A Poisson model estimates the probability of specific football scorelines by treating each team's goals as a series of independent random events occurring at a known average rate, then calculating how likely any given number of goals is for each side and combining those probabilities into a full scoreline grid.
Goals in football are relatively rare events spread across ninety continuous minutes, and they occur without a fixed maximum the way, say, a single at-bat in baseball does. That combination — a countable event, no natural ceiling, occurring independently at some underlying rate — is exactly the situation the Poisson distribution was built to describe. It does not know anything about tactics or form on its own; it only needs one number per team, an expected-goals rate, and it converts that single number into a full probability distribution across every possible goal count from zero upward.
The model's core assumption is that goals are independent of one another within a match: scoring one goal does not change the underlying rate at which the next goal is likely to arrive. That assumption is a simplification — teams that go a goal up sometimes sit deeper, changing their own and their opponent's true scoring rate for the remainder of the match — but it holds well enough in aggregate to make the model one of the most widely used baseline tools in football analytics and pricing.
Before building a scoreline estimate, two inputs are required for each team: an attack-strength figure, reflecting how many goals the team tends to score against average opposition, and a defence-strength figure, reflecting how many goals the team tends to concede. Both are normally calculated relative to a league average rather than in isolation, since a raw goals-per-game figure does not account for the fact that some leagues produce far more goals overall than others. Home advantage is typically folded in as a separate adjustment on top of these two figures, since home and away scoring rates differ systematically across almost every competition.
Sample size matters as much as the inputs themselves. A handful of matches produces attack- and defence-strength figures that swing sharply on the back of one or two unusual results, while a full season or more of data smooths those swings out at the cost of reacting more slowly to genuine changes in form, such as a new manager or a key injury. Most practical implementations strike a balance by weighting recent matches somewhat more heavily than older ones rather than treating every match in the sample as equally informative, though the exact weighting scheme varies between practitioners.
Consider two hypothetical sides, Team A hosting Team B. Suppose the calculation above produces an expected-goals rate of 1.6 for Team A and 1.1 for Team B once attack strength, defence strength, league average, and the home adjustment have all been combined. Neither number says anything on its own about the final score; each is fed separately into the Poisson formula to produce a full spread of probabilities for that team's goal count — some chance of zero, a larger chance of one, a slightly smaller chance of two, and a small but non-zero chance of three or more. The two independent spreads are then multiplied together cell by cell: the chance of a 1-1 draw is the chance of Team A scoring exactly one multiplied by the chance of Team B scoring exactly one, and so on for every other combination. Because the two teams' expected-goals rates are fairly close in this example, the resulting matrix usually spreads probability across several plausible outcomes — a one-goal home win, a draw, and a narrow away win can all sit within a similar range — rather than pointing decisively at one result, which is typical whenever two expected-goals rates are not far apart.
The finished matrix is a grid of probabilities, one for every realistic scoreline combination, that together sum to 100 percent. Reading it typically shows a skewed shape rather than a smooth one: low-scoring results such as 1-0 and 1-1 usually carry meaningfully higher individual probability than any single higher-scoring result, even though the combined probability of "some kind of high-scoring game" can still be substantial once every relevant high-scoring cell is added together. This is one of the more common points of confusion when reading model output — the single most likely exact scoreline is rarely a majority outcome on its own, because probability is spread thinly across dozens of plausible combinations.
The most frequent error is feeding the model a small, recent sample of matches as the attack- and defence-strength inputs, which makes the resulting probabilities extremely sensitive to a handful of unusual results. A team that has scored heavily in its last three matches against weak opponents will look far stronger in a short-sample model than a longer, opponent-adjusted sample would suggest. A second mistake is ignoring the goal-independence assumption entirely, applying the model to situations — such as a team already leading by three goals late in a match — where real-world scoring behaviour clearly departs from a flat, independent rate. A third mistake treats the model's output as a prediction of what will happen rather than a distribution of what could happen; the model is explicitly probabilistic, and a scoreline it assigns 8 percent to is not "wrong" if that scoreline occurs, any more than a fair six-sided die is wrong for landing on a specific number one time in six.
A related and more subtle mistake is comparing raw model output directly against market odds without first converting the odds to implied probability on the same scale. The two need to sit on identical footing — probability against probability — before any meaningful comparison of where a model and a market disagree becomes possible.
Basic Poisson models treat home and away goals as fully independent of each other, when in reality the score in a match can influence both teams' behaviour simultaneously — a phenomenon statisticians sometimes correct for with adjustments that allow low-scoring draws in particular to be modelled slightly differently. RubiScore's own expected-goals figures are one example of the kind of underlying-numbers input that feeds a more refined version of this model, since a rolling expected-goals rate typically reacts to genuine attacking and defensive quality faster than raw goals-scored data does. The model also has no built-in knowledge of injuries, suspensions, tactical matchups, or any single match's specific context; every input it receives has already been reduced to two numbers per team, attack and defence strength, so any information that is not captured in those two figures simply does not reach the model at all. More advanced versions layer in time-varying strength ratings, expected-goals-based inputs instead of raw goals, and correlation adjustments between the two teams' scoring rates, but the underlying logic — reduce two teams to expected-goals rates, then generate a full probability grid from those rates — stays the same across nearly all of them.
According to RubiScore's underlying expected-goals data, Poisson-style modelling is one of several structural tools used across the industry to translate team-level scoring rates into match-level probabilities, and it works as a starting framework rather than a finished forecast. This is educational information about how probability models work, not betting advice or a betting tip, and none of the figures above should be read as a recommendation to stake money on any outcome. Betting on football carries financial risk, is intended for adults aged 18 and over, and anyone concerned about their gambling should seek advice from an appropriate support service in their jurisdiction.