How good is it,
really?

The live model is the Hybrid Ensemble (Dixon-Coles goal expectation blended with a gradient-boosted outcome classifier). The classifier is stronger in isolation on one season, the blend is stronger on the other two — which is why the site serves the blend.

• 2023/24

  0.5828

• 2024/25

  0.6088

• 2025/26

  0.6286

Being right often is not the same as being right by the correct amount. A forecast that says 60% should happen about 60% of the time — the Brier score measures that, and it is the metric a model has to improve before it is allowed to ship. Latest full-season log loss: 1.0381; draw calibration gap: 0.0039.

• Q1 · MD 1–9

  53.3%

• Q2 · MD 10–19

  55.0%

• Q3 · MD 20–28

  40.0%

• Q4 · MD 29–38

  42.0%

The model is materially better in the first half of a season than the second. The leading explanation is that mid-season form pivots — a manager change, a January rebuild, a genuine hot streak — move faster than the recency window the ratings use. That gap is an open investigation, not a solved problem.

Decision on 2026-05-27: not shipped. A newer fit is not automatically a better one, and the gate exists to catch exactly this.

1. 1. Ratings, not vibes

   Each club carries an Elo rating plus separate attack and defence strengths, rebuilt from every completed fixture and weighted towards recent form. Home advantage is fitted, not assumed.

2. 2. Two models, blended

   A Dixon-Coles Poisson model turns those ratings into goal expectations and a full scoreline distribution. A gradient-boosted classifier predicts the outcome directly from the same features. The live forecast is a blend of the two — each covers the other’s weak spots.

3. 3. Context, carefully

   Season phase, table pressure, fixture congestion and squad availability adjust the baseline. These features are deliberately weighted low: they help in-season and hurt on seasons they were not calibrated against, so they are never allowed to dominate the ratings.

4. 4. Simulated seasons

   Title, top-four and relegation probabilities come from a 10,000-iteration Monte Carlo sweep of every remaining fixture. The sweep runs offline on a schedule; pages only ever read the resulting snapshot.

5. 5. Replayed and gated

   Every candidate model is replayed out-of-sample against completed seasons. It ships only if it improves Brier score and log loss without a season quarter regressing. Candidates that fail are documented and discarded.

• These are out-of-sample replay backtests, not a live frozen-prediction log. The World Cup 2026 log is the fully frozen record — every call written to disk before kickoff and never overwritten.
• Brier score has drifted upward season over season (0.583 → 0.609 → 0.629). Premier League results are getting harder to predict, or the model has drifted from the data distribution. This is under investigation at the 2026/27 season start.
• On historical seasons the plain baseline beats the fully-enriched context model. The context layer (pressure, congestion, manager, form weights) adds net noise against seasons it was not calibrated on; the hybrid blend recovers most of that, but the context layer needs regularisation.
• A run-in fatigue feature was tested and rejected — it made the model worse in both 2023/24 and 2024/25. It is not enabled.
• Squad-sentiment weighting cannot be validated before 2026-04-28, when the sentiment archive starts. It runs at deliberately conservative weights pending a full-season ablation.
• Exact scorelines are much harder than outcomes and are evaluated separately — treat any single projected score as the middle of a wide distribution, not a call.

This is a forecasting tool, not a tipping service. Every number here is a model output — a statement about how often something happens in simulation, not advice about what to do with it.