How WinSight works

Every prediction starts from the Dixon–Coles model. It estimates how strong each team is at scoring and at preventing goals, adjusts for home advantage, and produces a full grid of scoreline probabilities — the chance of 0–0, 1–0, 2–1, and so on. The match result, over/under, both-teams-to-score and correct-score numbers are all read off that single grid. That is why they always agree with one another: they are different views of the same underlying picture.

A team is not the same from one season to the next — players, managers and form all change. So the model weights recent matches more heavily than old ones, letting the distant past fade gradually rather than counting a result from five years ago as if it were yesterday's. How quickly the past fades is not guessed; it is tuned on the walk-forward backtest, where we can see which setting would have predicted best without ever letting the model see the future.

A newly promoted side has no recent record in its new division, so the model cannot yet judge it directly. It starts that team from the average of comparable teams — those around it in strength — and then lets the team's own results take over as they arrive. Within about ten matches the real evidence has replaced the starting assumption. Early predictions for these teams carry more uncertainty, and that uncertainty is meant to be there.

This is the plainest thing we can tell you: WinSight does not claim to beat the bookmakers, and you should be wary of anyone in this field who does. By the time a match kicks off, the betting market has absorbed almost everything that can be known about it — team news, injuries, and the weight of real money pushing prices toward the truth. Closing odds are the sharpest forecast in the sport, and no public model reliably beats them. On the same 16,322 matches, our honest walk-forward backtest trails Bet365's closing favourite by 2.16 percentage points of accuracy. We publish that gap rather than hide it.

So what is WinSight for? For probabilities that are honest about their own uncertainty, and for a record that is public, permanent, and never edited — including the predictions we got wrong. That record is the one thing a confident-sounding tipster can never show you, and it is the only thing we ask you to judge us on.

A 60% chance is not a promise. It means that if the same situation played out many times, we would expect it to happen about six times in ten — which also means it should fail about four times in ten. A single high-confidence prediction going the wrong way is not a broken model; a model whose 60%s only come true 40% of the time is. That difference is exactly what calibration measures, and it is why we publish our calibration curve rather than just a headline accuracy number. Read the probabilities as a careful estimate of uncertainty, not as a forecast of what must happen.