Behind the forecast
The model.
The numbers.
The uncertainty.
A guide to reading the weekly previews and understanding what the current goals model can tell you.
01 / From results to expected goals
A goals-based starting point.
The current model uses completed Premier League results from 2024/25, 2025/26 and the available 2026/27 matches. The first run used 816 completed matches, including the six played on 10 October 2026.
It estimates attack and defence strengths, gives more recent results greater influence and accounts for home advantage. Where team evidence is limited, estimates are pulled towards the league baseline rather than treating a small sample as conclusive.
Those estimates produce expected goals for each side. A Poisson score distribution then gives probabilities for possible scorelines. The displayed match outcomes and goal probabilities are calculated from that same score distribution.
Results source: OpenFootball’s CC0 dataset. The latest results in the initial run were checked separately.
02 / The three outcomes
A favourite can still lose.
Home win, draw and away win probabilities describe how the model distributes the possible outcomes. They are not confidence guarantees. A side can have the largest single share without having a greater-than-even chance of winning.
Illustration only / Not a match forecast
Here the home side has the largest share. The draw and away-win possibilities together still account for 55%. “Most likely outcome” therefore does not mean “likely to happen with certainty”.
Read the three outcomes together, alongside the snapshot time. Rounding can make displayed percentages differ very slightly from a total of 100%.
03 / One score versus many outcomes
Why a score and an outcome
can point different ways.
The most likely exact score is just one possibility. A team’s overall win probability adds together every winning scoreline. All draws are added together for the draw probability.
For example, 1–1 may be the most likely individual score while the combined probability of all home wins is larger than the combined probability of all draws. Both statements can be true.
A scoreline shown with a 12% probability still leaves 88% for other scores. Treat the exact-score preview as a compact summary of the distribution, alongside the broader match outcomes.
What about goal probabilities?
The wider output includes both-teams-to-score and over-2.5-goals probabilities. These are also derived from the score distribution, so they are connected to the same expected-goals estimates rather than a separate promise about the match.
04 / Two different clocks
The team pool rotates.
The forecasts are snapshots.
The free pool changes each Monday in UK time. That selects which teams are visible; it does not automatically rerun the model.
Each forecast records when it was generated. A card is hidden after its scheduled kickoff until a new next-match preview is published. Fixture schedules can change, so check the current kickoff when planning to follow a game.
If you would like a run at a different point before kickoff, email Jude with the fixtures and timing you have in mind. This preview currently shows pre-match snapshots.
05 / Read the evidence fairly
A baseline that needs
proper evaluation.
The current model uses goals. It does not incorporate xG, player availability, confirmed lineups, bookmaker prices or a calibrated ensemble of several models.
The initial settings have not yet been tested against a reserved evaluation set. The 816 historical matches were used for training; they are not an audited record of 816 predictions made before those matches.
A useful next evaluation would keep later matches out of training, record forecasts before kickoff and assess how well the probabilities match subsequent outcomes. Model changes and forecast times would need to be kept in that record.
No demonstrated accuracy advantage or profitable betting edge is claimed. You can use the preview to explore the research and ask Jude about its development.