If you have ever seen a model spit out "Argentina 38% to win, 31% draw, 31% Brazil," you have seen the Poisson distribution at work. It is the quiet engine under almost every scoreline forecast in football: assume each side scores goals like raindrops hitting a windowpane (rare, independent events arriving at some steady rate), and the whole grammar of match probabilities falls out of one little formula. The model is everywhere. What almost nobody does is the obvious next step: take real results and ask whether goals actually behave that way. So we did, first with 9,170 Premier League matches and then with all 64 matches of the 2022 World Cup, and the answer is the interesting kind of "almost": total goals fit the model closely, and the draws do not.
The model, in one line
The Poisson distribution gives the probability of seeing exactly k goals when the average rate is λ (lambda):
Poisson probability of k goalsP(k) = λk · e−λ / k!
That's it. Feed it a rate and it hands back the full distribution of outcomes. The appeal for football is that it needs only one number per team (an expected goals rate), and the assumptions sound exactly like a low-scoring sport: goals are rare, they don't (the theory says) depend on each other, and they fall at a roughly constant rate across the 90 minutes. Set λ to the average and you have a parameter-free prediction you can check against reality.
Do goals follow a Poisson? The test on 9,170 Premier League matches
Sixty-four World Cup games are a small sample, so we ran the test on the Premier League archive first: ESPN's public scoreboard for every match since 2002-03, pulled on 5 October 2026. That is 9,170 matches once ESPN's duplicated 2009-10 listing is removed. The tests use the 9,120 from the 24 complete seasons, 2002-03 to 2025-26. The 50 matches of 2026-27 in the file (to 20 September 2026) are left out of every fit, because part of a season can't support team ratings. For the record, they averaged 2.82 goals, with 16 draws and 5 goalless games.
Total goals: very close to a Poisson. The 9,120 matches produced 24,875 goals, a mean of 2.73 and a variance of 2.78, so the variance-to-mean ratio is 1.02. A Poisson gives exactly 1. At that mean the model gets every band from one to seven goals within 10%, and three-goal games land almost exactly: 2,010 played, 2,016.5 expected. The misses are at the ends: 666 goalless games against 596 expected (12% more), 308 six-goal games against 341, and 74 games of eight or more against 64. With this many matches even small misses show up in a formal test: a chi-square of 19.96 on 7 degrees of freedom, which a true Poisson would produce well under 1% of the time (p = 0.0057). The shape is right; the detail is not quite.
| Goals in match | Observed matches | Poisson expects | Gap |
|---|---|---|---|
| 0 | 666 | 596.3 | +11.7% |
| 1 | 1,567 | 1,626.4 | −3.6% |
| 2 | 2,167 | 2,218.0 | −2.3% |
| 3 | 2,010 | 2,016.5 | −0.3% |
| 4 | 1,444 | 1,375.0 | +5.0% |
| 5 | 746 | 750.1 | −0.5% |
| 6 | 308 | 341.0 | −9.7% |
| 7 | 138 | 132.9 | +3.9% |
| 8+ | 74 | 64.0 | +15.7% |
Season by season, the ratio wanders both ways. It runs from 0.84 (2023-24, the highest-scoring season at 3.28 goals a game) to 1.16 (2007-08), with 14 of 24 seasons above 1. Over 380 matches a true Poisson lands within about ±0.14 of 1 in 19 seasons of 20. Here 4 of 24 fall outside that range (2003-04 and 2007-08 above, 2019-20 and 2023-24 below), where chance alone would put about one, and 6 of the 8 seasons since 2018-19 sit below 1. So the league is not steadily overdispersed the way the World Cup sample below is; recent seasons, if anything, are a little underdispersed. One measurable pull in that direction: home and away goals in the same match are slightly negatively correlated (r = −0.085), which narrows the spread of totals. The basic model assumes they are independent.
Draws: the excess appears once teams differ. 2,250 of the 9,120 matches were drawn (24.7%). If every match in a season had that season's average home and away scoring rates (1.54 and 1.19 a game over all 24 seasons), independent Poissons would give 2,299 draws, slightly more than were played. But mismatched teams draw less often. Give each fixture its own pair of rates, from Maher's model of each team's attack and defence fitted to that season's results, and the expectation falls to 2,113 draws (23.2%). The league produced 137 more than that, about 3.4 standard deviations, and beat the model's draw count in 18 of 24 seasons.
