Data inputs
Each match prediction starts from a structured snapshot built at SSG time:
- Fixture metadata: kickoff time, venue, home/away assignment, league, round.
- Recent form for each side (last 5 matches, weighted by recency).
- Head-to-head record between the two clubs over their last meetings.
- Home/away split — how each side performs at home versus on the road this season.
- Probability distributions over win/draw/loss, expected goals, and over/under markets.
For background on this metric, see Expected goals.
Models
Predictions are produced by an ensemble of large language models that consume the structured inputs above and output a calibrated probability distribution plus a short reasoning paragraph.
No single model decides a prediction. The pipeline aggregates outputs across models and falls back to the best-calibrated source if a candidate disagrees beyond a configured threshold. For the underlying probability model, see Poisson distribution.
Calibration
Prediction confidence is calibrated against historical results, not against the model's internal certainty. A 60% home-win probability means the model is right roughly 60% of the time on similarly framed matches — not that the home side will win. For the statistical concept, see Calibration (statistics).
Known limitations
Football has irreducible variance. The model has no view on injuries, suspensions, or last-minute lineup news unless those signals are present in its training inputs. Predictions are best read alongside current news, not as a replacement for it.