Football probability glossary

The three ways a match can end: a home win (1), a draw (X), or an away win (2).

This is the core football market. WinSight gives each of the three outcomes a probability, and the three add up to 100%. It says nothing about the scoreline — only who finishes ahead. A 1–0 and a 4–0 are both simply "home win".

Whether the match will have three or more goals in total (over), or two or fewer (under).

The line sits at 2.5 so a match can never land exactly on it — there is always a clear over or under. WinSight derives this from its full grid of scoreline probabilities, adding up every scoreline on each side of the line.

Whether both teams score at least one goal — yes or no.

This is independent of who wins. A 1–1 is "yes"; a 3–0 is "no". Because it ignores the result, it often behaves differently from the 1X2 market for the same match.

The exact final score, such as 2–1.

The hardest market of the four: dozens of scorelines are possible, each with a small probability. WinSight shows the single most likely scoreline alongside how likely it is — often only 12–15%, which is normal for this market and honest about it.

A probability is our estimate of how likely something is (say, 60%). Betting odds are a price.

WinSight publishes probabilities, not tips and not prices. A probability describes uncertainty honestly; it never tells you what to stake or promises a return. Everything on this site is informational.

A plain-language label for how strongly the probabilities lean toward one outcome.

"High confidence" means one outcome is clearly the most likely — not that it is certain. Even a high-confidence call can go the other way; over many such calls, the less likely outcomes are supposed to happen sometimes. The label is a reading aid, never a promise.

Whether predicted probabilities match real-world frequencies: of all the times we say 60%, does it happen about 60% of the time?

This is the honest test of a probability model, and it matters more than raw accuracy. A model can be accurate yet overconfident. WinSight tracks its calibration openly; on backtest its calibration slope is about 0.94, where 1.0 is perfect. This is the part of the work we are proudest of.

A score for probability forecasts where lower is better; it rewards being accurate and honest about uncertainty at the same time.

The Brier score punishes overconfidence — claiming 90% and being wrong costs far more than claiming 55% and being wrong. We report it next to accuracy so a good headline number can't hide a badly-calibrated model underneath.

The betting market's final prices just before kick-off — the hardest benchmark in the sport.

By kick-off, the market has absorbed almost everything knowable: team news, money, and late information. Closing odds are extraordinarily hard to beat, and no public forecaster reliably does. We measure ourselves against them precisely because they are the honest, unforgiving bar.

Testing the model only on matches it had not yet seen, week by week, exactly as it would run live.

At each week the model knows only what it would have known at the time, then it is scored on the matches that follow. This prevents the classic mistake of using the future to "predict" the past, which is what makes most published accuracy figures untrustworthy.

The statistical model at the heart of WinSight, first published in 1997, which estimates each team's attacking and defensive strength.

From those strengths it builds a full grid of scoreline probabilities for a match — and every market on the site (1X2, over/under, BTTS, correct score) is read off that one grid. The method is public and well studied; our work is in fitting it carefully and reporting it honestly.