Poisson Goal Models

The Poisson distribution describes the probability of a given number of events happening in a fixed interval, when those events happen at a known average rate and are independent of each other. Typical examples: how many cars pass a given point per minute, how many raindrops land on a square metre per second, how many goals a team scores in a 90-minute football match.

The one parameter is lambda (λ) - the average number of events per interval. If a team's λ is 1.8 goals per match, the Poisson distribution gives you the probability of them scoring exactly 0 goals (16.5%), exactly 1 (29.7%), exactly 2 (26.7%), exactly 3 (16.0%), exactly 4 (7.2%), and so on down the tail. Adding up all the probabilities above 0 gives you the probability of them scoring at least once (83.5%).

Football goals are surprisingly well-described by Poisson. The actual distribution of team scores across a season is close to what a Poisson model with the right λ would predict. This is non-obvious - goals are not completely random or independent (scoring the first goal changes how both teams play), but the deviation from a pure Poisson is small enough that it's a good approximation most of the time.

A two-team Poisson model estimates a separate λ for each team - the home team's attacking strength adjusted for the away team's defensive strength, and vice versa. Then it assumes the two scores are independent Poisson draws and computes the joint probability of every possible scoreline. Sum the probabilities where home > away to get the home-win probability, where home = away for the draw, where home < away for the away-win.

The strength-of-schedule adjustment matters a lot. A team that has scored 2.0 goals per game but only against weak defences has a lower true λ than a team that has scored 2.0 goals per game against strong defences. The fix is to express each team's attacking and defensive strength relative to the league average, so that λ_home = league_avg × home_attack × away_defence × home_advantage. This is the Dixon-Coles-style formulation that most modern implementations use.

Once you have the scoreline probability matrix, you can derive every football betting market from it: 1X2 (sum over the triangles), Over/Under totals (sum over the diagonals), Both Teams To Score (sum over the quadrant where both teams score ≥ 1), Asian Handicap (shift the diagonals by the handicap line), correct score (read directly from the matrix). One Poisson surface, one source of truth.

The one place where pure Poisson demonstrably breaks is low-scoring matches. Pure Poisson with independent team scores under-predicts the frequency of 0-0 and 1-1 draws and over-predicts the frequency of 0-1 and 1-0 results. The deviation isn't huge (~1–2% on each scoreline) but it's systematic, and in a business where edges are measured in single percentage points, fixing it matters.

Mark Dixon and Stuart Coles published a 1997 paper that added a four-parameter correction to the Poisson joint distribution. The correction specifically inflates the probabilities of 0-0 and 1-1 and deflates 0-1 and 1-0, with no effect on any other scoreline. The mechanism is a correlation parameter that only activates on the 0-0 / 1-1 / 0-1 / 1-0 cells of the scoreline matrix.

Every serious Poisson-based football model uses some form of this adjustment today. The BetsPlug Poisson head runs a Dixon-Coles-style correction on top of the base Poisson joint distribution, with the correlation parameter fit per league (Serie A sits higher, reflecting the elevated draw rate; Bundesliga sits lower). The effect on 1X2 prediction accuracy is small but real - around +0.3 percentage points of log-loss improvement on holdout data.

Poisson assumes constant intensity across the 90 minutes, but football doesn't work that way. Teams score at different rates depending on the scoreline: leading teams defend, trailing teams attack, drawing teams vary in urgency. Pure Poisson over-predicts high-scoring matches because it ignores the 'killing the game' effect where a 2-0 leader parks the bus and both teams effectively stop trying.

The other failure mode is extreme mismatches. When a top-three side plays a bottom-three side, the raw Poisson numbers often give the underdog a 2–3% win probability, but historically those underdogs win more like 4–5%. The reason is that football is occasionally chaotic in a way the normal distribution assumptions underweight - red cards, flukes, first-minute goals changing the tactical context.

BetsPlug's ensemble works around these flaws by blending Poisson with Elo and logistic regression. The Poisson head gives tight probability estimates on balanced fixtures; the Elo head gives robust anchor estimates on mismatches; the logistic head picks up short-term form shifts Poisson doesn't see. The ensemble meta-model weighs them based on how well each has performed on similar fixture archetypes in the past.