How the FootyGrounds Match Predictor works
A plain-English explanation of the FootyGrounds prediction model: the data it uses, how strength, goals and form models are blended, how confidence is scored, and what the model cannot know.
What the model uses
- Historical results and league standings held in the FootyGrounds database
- Recent form, split into home and away performance
- Goals scored and conceded, used to build attack and defence rates
- League context, such as competitive balance and typical scoring levels
- Managerial changes, which temporarily widen the uncertainty around a club
The three components
- Strength ratings. An Elo-style rating that rises and falls with results, weighted by the quality of the opponent.
- Goals model. A Poisson model that turns attack and defence rates into a distribution of likely scorelines, which produces win, draw and away probabilities plus a most-likely score.
- Recent-form model. A shorter-horizon view of each club's last matches, weighted toward home or away performance as appropriate.
The three outputs are blended into a single set of probabilities. Where the components disagree, the blend stays closer to the middle and the confidence score falls.
The confidence score
Every prediction carries a FootyGrounds confidence score out of 100. It starts from the strength of the modelled probability, then is reduced when the component models disagree, when there are few recent matches to learn from, or when the data for a club is thin. The score is capped at 95 — the model never claims certainty. Data quality is published alongside every prediction as limited, good or excellent.
What the model cannot know
- Late injuries, illness and suspensions confirmed close to kick-off
- Rotation, rested players and cup priorities
- Weather, pitch conditions and refereeing decisions
- Dressing-room and off-field events not reflected in results
- Individual moments of brilliance or error, which decide many matches
Football is a low-scoring sport with high variance. A 70% probability still means the outcome fails roughly three times in ten.
Transparency and accountability
Predictions are locked at kick-off so they cannot be changed retrospectively, then settled against the real result. Every published prediction, including the ones that were wrong, stays available on the history page, together with accuracy by competition and a calibration table showing whether outcomes predicted at, say, 70% actually occur about 70% of the time.