Skip to content

Glossary

Every term the documentation uses, one line each, with a link to the page that develops it.

The project's naming discipline is deliberate: several terms have an "avoid" note because an alternative word is ambiguous in this domain, and consistency here is what keeps the code and the docs describing the same thing.

A

Abstention
A method declining to produce a forecast — returning NaN or degrading to a named base — because it lacks the data to fit. Recorded as fit_status, not hidden. → Notation §6
Anchoring
Short-lead correction of a blend toward the latest observed residual, decaying exponentially with lead. Avoid: nowcasting — the minutely product is a nowcast; the technique is anchoring. → Combination §7
Anytime-valid
A test whose error guarantee holds no matter how often or when you look at it. Necessary because this leaderboard is regenerated nightly. → Model selection §2
Availability mask
The per-row pattern of which sources have a usable forecast. Every blender renormalizes its weights over it. → Combination §1

B

Bias
Mean signed error — being wrong in a consistent direction. Reported as its own column because, unlike error generally, it is correctable. → Concepts §4
Blender
Any method implementing fit(train) / predict(matrix) -> point (+quantiles), baselines included. Avoid: model — overloaded with provider models.
Blending
Combining grounded sources into one forecast via weights or a learned stacker. Avoid: ensembling, averaging as method names. → Combination
BOA
Bernstein Online Aggregation — an online expert algorithm with a second-order update. → Combination §5
Brier score
Mean squared error of a probability forecast against the binary outcome. The proper score for PoP. → Verification §2

C

Calibration
A probabilistic forecast whose stated probabilities match observed frequencies. In this project the word is reserved for probabilistic calibration — correcting a point forecast toward the station is called grounding. → Calibration
Conformal prediction
Interval construction with finite-sample coverage guarantees requiring only exchangeability, not a correct model. → Uncertainty
Coverage
The fraction of truths falling inside a stated interval; compared against its nominal level.
CRPS
Continuous Ranked Probability Score — the proper scoring rule for a full predictive distribution; reduces to absolute error for a point forecast. → Verification §2
CSGD
Censored Shifted Gamma Distribution — the precipitation distribution whose censored mass at zero is the probability of a dry hour. → Calibration §2

D

Degraded forecast
A forecast emitted without compatible promoted evidence. Uses fit-free equal weight, sets status = "degraded", and records why. A valid product with weaker guarantees, not an implicit trained model. → FAQ
Diebold–Mariano (DM)
Test of equal expected forecast loss on paired loss differentials. A test, not a decision procedure. → Verification §4
Dressing
Attaching quantiles to a point forecaster from the empirical distribution of its own residuals. → Calibration §7

E

e-BH
The e-value analogue of Benjamini–Hochberg; controls false discovery rate under arbitrary dependence. → Model selection §3
e-process / e-value
A nonnegative process with expectation ≤ 1 under the null; accumulated wealth is evidence against it, and Ville's inequality makes reading it at any time valid. → Model selection §2
EMOS
Ensemble Model Output Statistics — a Gaussian predictive distribution whose mean and spread are regressed on the ensemble. → Calibration §1
Evaluation run
One immutable production of score rows, identified by dataset fingerprint, source set and kind, product, window, truth semantics, method set, code version, and config fingerprint.
EWA
Exponentially Weighted Average forecaster — the classic multiplicative-weights online expert algorithm. → Combination §5

F

Fixed share
Redistributing a small fraction of expert weight uniformly each round, so a recovering source can climb back. Not decoration — without it the aggregators lock onto the wrong expert. → Combination §5
Fingerprint
A content hash identifying a dataset, configuration, or code version. Serving refuses evidence whose fingerprints do not match the live system.
fit_status
A fitted head's self-report: unfitted, insufficient_rows, insufficient_wet_rows, gaussian_fallback, converged, fit.

