Glossary
The scores, methods and terms used on the site. On every page, the first use of a term shows its explanation when you point at it (or reach it with the keyboard).
- Benjamini–Hochberg
- Benjamini–Hochberg correction for testing many methods or targets at once: controls the expected share of false findings among those confirmed (less strict than Holm). The method selection uses it since 17 October 2026.
- BSS
- Brier skill score of the probability of one category (e.g. below normal) against climatology. Positive is better than climatology.
- CHIRPS
- Climate Hazards Center InfraRed Precipitation with Stations: the gridded observed rainfall (0.05°, averaged to 0.25°) used for calibration and verification.
- climatology
- The reference forecast: the observed frequency of each tercile in the training years (about one-third each), or the past observed totals for amounts.
- climatology blend
- The final probabilities mix the model probabilities with climatology: p = (1 − λ)·p_model + λ·p_climatology, one weight λ per target fitted on past years.
- coverage
- Share of observations inside the forecast 10–90% range; 80% is calibrated, less means the range is too narrow.
- cross-validated
- Each year is forecast with parameters fitted without that year (leave-one-year-out), so the score is not inflated by fitting to the year itself.
- CRPS
- Continuous ranked probability score of a rainfall-amount forecast (ensemble or distribution) against the observed total, in mm. Lower is better.
- CRPSS
- Continuous ranked probability skill score for rainfall amounts: 1 − CRPS of the forecast / CRPS of climatology. Positive is better than climatology.
- EMOS
- Ensemble model output statistics: a regression that turns the raw ensemble into a full rainfall distribution (here a censored shifted gamma).
- exploratory
- Scores of years that were seen while the method was chosen (2017 onward); useful context, but weaker evidence than the cross-validated 1993–2016 scores.
- Holm
- Holm correction for testing many methods or targets at once: p-values are adjusted so that a confirmation is not a chance result of the number of tests.
- leave-one-year-out
- Each year is left out in turn; everything is fitted on the other years and the left-out year is scored.
- Mean-Variance Scaling
- MVS, the operational amount correction: each member is shifted and scaled so the model mean and spread match CHIRPS in every grid cell (spread ratio limited to 0.5–2).
- MVS
- Mean-Variance Scaling, the operational amount correction: each member is shifted and scaled so the model mean and spread match CHIRPS in every grid cell.
- pinball loss
- The score of one forecast quantile (e.g. the median) against the observation; lower is better.
- rainfall domain
- The area where the season matters climatologically (e.g. FMAM ≥ 100 mm and ≥ 20% of annual rainfall); used for summaries and maps, not for calibration.
- reliability
- Whether events forecast with probability p happen p of the time; a reliability diagram plots observed frequency against forecast probability.
- RPS
- Ranked probability score of tercile probabilities against the observed category (cumulative, so near misses cost less). Lower is better.
- RPSS
- Ranked probability skill score: 1 − RPS of the forecast / RPS of climatology. Positive is better than climatology; +0.05 means a 5% lower score.
- SEAS5
- ECMWF seasonal forecast system 5 (system 51): 25 members in the 1993–2016 hindcasts, 51 from 2017.
- shadow forecast
- A forecast made every cycle with an alternative method and scored against observations, but never published; it collects independent evidence.
- spread–error
- Ratio of the ensemble spread to the error of the ensemble mean; 1 means the spread matches the error, below 1 too narrow.
- tercile
- One of three equally likely categories of seasonal rainfall in the reference period: below normal, near normal, above normal.