Methodology

Separate facts, features, and forecasts.

The site treats official records as source data, derived indicators as features, and predictions as model outputs. Each layer has its own provenance and review rules.

Publication rules

Candidate and result facts must link to a primary or named secondary source. Model features must describe their source data and transformation. Forecast outputs must include uncertainty, a timestamp, and the model version that produced them.

Initial model boundary

The first usable model should be deliberately narrow: constituency-level contest probabilities, vote-share intervals, and national seat aggregation. Local, devolved, mayoral, and by-election models can reuse the same schema once the source history is stable.

Backtesting

Backtests are part of the product, not an internal notebook. The public record should show calibration, error distribution, biggest misses, and whether model changes improve the archived result.