G

Grounding
Per-source correction toward the station, fitted per variable × lead bucket. A bias correction by default; the slope is opt-in. Avoid: calibration (reserved), MOS (in code). → Grounding

I

IDR
Isotonic Distributional Regression — a nonparametric predictive distribution assuming only that outcomes are stochastically increasing in the forecast. → Calibration §3
Issue time
In docs: when a forecast was made — the information boundary. In this project's vocabulary, distinguish source retrieved at (the provider's fetched_at), source available at (the collector run's completed_at), and forecast issued at (when this system emits a product).

K

\(k_{\text{eff}}\)
Effective number of independent sources, \(k/(1 + (k-1)\rho)\). Eight providers measured 1.8 here. → Combination §2

L

Leakage
Any path by which information unavailable at issue time reaches a model. Assumed present until proven absent. → Verification §6
Lead
Distance from forecast-issued time to valid time. Always recomputed from timestamps, never read from an upstream horizon_hours column. Avoid: horizon (in code).
Lead bucket
One of the fixed left-closed lead intervals used to stratify fitting and evaluation. → Notation §2

M

MAE
Mean Absolute Error — the promotion metric. Minimized by the conditional median.
MCS
Model Confidence Set — the set of methods that cannot be ruled out as best, at a stated level. → Model selection §1
Model release
A promoted mapping from product × variable × lead bucket to a method, tied to compatible live evidence and a training cutoff. Serving consumes releases, not leaderboards. → ADR 0005

O

Observation
One raw station sample (~1/minute), imperial units, unvalidated. Becomes truth only after QC and normalization.

P

PIT
Probability Integral Transform — the forecast CDF evaluated at the observed value. Uniform if calibrated; its histogram names the failure mode. → Verification §2
Pinball loss
The proper scoring rule for a single quantile level.
PoP
Probability of Precipitation — here, the probability that hourly accumulation reaches 0.254 mm. → Concepts §5
Product
An emitted forecast bundle with its own temporal contract: minutely (next 60 minutes), hourly (next 48 hours), daily (next 10 local days).
Provenance wall
The rule that live and backfilled (synthetic) rows are never pooled, enforced at the filesystem level. → Verification §7
Provider
An upstream forecast API, by slug (e.g. open_meteo).

Q

QC
Quality control. On the station: bounds, spike, and flatline filters producing a per-minute bitmask. A flagged sample becomes null, never corrected. → Truth and QC

R

Rolling origin
The backtest protocol: repeatedly train on what was knowable at an origin and test on what was issued just after it. → Verification §5

S

Self-verification
Scoring the system's own emitted products against later truth, alongside providers and backtest expectations. → Verification §8
Skill
Relative accuracy against a named reference: \(1 - \text{MAE}_{\text{method}}/\text{MAE}_{\text{ref}}\). → Verification §3
SNHT
Standard Normal Homogeneity Test — the change-point statistic used for station drift, most sensitive near the series end. → Truth and QC §5
Snapshot
The as-of view of all sources at one issue time: each source's latest forecast at or before that moment, within the staleness cap. Avoid: vintage.
Sleeping expert
A source outside its horizon: absent from the round, neither updated nor penalized. Why ragged provider horizons need no special casing.
Source
One provider+model forecast stream — the unit a blender weighs. Avoid: expert (except in the online-experts method), member.
Source kind
Whether rows came from the live archive (live) or a backfill (synthetic). Never pooled silently.
Spread
Ensemble standard deviation, used as a predictor of forecast uncertainty. → Uncertainty §3

T

Truth
The QC'd, metric-normalized, time-aggregated observation series used for training and scoring. Either present and trusted, or null — never imputed.
truth_known_at
When a row's truth became knowable. Training uses this, not issue time — the distinction is the core leakage defence. → Verification §5
Truth semantics
Which aggregation an hourly value means: instantaneous (inst, ±5-min centered mean) or interval mean (mean). Measured per provider, not assumed. → ADR 0003

V

Valid time
The moment or interval a forecast is about. Avoid: target time, forecast time.
Ville's inequality
The result making e-processes anytime-valid: a nonnegative martingale exceeds \(1/\alpha\) with probability at most \(\alpha\).

W

Winner's curse
The selected method's reported score is optimistically biased, because selection favours methods that got lucky. Corrected explicitly. → Model selection §4