Mathematical Methodology
The scoring model is a pair of independent Poisson distributions over home and away goals, multiplied by the Dixon-Coles correction term (ρ) that redistributes probability across the four lowest scorelines. For the five leagues with an expected-goals record deep enough to fit on, ρ and the home-field scale are estimated from historical results; elsewhere the model uses a shared default. Home and away expected-goals rates (λhome, λaway) come from attack/defence strength ratings built on a rolling, recency-weighted window — from expected goals where those records exist, from goals scored and conceded where they do not. The λ pair defines a full scoreline probability grid, from which every published market (1X2, over/under, BTTS, correct score, double chance, handicap) is derived directly — no market is estimated independently of the others.
Where bookmaker odds exist for a market, the de-vig'd market-implied probability is blended with the model probability (roughly 20–30% model weight at launch — a market-dominant blend); this acknowledges that markets carry real information the model doesn't see, and treats the model as a corrective lens on the market rather than the other way round.
Full backtest methodology and results: accuracy.