1. The unit of analysis is a fixture
Exact Predictors estimates a specific home team against a specific away team in a named competition and season. Home and away scoring rates remain separate because venue changes both attacking output and defensive pressure. Competition averages provide the baseline, so a scoring rate is interpreted relative to the environment in which it was produced.
Inputs are captured as a snapshot. That snapshot ties the forecast to the information available at generation time and prevents a later source update from silently changing the historical explanation.
2. Attack, defence, and expected goals
The model estimates attacking and defensive strength from goals scored and conceded, normalized by league rates and split by venue. The home attack is combined with the away defence to produce a home expected-goal rate; the reverse produces the away rate. Longer-run priors can stabilize early-season estimates when current-season samples are too small to stand alone.
Expected goals here means the model's average goal count for the matchup. It is not a claim that a fractional goal can occur and it is not the same as a single most likely score. It is the rate parameter used to distribute probability across many possible scores.
3. The score matrix
A Poisson distribution converts each expected-goal rate into probabilities for zero, one, two, and further goals. Multiplying the two team distributions creates a matrix of exact scores. Adding the relevant cells produces home-win, draw, away-win, both-teams-score, and goal-total probabilities.
Independent Poisson misses a small but important feature of football: low scores influence one another. Exact Predictors applies a Dixon-Coles correction to 0-0, 1-0, 0-1, and 1-1. When a reliable competition-specific correction is unavailable, the model uses a guarded fallback instead of accepting a stale or mismatched fit.
4. Context adjustments are bounded
Recent scoring form, active streaks, squad availability, competition maturity, and squad-quality priors can move the two expected-goal rates. Adjustments are deliberately bounded so one short run or one missing player cannot replace the season-level evidence. A match page lists the adjustments that actually fired rather than presenting them as hidden judgment.
5. Probability and confidence answer different questions
Outcome probability describes the model's distribution: for example, the combined share of score-matrix cells in which the home team finishes ahead. Confidence is an internal 0–100 evidence-quality indicator based on model conditions and residual uncertainty. It is not the observed historical success rate and must not be read as a second probability.
6. AI commentary is downstream of the statistical evidence
When available, a language model receives the structured match statistics and forecast outputs and produces a constrained scenario. It can emphasize a different plausible score, but it is using the same underlying match evidence. It is therefore a separate analysis, not an independent data source. Structured-schema validation rejects malformed output; individual scenarios are not manually edited before publication.
7. Evaluation happens after settlement
Stored forecasts are compared with final scores. Outcome accuracy checks home/draw/away classification, while exact-score accuracy requires both goal totals to match. Brier score evaluates the full probability vector rather than only its largest category. Calibration asks whether events assigned similar probabilities occur at similar rates over a sufficiently large sample.
Read the performance guide for the evaluation protocol, or inspect the published result records match by match.
Limits
- Football is low-scoring, so individual outcomes remain noisy.
- Early-season and tournament samples require stronger priors.
- Source gaps can remove a fixture or reduce confidence.
- Squad information can change after a forecast is generated.
- A model can be calibrated overall and still be wrong on a match.