Sport News

By Martin · Published 3rd November 2026 · Last updated 3rd November 2026

Quick answer: The Dixon-Coles model is an improved version of the basic Poisson football model that corrects a specific flaw: standard Poisson underestimates low-scoring draws like 0-0 and 1-1, and slightly misprices 1-0 and 0-1 results. It does this by adding a correction factor to those four low-scoring scorelines and by weighting recent matches more heavily than old ones. The reason the correction is needed is that basic Poisson assumes the two teams score completely independently, which is not quite true in tight, low-scoring games. Dixon-Coles makes football prediction models noticeably more accurate on exactly the results that decide the most matches. AMpredict uses correlation corrections of this kind within its wider model.

If you build a basic Poisson model and compare its scoreline predictions to real football, one gap jumps out straight away.

It does not produce enough draws. Specifically, it underestimates how often matches finish 0-0 and 1-1, the tight, low-scoring results that happen far more in reality than pure maths expects. For anyone relying on the model, that is a meaningful blind spot, because those exact scorelines decide a huge number of matches and heavily influence the draw and under markets.

In 1997, two statisticians, Mark Dixon and Stuart Coles, published a fix. Their model kept everything good about Poisson and added a targeted correction for precisely this flaw. This guide explains, in plain English, why basic Poisson gets draws wrong, how the Dixon-Coles correction works, and why it matters for prediction. It follows directly from our explainer on how the Poisson distribution predicts football scorelines.

What is the Dixon-Coles model?

The Dixon-Coles model is a statistical football prediction model, published in 1997, that improves on the basic Poisson model by correcting its underestimation of low-scoring results and by giving more weight to recent matches. It keeps the Poisson framework for generating scoreline probabilities but adjusts the four lowest scorelines and adds time-weighting, making it more accurate than plain Poisson.

The model was introduced in a now-famous academic paper by Mark Dixon and Stuart Coles, and it has become one of the most influential approaches in football modelling.

Its importance is that it was one of the first models designed specifically to predict and profit from football results, rather than just describe them. It made two targeted changes to standard Poisson. The first is a low-score correction, which fixes the draw problem. The second is time-weighting, which makes recent form count for more than results from months ago. Both changes are small in isolation but together they produce a measurably better model.

Crucially, Dixon-Coles did not throw out Poisson. It built on it. Everything covered in the step-by-step calculation of a mathematical prediction still applies; Dixon-Coles simply adds two refinements on top.

Why does basic Poisson underestimate draws?

Basic Poisson underestimates low-scoring draws because it assumes the two teams score completely independently of each other, when in reality their scoring is slightly linked, especially in tight games. When a match is goalless or 1-1 late on, both teams often become more cautious, which produces more low-scoring draws than a model assuming independence predicts.

Independence is the core assumption that breaks down.

In the basic model, the probability of a 1-1 draw is simply the home team’s chance of scoring one goal multiplied by the away team’s chance of scoring one. That multiplication treats the two events as unrelated. But football is not unrelated. In a tense 0-0 or 1-1, teams frequently prioritise not losing over trying to win, sitting deeper and taking fewer risks. That shared caution makes those exact scorelines more common than the independent maths expects.

The result is a predictable error. Across a large sample, a basic Poisson model will forecast fewer 0-0 and 1-1 results than actually occur, and slightly misprice 1-0 and 0-1 too. These are not random misses; they cluster on the four lowest scorelines every time, which is exactly what makes the error correctable. This is the kind of subtle, repeated pattern that separates a refined model from a crude one, the same principle behind why possession and other surface stats mislead: the obvious model misses what the careful one captures.

How does the Dixon-Coles correction work?

The Dixon-Coles correction works by applying an adjustment factor, usually written as the Greek letter tau, to the four lowest scorelines: 0-0, 1-0, 0-1, and 1-1. This factor nudges the probabilities of 0-0 and 1-1 upward and 1-0 and 0-1 slightly downward, bringing the model’s low-score predictions into line with how often those results actually happen.

The correction is surgical: it touches only those four scorelines and leaves everything else in the Poisson grid unchanged.

Here is the effect in practice. Take a fixture where a basic Poisson model produces the low-score probabilities in the first column below. The Dixon-Coles correction adjusts them to the second column:

The two draw scorelines get lifted, the two single-goal wins get trimmed slightly, and the total probability still balances across the grid. The size of the adjustment is set by the tau parameter, which is estimated from historical data rather than guessed. The exact value depends on the league, because some leagues are tighter and lower-scoring than others, so the correction is calibrated to the competition being modelled.

What is time-weighting in the Dixon-Coles model?

Time-weighting in the Dixon-Coles model means giving more importance to recent matches than to older ones when estimating a team’s strength, on the logic that a game from last month tells you more about a team’s current level than a game from last season. Each past match is given a weight that decays the further back in time it sits.

