How the KrystalPredict Football Prediction Model Works

Model inputs

The codebase trains on historical match results. Team strength is represented through Elo ratings, while the goal model estimates score distributions. Competition services only use data available before the match when constructing their historical evaluation path.

TODO: connect this explanation to the exact production model inputs when the production data pipeline publishes a versioned input specification.

Probability calibration and fair odds

A 60% model output means the model allocates 0.60 of its probability mass to that outcome under its assumptions; it is not a certainty. Fair odds are simply the reciprocal of a valid probability and exclude spreads, commissions, limits, and market movement.

Kelly sizing and odds comparison

Kelly calculations are kept separate from probability generation. They require both a probability estimate and an odds price. The tool returns zero where the formula is non-positive and presents fractional variants to illustrate lower-volatility sizing. This is educational mathematics, not advice to take an action.

Backtesting and limitations

Historical evaluation can report sample size, date range, log loss, calibration, and strategy results only where the backend provides the underlying records. It cannot establish future performance. The model does not know future line-ups, injuries, tactical decisions, weather, officiating events, or data-feed errors.

Responsible use

Review uncertainty, check the data timestamp, never chase losses, and do not risk more than you can afford to lose. KrystalPredict does not accept bets or guarantee outcomes. 18+ only and local laws apply.