How the MatchSense Prediction Model Works

A prediction is only worth as much as the process behind it, so this page describes that process rather than asserting an accuracy figure. Nothing here is proprietary in a way that would prevent you from checking the reasoning.

Step 1: team strength

Every team carries a rating that updates after each result. The update is Elo-derived: beating a strong opponent moves the rating more than beating a weak one, and the margin of victory is taken into account with diminishing returns, so a 6-0 does not count as three times a 2-0.

Ratings decay toward the mean across a long off-season and are seeded from the previous campaign rather than reset, which is why a newly promoted side does not start level with a title contender.

Step 2: from strength to expected goals

Team strength alone does not price a market. The model converts it into two expected goal figures — one for each side — using separate attacking and defensive parameters plus a home advantage term fitted per competition, because home advantage is not a universal constant and differs measurably between leagues.

Step 3: from expected goals to a scoreline distribution

Two expectations are turned into the probability of every plausible scoreline using a Dixon-Coles variant of the bivariate Poisson model.

Plain Poisson is the textbook approach and it has a known flaw: it underestimates low-scoring draws, 0-0 and 1-1 in particular, because it treats the two teams' scoring as fully independent when in reality low-scoring matches are correlated events. Dixon-Coles applies a correction in exactly that region, and additionally weights recent matches more heavily than old ones.

Once you have a full scoreline distribution, every football market falls out of it by summation:

• Match result: sum the cells where home > away, home = away, away > home
• Over 2.5: sum every cell whose goals total three or more
• Both teams to score: sum every cell where both values are at least one
• Correct score: read the cell directly

This is why our market probabilities are internally consistent with each other rather than being estimated separately.

Step 4: context the goal model cannot see

A goal model knows about goals. It does not know that a team played 72 hours ago, flew 2,000 kilometres, or lost its first-choice goalkeeper in the warm-up.

A gradient-boosted ensemble adds those features on top: rest days and fixture congestion, travel distance, lineup and injury information where available, competition context such as a dead rubber or a cup rotation, and market signals. The ensemble learns where the goal model is systematically wrong and corrects it, rather than replacing it.

Step 5: calibration

Raw model output is not published directly. It is passed through a calibration layer fitted on historical predictions and outcomes, so that predictions stated at 70% actually occur about 70% of the time.

This step matters more than model sophistication. An uncalibrated model with excellent discrimination will still lose money when staked, because expected value calculations inherit its overconfidence. See how to read football prediction probabilities for how to verify calibration yourself.

Step 6: expected value screening

The final stage compares the calibrated probability against the market price with the bookmaker margin removed. Only where the model and the market disagree by a meaningful margin does a pick get flagged as value.

The consequence is deliberate: most fixtures produce no recommendation. A model that finds an edge in every match has not found edges, it has found noise. The method is described in expected value in football betting explained.

Step 7: grading and retraining

Every published pick is graded against the final result. Those grades feed the accuracy record, the calibration re-fit, and the retraining trigger that decides when a model has drifted far enough to need rebuilding.

If you want to judge the output rather than the description, the honest place to look is the record over hundreds of graded picks and the calibration curve behind it — which is exactly what how to evaluate a football prediction service tells you to demand from anyone, including us.