This addresses a weakness in basic Poisson, which typically treats every match in the sample equally.

Treating a result from eight months ago as just as relevant as last weekend is clearly wrong, because teams change: new signings, new managers, injuries, form swings. Dixon-Coles applies a decay function so that recent results dominate the strength estimate and old results fade in influence. A match from a week ago might carry close to full weight, while one from a year ago carries very little.

The practical effect is a model that reacts faster to a team’s current level rather than being anchored to stale history. This is the same logic that underpins regression to the mean and reading current form correctly: recent, relevant data beats a long but outdated record. The decay rate, like tau, is estimated from data rather than set arbitrarily.

Why does the draw correction matter for prediction?

The draw correction matters because low-scoring draws decide a large share of matches and dominate the draw and under-goals markets, so getting their probability right has an outsized effect on prediction accuracy. A model that underestimates 0-0 and 1-1 will systematically misprice those outcomes, leaking accuracy on exactly the results that occur most often in tight fixtures.

Draws are not a rounding error in football; they are a major outcome.

In a typical league season, roughly a quarter of matches end in a draw, and a large share of those are the low-scoring kind the correction targets. A model that is systematically too low on 0-0 and 1-1 will therefore misjudge the draw market again and again, and will also get the under 2.5 goals market wrong, since those scorelines sit firmly in under territory. Over hundreds of fixtures, that systematic error compounds.

Fixing it has a real payoff. By bringing the low-score probabilities into line with reality, Dixon-Coles improves accuracy precisely where basic Poisson is weakest, which is why it became a benchmark model. It is a clear example of how refining the maths, rather than adding more guesswork, is what genuinely lifts a model, the same theme running through the complete guide to mathematical prediction.

What are the limits of the Dixon-Coles model?

The limits of the Dixon-Coles model are that it still cannot see current context, it only corrects low-scoring results rather than every imperfection, and it remains a goals-based statistical model that knows nothing about team news or motivation. It is a better model than basic Poisson, but it is still pure maths, and pure maths has a ceiling.

Dixon-Coles fixes a specific, well-defined flaw. It does not make the model omniscient.

The model still works entirely from historical results and expected goals. It does not know that a key striker is injured, that a team is resting players before a cup final, or that a side has nothing left to play for. Those context factors decide many matches and sit completely outside any Poisson-based model, which is why even a strong statistical model plateaus and why human review still beats pure AI. It also only corrects the four lowest scorelines; other subtle imperfections in the independence assumption remain.

So Dixon-Coles is best understood as a sharper tool, not a complete system. It belongs inside a wider model that adds context and judgement, rather than being used alone. No statistical model, however refined, escapes football’s irreducible randomness, which is why sure wins are mathematically impossible no matter how good the maths gets.

How does AMpredict handle the draw problem?

AMpredict handles the draw problem using correlation corrections of the Dixon-Coles type within its wider model, rather than relying on uncorrected Poisson. The low-score adjustment and recency-weighting are built into the mathematical layer, which is then combined with 250+ data points, AI pattern recognition, and human review before any prediction is published.

The draw correction is one of many refinements inside the system, not the whole of it.

AMpredict’s mathematical layer already accounts for the fact that tight games produce more low-scoring draws than naive independence suggests, and it weights recent form above stale results, the two things Dixon-Coles introduced. But it does not stop there. Those corrected probabilities feed into the full three-layer methodology, where AI checks them against historical patterns and human experts adjust for the current context the maths cannot see. That is what turns a good statistical model into a tracked 89% accuracy on High Confidence picks.

You can see the finished probabilities, including honest draw and under markets, on the AMpredict homepage, with the full range across confidence tiers inside the VIP packages. If you want to build or study a Dixon-Coles model yourself, the expected-goals data it needs is published free by FBref and Understat, and the expected-goals methodology is documented by Opta’s analysts.

The Bottom Line

The Dixon-Coles model exists to fix one clear flaw in basic Poisson: it underestimates low-scoring draws like 0-0 and 1-1, because it wrongly assumes the two teams score completely independently. The fix is a targeted correction to the four lowest scorelines, plus time-weighting that makes recent matches count for more than old ones. Both changes are small but they sharpen the model exactly where it is weakest, on the tight results that decide so many games.

It matters because draws are a major outcome, roughly a quarter of all matches, and a model that misprices 0-0 and 1-1 leaks accuracy on the draw and under markets every week. Correcting that is a real, measurable improvement, which is why Dixon-Coles became a benchmark in football modelling.

But it is still pure maths with a ceiling. It knows nothing about injuries, rotation, or motivation, so it works best inside a wider system. At AMpredict, corrections of this kind sit within the full model, combined with hundreds of data points, AI, and human review into calibrated forecasts tracked at 89% on High Confidence picks.

Want draw and under markets priced properly without building the model yourself? See today’s predictions on the AMpredict homepage, or view the membership plans to get the full model on every fixture.