Poisson vs Dixon-Coles vs Bivariate Poisson for Football Scores
Independent Poisson, Dixon-Coles and bivariate Poisson models can all turn team goal rates into a complete score probability matrix. Their main difference is not how they estimate team strength, but how they represent dependence between the two teams’ goal counts. Independent Poisson assumes none, Dixon-Coles changes four low-score cells, and the standard bivariate Poisson introduces a shared scoring component across the entire matrix.
- Independent Poisson: simplest and most transparent; home and away goals are conditionally independent.
- Dixon-Coles: preserves Poisson marginals but reallocates probability among 0-0, 1-0, 0-1 and 1-1.
- Bivariate Poisson: adds a shared Poisson component and allows positive covariance across many scores.
- Dixon-Coles can represent limited positive or negative local dependence; the standard shared-component bivariate Poisson allows only non-negative covariance.
- None of the three automatically solves changing team strength, red cards, game state or non-Poisson marginal dispersion.
- The best model must be selected out of sample, not by which one fits one match most closely.
Football score modelling has two separate layers
Discussions of “Poisson versus Dixon-Coles” often mix together two different questions. For an introduction before the technical comparison, read Maher and Dixon-Coles models in plain words.
The rate model can use historical attack and defence strengths, projected xG, player and lineup variables, or current betting odds. The joint distribution can then be independent Poisson, Dixon-Coles, bivariate Poisson or another model.
For example, the expected goal rates could be produced by Maher’s attack–defence model and then combined with Dixon-Coles. Rates can also be recovered from market odds. If the purpose is to fit those odds, the chosen joint distribution must be active during fitting; simply inserting rates fitted under another distribution need not preserve the market fit.
Changing the joint distribution does not automatically improve the goal-rate estimates. A sophisticated dependence model with poor λ inputs can forecast worse than a simple independent model with well-calibrated inputs.
A terminology warning: Dixon-Coles is also bivariate
“Bivariate” simply means that a model describes two variables jointly. In that general sense, Dixon-Coles is a bivariate model with Poisson marginal distributions.
In football analytics, however, bivariate Poisson usually refers more specifically to the shared-component construction:
where the three X variables are mutually independent Poisson variables. This page uses “bivariate Poisson” in that narrower standard sense.
The distinction matters because Dixon-Coles and shared-component bivariate Poisson create dependence in very different ways even though both have Poisson marginals.
Independent Poisson, Dixon-Coles and bivariate Poisson at a glance
| Property | Independent Poisson | Dixon-Coles | Shared-component bivariate Poisson |
|---|---|---|---|
| Marginal goal distributions | Poisson | Poisson | Poisson |
| Dependence parameter | None | ρ | λ3 |
| Dependence location | Nowhere | Only 0-0, 1-0, 0-1 and 1-1 | Across the full score matrix |
| Possible covariance | Zero | Small positive or negative, subject to bounds | Non-negative only in the standard construction |
| Changes marginal means? | No | No | λ3 contributes to both marginal means |
| Changes total-goals distribution at fixed marginal means? | Baseline | Yes — but only totals 0, 1 and 2; O/U 2.5 and higher half-goal lines are unchanged | Yes |
| Extra computational cost | Low | Low | Moderate |
| Main use case | Transparent baseline | Low-score and draw correction | Broader positive dependence |
| Main structural limitation | No conditional dependence | Only four scores are changed directly | Cannot represent negative covariance in its standard form |
Model names do not determine quality. A bivariate Poisson model is not automatically “more advanced and therefore better.” Its additional parameter can improve fit, do nothing, or overfit depending on the data and estimation procedure.
Model 1: independent Poisson
Let:
If H and A are conditionally independent:
What independence means
Conditional independence does not claim that the two teams have no relationship in the real world. The expected rates can already depend on both teams:
Independence states that after those rates and covariates are fixed, the remaining goal-count variation is independent.
Useful mathematical properties
- The score matrix is an outer product of two one-dimensional Poisson vectors.
- Total goals follow Poisson with rate μhome + μaway.
- Goal difference follows a Skellam distribution.
- Each marginal mean equals its variance.
- Market probabilities are fast and stable to calculate.
