms and category Brier skill (cross-validated and exploratory).
- Display or explanatory-text change: PDF bulletin and one-click ZIP package per cycle; side-by-side forecast vs observed anomaly maps; release archive and change log; research label on phones; keyboard focus kept when leaving an unavailable view.
Cycles: FMAM 2026, JJAS 2026, ONDJ 2026/27
Release 2026-10-08.2 · 2026-10-08
Area-specific historical skill and clearer uncertainty
- New evidence or diagnostics (method unchanged): Historical skill computed for each cycle's rainfall domain and the R0-R3 regions from stored cross-validated probabilities (no refitting).
- Display or explanatory-text change: Skill labels reflect intervals and Holm-adjusted p-values; verification narrative always reports the rainfall error; climatology defined as observed tercile frequencies; unavailable verification explained; separate verification panels.
Cycles: FMAM 2026, JJAS 2026, ONDJ 2026/27
Release 2026-10-08.1 · 2026-10-08
Redesign around the reader: one cycle selector, outlook first, verification narrative
- Display or explanatory-text change: Cycle and area selectors drive every section; shareable view links; outlook and map first; status panel with precise target statuses; verification narrative and paired anomaly chart; per-cycle CSV downloads.
Cycles: FMAM 2026, JJAS 2026, ONDJ 2026/27
Release 2026-10-08.0 · 2026-10-08
Results site before the review-driven redesign
- New verification observations: FMAM 2026 March, April and May verified against official CHIRPS v2.0 observations.
- Rainfall-domain presentation update: ONDJ shown on the ONDJ R3 (Deyr) rainfall domain (575 cells) instead of the earlier >=120 mm / 20% rule.
Cycles: FMAM 2026, JJAS 2026, ONDJ 2026/27