| Scoreline | Played | One rate per season expects | Team rates expect |
|---|---|---|---|
| 0-0 | 666 | 604.4 | 673.4 |
| 1-1 | 1,000 | 1,085.3 | 964.5 |
| 2-2 | 487 | 494.7 | 387.4 |
| 3-3 or higher | 97 | 114.9 | 87.8 |
| All draws | 2,250 | 2,299.2 | 2,112.9 |
The 0-0 count, and where the extra draws really are. Against a single Poisson there are 666 goalless games where 596 were expected. With team rates the model expects 673 and there were 666: the 0-0 excess disappears, because low-scoring fixtures between defensive sides produce the blanks a single rate can't. The surplus of draws sits at 1-1 (1,000 against 964) and, mostly, 2-2 (487 against 387, about a quarter more). Dixon and Coles' correction, described below, adjusts only 0-0, 1-0, 0-1 and 1-1. Two of its moves fit this data: 1-1 is under-predicted and the narrow away win over-predicted (661 0-1 results against 758 expected). But 0-0 already fits, and the biggest gap, 2-2, is outside the four cells it touches.
Limits. The team rates are fitted on the same season they are tested on, which flatters the model a little; a model fitted only on earlier games would be a stricter test. This is one league, full-time scores as ESPN records them, and the plain Maher model (no form, no time weighting). The ±0.14 range and the chi-square p-value are large-sample approximations. The two numbers to keep: a variance-to-mean ratio of 1.02 for total goals, and 137 more draws than a team-strength Poisson allows. For how the same goal-rate idea runs minute by minute inside a match, see how win-probability models work.
Fitting it to 2022
The World Cup section below was published in June 2026 and is kept as written, apart from small edits where the Premier League test changes a general claim (two are marked as corrections). The honest test is simple. Across all 64 matches the 2022 World Cup produced 172 goals, a mean of 2.69 goals per match (add every home_score and away_score, divide by 64: no model in that step). Set λ = 2.69, compute the Poisson probability of a 0-goal game, a 1-goal game, and so on, multiply each by 64, and you get how many matches of each total the model expects. Lay that next to what actually happened.
| Goals in match | Observed matches | Poisson expects |
|---|---|---|
| 0 | 7 | 4.4 |
| 1 | 10 | 11.7 |
| 2 | 17 | 15.7 |
| 3 | 14 | 14.1 |
| 4 | 4 | 9.5 |
| 5 | 6 | 5.1 |
| 6 | 3 | 2.3 |
| 7+ | 3 | 1.2 |
Observed counts from StatsBomb Open Data, 2022 FIFA World Cup; expected counts = 64 × Poisson(λ = 2.69).
Look at the middle of that table and the model is genuinely good. Two-goal games: 17 observed, 15.7 predicted. Three-goal games: 14 observed, 14.1 predicted. The single most common scoreline band and the overall right-skewed shape (most games clustering at two or three, a long thin tail of high-scoring ones) come straight out of the formula. For a one-parameter model that knows nothing about who was playing, that is a real success, and it is why the Poisson has survived as the default since Michael Maher formalized it for football in 1982.
Where it breaks, and it breaks in a revealing way
Now read the extremes. There were 7 goalless or near-goalless games where the model expected about 4.4. There were 3 matches of seven or more goals where it expected barely 1. And right in the middle of the upper range, four-goal games, reality came in under the model: 4 observed against 9.5 expected. The pattern is unmistakable: too many blanks, too many blowouts, not enough of the tidy mid-range games the Poisson loves. The distribution has fatter tails than pure randomness allows.
There's a precise name for this. A Poisson distribution has a defining property: its variance must equal its mean. Here the mean is 2.69 but the variance of match totals is 3.53: a variance-to-mean ratio of 1.31. These 64 games are overdispersed: more spread out than the model permits. That single number, 1.31 instead of 1.00, is the whole failure quantified.
Why the 2022 goals are overdispersed (the part worth keeping)
Two of the Poisson's assumptions are wrong in football, and each pushes the variance up.
First, λ is not the same in every match. We fit one rate, 2.69, to all 64 games, but Spain–Costa Rica is not Croatia–Morocco. A lopsided fixture has a high expected rate and tends to produce a rout; two cagey, well-matched defenses have a low rate and grind to 0–0. When you average genuinely different rates into one number, the result is mathematically guaranteed to be overdispersed: a mixture of Poissons always has variance greater than its mean. The extra blanks and the extra blowouts aren't noise: they're the fingerprints of team quality the single-λ model deliberately ignored. (It's the same heterogeneity that makes a tournament's 64 results spread the way they do.)
Second, goals within a match aren't independent. The Poisson assumes the next goal is as likely after a 3–0 as after a 0–0. Football doesn't work like that: a team three up empties the bench and protects the lead, while a team chasing the game throws numbers forward and either rescues it or gets picked off on the break. Game state couples the goals together, thickening both tails: the comfortable winner adds a fourth and fifth, or the cagey pair never breaks the seal.