Strengths
- transparent assumptions;
- only two match-level goal-rate parameters;
- easy estimation, simulation and debugging;
- strong baseline for testing whether extra dependence is justified;
- coherent 1X2, totals, BTTS and exact-score outputs.
Limitations
The model cannot represent residual dependence after the rates are fixed. If 0-0 and 1-1 are systematically overrepresented, or if game-state interaction affects many scores, the product of the marginal distributions can misallocate probability.
It also retains Poisson marginal dispersion. Dependence and marginal overdispersion are separate problems: adding dependence does not automatically fix a poor marginal count model.
See the Poisson football model guide for a full derivation.
Model 2: Dixon-Coles
Dixon-Coles starts with the independent Poisson matrix and multiplies four low-score cells by a correction function τ:
| Score | τ multiplier | Effect when ρ < 0 |
|---|---|---|
| 0-0 | 1 − μhμaρ | Increases |
| 1-0 | 1 + μaρ | Decreases |
| 0-1 | 1 + μhρ | Decreases |
| 1-1 | 1 − ρ | Increases |
| Every other score | 1 | Unchanged directly |
What the correction preserves
The four probability changes cancel. Total probability remains 1, and the marginal home and away goal distributions remain Poisson with the same means.
Under the sign convention above, define:
For negative ρ, δ is added to 0-0 and 1-1 and removed from 1-0 and 0-1. The resulting covariance is δ, which is usually small because the adjustment is confined to four cells.
Strengths
- targets the best-known local failure of independence in low-scoring men’s football;
- retains simple Poisson marginals;
- adds only one dependence parameter;
- changes draw and BTTS probabilities without rebuilding the whole marginal model;
- is inexpensive enough for grid search, simulation and live tools.
Limitations
- only four score probabilities are modified directly;
- dependence at 2-2, 3-2, 3-3 and other scores is not modelled directly;
- the feasible range of ρ depends on the two goal means;
- the implied correlation range is limited;
- Poisson marginal mean–variance equality remains;
- a ρ fitted to one league or period may not transfer to another.
The historical Dixon-Coles framework also estimates attack, defence and home advantage with recency weights. The time-decay setting must be selected and validated as part of the estimation procedure. The four-cell correction alone does not fit those historical parameters.
See the Dixon-Coles model guide for the full specification.
Model 3: shared-component bivariate Poisson
Let three latent variables be independent:
Define the observed goals as:
The shared component X3 is added to both teams. It creates positive dependence.
Because λ3 must be non-negative, the standard shared-component construction cannot represent negative covariance.
When comparing models at fixed marginal means μhome and μaway, the shared component must also satisfy 0 ≤ λ3 ≤ min(μhome, μaway), because λ1 = μhome − λ3 and λ2 = μaway − λ3 must remain non-negative.
Joint probability formula
Setting λ3 = 0 reduces the model exactly to independent Poisson.
The total-goals distribution is no longer Poisson
Total goals are:
The mean and variance are:
Positive λ3 therefore makes total goals more dispersed than a Poisson variable with the same mean. Over/Under probabilities are no longer determined by the total expected goals alone.
The shared component cancels from goal difference
The shared component adds one goal to each side and disappears from the difference. However, if marginal means are held fixed while λ3 rises, λ1 and λ2 must fall, narrowing the goal-difference distribution and often — though not universally — increasing draw probability.
Strengths
- dependence affects the full matrix rather than only four cells;
- covariance has a direct latent-variable interpretation;
- independent Poisson is nested at λ3 = 0;
- the covariance parameter can be modelled with match covariates;
- higher-score draws and joint scoring behaviour can change.
Limitations
- standard λ3 permits only non-negative covariance;
- the shared-goal mechanism is a mathematical construction, not necessarily a literal football process;
- the likelihood is more expensive because each cell contains a finite sum;
- marginal goal counts are still Poisson;
- λ3 can be weakly identified in small samples;
- more flexible fit can become overfitting without regularization or validation.
Positive covariance does not mean that one team’s goal directly causes the other team to score. It means that unobserved or modelled match-level factors can move both goal counts together.
