Six models,
one consensus
Each model sees the match from a different angle. The ensemble combines them with weights learned by the auto-calibration loop.
Poisson
Models each team's goals as Poisson processes derived from their attack and defense strength. The backbone of the scoreline matrix.
Dixon-Coles
Adjusts Poisson on low scorelines (0-0, 1-0, 0-1, 1-1), where independence breaks down, and applies time decay to weight recent form more.
Dynamic Elo
A per-team rating that rises or falls after each result, factoring in home advantage. Captures form and hierarchy continuously.
Bayesian
A Gamma-Poisson model with a conjugate prior that quantifies uncertainty when data is scarce (teams with few matches, early season).
Market consensus
De-vigs odds from multiple bookmakers → the "fair" implied probability. The market aggregates information no single model has.
XGBoost
900 decision trees trained on 124,000+ real historical matches. Captures non-linear interactions that analytical models miss.
How they combine
It's not a simple average. Each model contributes a probability distribution over scorelines; the ensemble weights them by how well each has calibrated on recent history. Then NEMESIS audits the result and recalibrates confidence before showing it to you.