Correction (November 5, 2026): the Premier League test above shows these two forces need not add up to overdispersion. Over 9,120 league matches the variance-to-mean ratio is 1.02, and home and away goals are slightly negatively correlated (r = −0.085), which pulls the spread of totals down rather than up. The 2022 ratio of 1.31 comes from 64 games, where a few routs move the variance a long way.
How the pros patch it
Nobody who builds these models for a living uses a single league-wide λ. The standard upgrade is to give each team its own attack and defence strength, so every fixture gets its own pair of rates, which is exactly what a real scoreline model does, and what cures most of the heterogeneity problem. On top of that, the Dixon–Coles model (1997) adds a small correction specifically to the low-score cells (0–0, 1–0, 0–1, 1–1) to fix the dependence the plain Poisson gets wrong. The fix targets the low scores because that is where the World Cup chart above shows the gap. Correction (November 5, 2026): in the Premier League, once each team has its own rates, 0-0 fits (666 played, 673 expected) and the extra draws sit at 1-1 and 2-2; see the test above.
You can feel the per-team version directly. Drop two expected-goals rates into the model below and it returns the win/draw/loss probabilities and the likeliest scorelines, the same Poisson, but with a rate for each side instead of one for the whole tournament:
Interactive scoreline model: enter each team's expected goals to see the Poisson outcome probabilities.
A quick worked check
If you want to confirm the table by hand, take the goalless case. At λ = 2.69, P(0) = e−2.69 = 0.068, so over 64 matches the model expects 64 × 0.068 ≈ 4.4 blank games. We saw 7. That gap of nearly three matches, about 60% more shutouts than predicted, is small in raw count and large in meaning: it's the low-scoring, low-λ fixtures that a single average rate can't represent. Run the same arithmetic for k = 2 (P = 0.246, expecting 15.7, observing 17) and the middle of the distribution lands almost exactly. The model is right where the matches are average and wrong where they aren't.
Mostly random, but not quite
"Are goals random?" has a satisfying answer: mostly, but not quite. The Poisson, the simplest possible model of randomness, reproduces the shape of real World Cup scoring well enough that you understand why the entire field is built on it. But it fails in a specific, honest direction. Over 9,120 Premier League matches the total-goals ratio is 1.02, close to the model's 1, while the 64 World Cup games reach 1.31; the gap that holds up in the big sample is the draws, 137 more than a team-strength Poisson allows. The good forecasters don't throw the Poisson out; they keep its skeleton and add back exactly the two things it pretends away: that teams differ, and that the score so far changes what happens next.
Reproduce it
Every figure in the Premier League section (added November 5, 2026) comes from our analysis script. It reads data_layer/epl_goal_times_2002_2027.json, ESPN's public Premier League scoreboard pulled on 5 October 2026, through our archive loader, which drops the 357 duplicated 2009-10 rows. It keeps the 24 complete seasons, fits the single Poisson, the per-season rates and Maher's team model (iterative proportional fitting), writes every number to data/stats/do-football-goals-follow-a-poisson-epl.json and draws the chart. No network, nothing hand-entered.
The 2022 World Cup numbers come from one results file and the Poisson formula. Total goals per match is home_score + away_score; λ is their mean; the expected counts are 64 × λk e−λ / k!; the overdispersion is just variance / mean. The observed-versus-fitted chart is regenerated by our chart script against the 2022 World Cup match file in StatsBomb’s open-data repository: no network, nothing hand-entered.
Where these numbers come from
- Premier League section (added November 5, 2026): every completed Premier League match 2002-03 to 2025-26 (9,120 of the 9,170 in the file) in /data_layer/epl_goal_times_2002_2027.json. ESPN public scoreboard (soccer/eng.1), pulled 5 October 2026. Computed by our analysis script.
- Match results: the 2022 World Cup match file in StatsBomb’s open-data repository (free public dataset; attribution required), 2022 FIFA World Cup, all 64 matches. Charted by our chart script.
- M. J. Maher (1982), "Modelling association football scores": the foundational Poisson treatment.
- M. Dixon & S. Coles (1997), "Modelling association football scores and inefficiencies in the football betting market": the low-score dependence correction.
- Companion: Inside the 2022 World Cup: what all 64 results say (the descriptive look this one stress-tests).
- How-to: Build a Poisson goals model in Python (the per-team version, in code).
- Background: Expected goals (xG) explained (where each team's λ comes from in practice).
Published by SoccerAnalytics from public data and the sources named in this piece. How this site is made.