We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Prediction methods
Every method implements one walk-forward behaviour, so new approaches drop in without touching the backtest engine. This is the "naive" baseline set; advanced models come later.
Base Rate
base_rate
Predicts the historical home/draw/away frequency, split by neutral vs non-neutral venue.
Elo
elo
World-football Elo: home-field advantage, goal-difference-weighted updates, and match-importance-weighted K (World Cup 60 → friendly 20).
Poisson
poisson
Independent-Poisson goals model from running team attack/defense strengths; predicts scorelines.
Dixon-Coles
dixon_coles
Poisson goals model with the Dixon-Coles low-score correction (better draws) and recency time-decay.
Bivariate Poisson (diagonal-inflated)
bivariate_poisson
Dixon-Coles goals model with a diagonal inflation lifting all drawn scorelines — the bivariate-Poisson draw correlation as a tunable.
Calibrated Elo
calibrated_elo
Elo forecasts remapped through online histogram-binning calibration learned walk-forward from past matches — better-calibrated probabilities (Brier/log-loss) and fairer odds.
Squad Value
squad_value
Transfermarkt squad value (Peeters 2018): log(home/away top-23 value) → logistic win probability with an Elo-style draw model. Falls back to base rates where a team has no valuation.
Elo + Value
elo_value
Online multinomial-logistic blend of Elo rating and Transfermarkt squad value (Peeters 2018), fitting the value-vs-Elo weighting walk-forward. Warm-started to Elo so it never starts worse.
Elo + Value + Venue
elo_venue
EloValue plus venue effects: per-team travel/jetlag (host-country streak + arrival timezone jump) and altitude gap (venue elevation above each team's home base — matters on neutral high-altitude venues like Mexico City), fitted walk-forward.
Elo + Value + Importance
elo_value_importance
Elo + squad-value blend extended with game-importance interactions (trust the favourite less in dead rubbers, more in high-stakes games) and a dead-rubber draw bias. Warm-started to Elo+Value so it never starts worse.
Calibrated Elo + Venue
calibrated_elo_venue
EloVenue (Elo + squad value + travel/altitude) remapped through online histogram-binning calibration — the richest model with corrected confidence (Brier/log-loss) and fairer odds.
Ensemble (Venue + Dixon-Coles)
ensemble_venue_dc
Equal-weight linear opinion pool of the EloVenue blend (strength: ratings, squad value, travel/altitude) and the Dixon-Coles goals model (scoreline structure) — two families pooled for steadier calibration.