January initialization · FMAM and its months
Can a different method improve the tercile probabilities, or give better rainfall amounts and ranges, than the published pipeline (MVS-corrected members → smoothed counts → climatology blend)? Every parameter is fitted only on training years within each left-out year; the published forecasts are unchanged.
| Phase | Methods | Compared with | Confirmed better |
|---|---|---|---|
| Phase 1 · tercile probabilities | Pseudocount smoothing (no blend), Temperature scaling, Regularized Dirichlet calibration, Regularized multinomial logistic regression, Ordinal (cumulative-logit) logistic regression | Published probabilities (climatology blend), RPS | none |
| Phase 2 · binary calibrators | Platt (sigmoid) calibration, one-vs-rest + normalization, Beta calibration, one-vs-rest + normalization, Isotonic regression, one-vs-rest + normalization | Published probabilities, RPS | none |
| Phase 2 · distribution models, terciles | EMOS, censored shifted gamma, Heteroscedastic extended logistic regression, Regularized quantile regression | Published probabilities, RPS | none |
| Phase 2 · distribution models, amounts | EMOS, censored shifted gamma, Heteroscedastic extended logistic regression, Regularized quantile regression | MVS-corrected ensemble, CRPS | EMOS, censored shifted gamma (FMAM) |
| Phase 3 · distribution models, terciles | Bayesian joint probability, Bayesian model averaging for precipitation, Quantile regression forest | Published probabilities, RPS | none |
| Phase 3 · distribution models, amounts | Bayesian joint probability, Bayesian model averaging for precipitation, Quantile regression forest | MVS-corrected ensemble, CRPS | none |
Amount and range shadow (EMOS, not published): 2026: 5 of 5 targets observed. Scores and maps →
Explore the evidence
- Phase 1
Six tercile-probability methods with the amount correction fixed, and the selection-aware procedure. - Phase 2
One-vs-rest binary calibrators, and the first rainfall-distribution models (EMOS, extended logistic, quantile regression). - Phase 3
Bayesian joint probability, BMA for precipitation and quantile regression forests. - Amount and range shadow
EMOS medians and 10–90% ranges for the forecast year, never published, scored season by season. - Downloads
Every result file and the full write-up.