The model's FPL tools.
Player projections, captain shortlist, optimal XI, fixture ticker — model estimates updated before each deadline. All projections are labelled as such: the model has uncertainty and will be wrong sometimes. That's why every decision is committed to the public repo before you can see it.
The squad publishes when the model locks, before each deadline — committed to the public repo. The git timestamp is the proof it was not changed afterwards.
The model's gameweek scores appear here as they are graded — every squad committed before the deadline, every score published afterwards, including the bad ones.
The model plays by your rules — one squad, real budget, real transfers, hits paid. See how you did against it.
The same Dixon-Coles model, fitted on Champions League results. Half of this season's field is newly qualified and has no European record to learn from, so every fixture carries how much the model actually knows about it.
The same Dixon-Coles model — the same commitment to transparent, pre-deadline forecasts — now runs for FPL. Here's what it did when every prediction was already on the record:
Built on a proven forecasting engine.
We tested three models on GWs 29–34 of 2025/26 — weeks no model trained on. The Dixon-Coles projection won. The ML model performed no better than a naive baseline on an honest out-of-sample test, so we didn't ship it. The same engine ran all 104 WC2026 fixtures and now runs FPL 2026/27.
Built on football analytics.
TheModelSays uses a Dixon-Coles Poisson model — the same statistical framework used in academic football research and professional betting markets.
Dixon-Coles Poisson model
At the core is a Dixon-Coles model — the statistical standard for football prediction. It estimates each team's attack strength and defensive weakness from weighted historical fixtures. A low-score tau correction improves accuracy on 0-0, 1-0, 0-1, and 1-1 scorelines, which a naïve Poisson model under-predicts.
FPL-specific feature engineering
Player rates (goals, assists, clean sheets, xG, xA) are computed from rolling history. A cold-start path bootstraps from the prior season's archive when live history is thin — no player projects 0.0 in GW1. Positional priors handle promoted-side players and new signings.
50,000 Monte Carlo simulations (WC) / per-fixture matrices (FPL)
The WC tournament simulation ran 50,000 bracket paths. For FPL, the DC matrix runs per fixture: scoreline probabilities convert to expected goals, assists, and clean-sheet rates per position — no black-box ML, every number traceable to the model.
Walk-forward retraining — tested and dropped
We retrained after each WC matchday, weighting 2026 results at 4×. It made things worse: the frozen pre-tournament model called 66.3% of outcomes correctly vs the retrained version's 59.6%. Walk-forward retraining is dropped. The same decision logic applies to FPL: refit only on sufficient held-out evidence.
TheModelSays is an independent, free-to-use football prediction tool. The WC2026 model is complete — 104 matches, all on the record at themodelsays.com/wc2026-record. FPL 2026/27 is live. Not affiliated with FIFA or the Premier League. All predictions are generated by a statistical model for informational and entertainment purposes only. Follow on TikTok and YouTube.
TheModelSaysDixon-Coles predictions for FPL 2026/27. Every squad committed before the deadline. Every score graded in public.