Skip to content

Methods

Rigorous specifications of every estimator, score, and decision procedure in the system, as implemented — with the constants, the guards, and the places where the code knowingly departs from the textbook.

Theory and concepts is the narrative companion: it owns why and which. These pages own what exactly. No formula appears in both.

If you want the plain-language version first, read Concepts; if a term is unfamiliar, the Glossary has it.

  • Notation — symbols, indexing, the forecast object, column naming as a type system, and why abstention is a first-class outcome. Start here.
  • Grounding — affine correction toward the station, regression dilution, slope shrinkage, empirical-Bayes pooling across lead buckets.
  • Combination — the availability algebra, the \(k_{\text{eff}}\) ceiling, every weighting scheme, GBM stacking, online expert aggregation, anchoring, and the minutely path constructions.
  • Calibration — EMOS, CSGD, isotonic distributional regression, PoP calibration, PIT/CQR recalibration, quantile dressing.
  • Uncertainty — adaptive conformal prediction, distributional conformal coverage, spread–skill, cross-variable coherence.
  • Verification — proper scoring rules, Diebold–Mariano with HAC and HLN, the rolling-origin protocol, four leakage defences, the provenance wall.
  • Model selection — Model Confidence Set, betting e-processes, BH and e-BH, the winner's curse, and the promotion gates.
  • Precipitation — mixed discrete–continuous handling, reset-aware accumulation, cross-source QC guards, sparse shrinkage.
  • Truth and QC — station quality control, the aggregation ladder, neighbour cross-checks, SNHT/Pettitt change points, the radiation-shield error model.
  • Bibliography — every reference, matched to its module, plus the papers deliberately declined and why.

Three properties worth knowing before reading any of it

Nothing here is chosen by argument. Every method on these pages is registered, scored on the same rolling-origin folds, and promoted or not by the machinery in model selection. Where two approaches compete, both are usually registered and the leaderboard arbitrates. This documentation explains what the verdict means; it does not substitute for it.

Abstention is designed for. A method that cannot fit returns NaN or degrades to a named base, and records why. Thin slices show missing methods rather than noisy ones.

The evidence is dated and local. Every measured number in this documentation comes from one station in Crestline, CA, over a specific window. Read Limitations before generalizing any of it.