How our football forecasts work
Last updated:
Every LikelyGoals forecast comes from one statistical model, applied the same way to every match in every league we cover. This page explains the principles behind it and how it turns past results into probabilities.
The data behind it
The model is built on completed league matches in the same competition, using the result after 90 minutes. Cup ties, friendlies and abandoned or awarded games are left out. The current season carries the most weight. The previous season helps early on, while only a few rounds have been played, and only for teams that stayed in the same league; a promoted team starts from the league average.
Step 1: how much a league scores
For each league we measure the average number of goals scored by home teams and by away teams per match. Early in a season these averages are blended with last season's, and the current season takes over as more matches are completed.
Step 2: how strong each team is
Each team gets an attacking strength and a defensive strength, both measured against the league average. A team that scores 20% more than an average side has an attacking strength of 1.2; a team that concedes 20% less has a defensive strength of 0.8. Home and away results are combined, with each one compared against the matching home or away average.
Small samples can mislead, so each strength is pulled towards a starting estimate until enough matches have been played. For a team that stayed in the league, that estimate is its strength from last season; for a promoted team it is the league average.
Step 3: expected goals
The model combines these numbers into expected goals for each side. The home side's figure is the league's home-scoring average, multiplied by the home team's attack and by the visitors' defence; the away figure works the same way the other round. These are model expected goals, calculated from goals scored and conceded. They are not the shot-based xG figures published by data companies.
Step 4: from expected goals to probabilities
From the two expected-goals figures the model calculates the probability of every scoreline from 0-0 up to ten goals a side, using the Poisson distribution, a standard way of modelling how often goals occur. Plain Poisson slightly misjudges low scores such as 0-0 and 1-1, so a well-known adjustment, the Dixon-Coles correction, is applied to them.
Every figure we publish comes from this one table of scorelines: the home win, draw and away win probabilities, the chance of more than 2.5 goals and the chance of both teams scoring. That keeps the numbers consistent with each other.
Step 5: choosing the forecast
The forecast is normally the most likely of the three results. Draws are the exception. They are rarely the single most likely outcome, so a model that only ever picks the top result almost never forecasts one. We therefore forecast a draw when its probability is high enough and close enough to the leading result.
The predicted score is the most likely scoreline within the forecast result, so a home-win forecast always comes with a home-win score. Alongside it we show the double chance with the highest probability.
How it was tested
Before going live, the model was run on 3,216 matches from the 2025/26 season across our ten leagues, in date order, so that each forecast used only results available at the time. Its settings were chosen on the first 60% of those matches and then checked on the remaining 40%, which played no part in choosing them.
What we keep to ourselves
This page describes the principles of the model. The exact way it is tuned and weighted is our own work, developed and tested in-house, and we do not publish those details.
When a forecast is final
Once a forecast is published, it does not change. If the match is moved by more than a week, the forecast is void and a new one is published closer to the new date.
Forecasts are probabilities, not promises. LikelyGoals is for adults aged 18 and over; if you bet, only stake what you can afford to lose.