The ensemble
A Dixon-Coles bivariate Poisson model is fitted separately for each division — maximum likelihood with analytic gradients, exponential time decay (ξ = 0.0015 per day), the low-score τ correction, and ridge-regularised attack and defence. A sequential Elo system with goal-difference weighting and season regression captures cross-division strength that a per-league Poisson cannot. A gradient booster adds rolling goals, shots on target, form, rest days and head-to-head.
Their average is the model ensemble. An Elo-to-1X2 calibration map keeps the rating component honest, and every input is computed as of the publish timestamp — no leakage, tested rather than assumed.
The market is a feature
Closing prices are the strongest single predictor in football. We de-vig them with the power method and blend them with the ensemble. The weight is learned per division by minimising RPS on a held-out season.
Across all 17 divisions the optimisation returned 0.80–1.00 — “trust the price almost everywhere”. That matches the ten-season backtest, where the optimal weight landed at 0.9–1.0 in every tunable season. We cap it at 0.85 because the anchor is learned on closing prices while this site publishes days earlier, when the price is softer. That 15% tilt is the entire methodological claim.
Closing-line value
For any forecast where the model found a positive edge at publication, we keep the price we published at and compare it with the price at kick-off. If our number was right, the market should move toward us and the closing price should be shorter. The average of log(publish / close) is CLV — a far lower-variance signal than profit and the metric professionals actually use.
The ledger
Each forecast is written once with the previous row's hash folded into its own:
row_hash = sha256(prev_hash + "|" + canonical_json(forecast))
There is no update path and no delete path. Daily Merkle roots are published so a third party can verify, years later, that a given probability existed at a given moment. That is the difference between “our model is good, trust us” and a record anyone can check.
Scoring
The primary metric is the Ranked Probability Score, which rewards calibrated distributions and penalises both overconfidence and hedging. Accuracy is reported next to it, but accuracy treats a near miss and a wild miss as identical — the shortcut that makes prediction sites look better than they are.
Limitations
- Public data only — no licensed xG, lineups or injury feeds in this build.
- The replay is walk-forward by season, not by match week.
- Closing-line value is computed only where both the publish price and the close are known.
- We reach statistical parity with the closing line, not dominance. Anyone claiming otherwise is selling something.
The API
Every forecast is JSON at /api/v1/forecasts; the ledger verifies at /api/v1/ledger/verify; the record is at /api/v1/track-record. These pages are generated from the same store, so there is one source of truth.