Full football match payload (probabilities + distributions/CDFs)

Three-way result + over/under + both-teams-to-score outcomes.

NOTE — model independence: home_win/draw/away_win are the output of the 1X2 softmax classifier; over_2_5/under_2_5 are the output of an independent over/under binary classifier; btts_yes/ btts_no are the output of a third independent classifier. The three families do NOT share a joint likelihood: a high P(home_win) need not imply a correspondingly high P(over_2_5), and the 1X2 marginal need not equal the sum of any exact-score family. Apparent inconsistencies between these probabilities on the same match are EXPECTED BY DESIGN. To assess the reliability of any single probability call GET /v1/calibration/football/<model_id> (or its MCP equivalent get_model_calibration). Discover model ids via GET /v1/models.

Probability mass functions for count-style targets.

Probability mass functions for count-style targets.

Concept-level attribution of the winner prediction.

Computed with gradient-based attribution on the production neural network and aggregated into human concepts; the factors approximately sum to the gap between the served win probability and an even matchup. For football the factors compare the two sides directly (the draw probability is modelled separately). Absent on rows predicted before the explainability rollout (June 2026).

Concept-level attribution of the winner prediction.

Computed with gradient-based attribution on the production neural network and aggregated into human concepts; the factors approximately sum to the gap between the served win probability and an even matchup. For football the factors compare the two sides directly (the draw probability is modelled separately). Absent on rows predicted before the explainability rollout (June 2026).

Realized result once the match has been played and reconciled (the per-league football_in_season_performance* table). Absent until the match is played.