How United Matchday Predictions Work

What the model reads

• Team scoring record, split by home and away matches
• Team defensive record, split by home and away matches
• Recent form over the last five and last ten matches
• An internal Elo-style team rating updated after every completed match
• The scoring environment of the competition itself (average home and away goals per match)

How the numbers are produced

Each team gets an attacking strength and a defensive weakness relative to its competition average. A team scoring 1.8 goals per home match in a league averaging 1.5 home goals has a home attacking strength of 1.2. Those strengths are combined with the competition baseline, a home-advantage multiplier and the rating difference between the two clubs to produce an expected goals figure for each side.

Expected goals are fed into a Poisson distribution, which produces the probability of every plausible scoreline. Summing those cells gives the home win, draw and away win probabilities, the over/under goal markets, both-teams-to-score and the single most likely scoreline — the predicted score you see on the page.

Weighting and sample size

Season-long performance carries the most weight, with the last ten and last five matches layered on top so recent form matters without overwhelming the longer record. Early in a season — or for a promoted side with little data — the model pulls each team back toward the competition average and lowers its confidence rather than pretending the sample is stronger than it is.

Confidence

Confidence reflects the strength and consistency of the statistical signal — sample size, separation between outcomes and how much data the competition has produced. It is not a guarantee that the predicted result will happen.

Locking and grading

A prediction is recalculated periodically until kickoff, with a final refresh a few hours before the match. At kickoff it is locked and never changes again. After full time the result is imported and the prediction is graded automatically on outcome, exact score, both teams to score and over/under 2.5 goals. Incorrect predictions stay published — the full record is on the prediction record page.

Model versions

The current model is UM Prediction Model v1.0. Every prediction stores the model version that produced it. When the model changes, historical predictions are not recalculated — they stay attached to the version that made them, so the published record remains honest.

Limits

Predictions are probabilities, not forecasts of certainty. The model does not currently read injuries, lineups, tactical changes or managerial news, and it will not claim to. Where reliable structured data for those factors becomes available, it will be added and documented here.