Worked comparison with the same marginal goal means
Use the same home and away marginal means for all three models:
The model settings are:
| Model | Parameters | Dependence |
|---|---|---|
| Independent Poisson | μh = 1.70, μa = 1.10 | Cov = 0 |
| Dixon-Coles | μh = 1.70, μa = 1.10, ρ = −0.03 | Cov ≈ 0.0034 |
| Bivariate Poisson | λ1 = 1.55, λ2 = 0.95, λ3 = 0.15 | Cov = 0.15; Corr ≈ 0.110 |
The bivariate settings preserve the same marginal means: 1.55 + 0.15 = 1.70 and 0.95 + 0.15 = 1.10. To reproduce the independent and Dixon-Coles columns, enter 1.70 and 1.10 in the xG to Fair Odds Calculator and use ρ = 0 or −0.03. That tool does not calculate the shared-component bivariate column.
Derived market probabilities
| Market | Independent Poisson | Dixon-Coles ρ = −0.03 | Bivariate Poisson λ3 = 0.15 |
|---|---|---|---|
| Home win | 51.40% | 51.06% | 51.30% |
| Draw | 24.01% | 24.69% | 25.54% |
| Away win | 24.58% | 24.24% | 23.15% |
| Over 2.5 | 53.05% | 53.05% | 52.13% |
| BTTS Yes | 54.53% | 54.87% | 55.51% |
Dixon-Coles raises draw probability by 0.68 points and leaves Over 2.5 unchanged at fixed marginal means. The bivariate model raises draw probability more broadly and also changes the totals distribution.
Selected exact scores
| Score | Independent Poisson | Dixon-Coles | Bivariate Poisson |
|---|---|---|---|
| 0-0 | 6.08% | 6.42% | 7.07% |
| 1-0 | 10.34% | 10.00% | 10.95% |
| 0-1 | 6.69% | 6.35% | 6.71% |
| 1-1 | 11.37% | 11.71% | 11.46% |
| 2-0 | 8.79% | 8.79% | 8.49% |
| 2-1 | 9.67% | 9.67% | 9.71% |
| 2-2 | 5.32% | 5.32% | 5.47% |
| 3-2 | 3.01% | 3.01% | 3.31% |
The table shows the structural distinction. Dixon-Coles changes exactly four listed cells; 2-0, 2-1, 2-2 and 3-2 remain identical to independent Poisson. The bivariate model can change probabilities throughout the score matrix, because the shared component is not restricted to four predefined cells.
This example does not rank the models. The parameters were chosen to illustrate their mechanics. Predictive superiority can be established only on unseen match data with parameters estimated before the outcomes are known.
Why totals, draws and BTTS respond differently
Independent Poisson
Total goals follow Poisson with rate μhome + μaway. Once the two means are fixed, the entire totals distribution is fixed.
Dixon-Coles
The standard correction moves equal probability mass within four Under 2.5 cells. Therefore, at fixed marginal means:
- Over/Under 2.5 is unchanged;
- draw probability changes through 0-0 and 1-1;
- BTTS changes through 1-1;
- home and away win probabilities change through 1-0 and 0-1.
Lower lines can change. For example, 0-0 is Under 0.5 while 1-0, 0-1 and 1-1 are Over 0.5. The statement that “Dixon-Coles does not affect totals” is therefore too broad; it specifically leaves Over/Under 2.5 and higher half-goal lines unchanged at fixed λ because all four affected scores lie on the same side of those lines.
Bivariate Poisson
The shared component adds goals in pairs. Total-goal variance exceeds the total-goal mean by 2λ3. Consequently, two matches with the same expected total can have different Over/Under probabilities if their covariance parameters differ.
BTTS is also no longer equal to the simple independent product:
The final term is P(0-0). Under independence with the same marginal means it would be e−(μhome+μaway); the bivariate value is larger when λ3 > 0.
Changes in draw and one-goal-margin probabilities also matter for handicap settlement. The Asian Handicap fair-odds guide explains how full wins, half wins, pushes and losses enter the price; its calculator compares independent Poisson with Dixon-Coles.
Parameter estimation and computational complexity
Independent Poisson
In a historical team model, attack, defence and home-advantage parameters are estimated by maximizing a sum of two univariate Poisson log-likelihoods per match. The calculations are straightforward and scale well.
Dixon-Coles
The log-likelihood adds log τ for the four affected scores. The original formulation also uses exponential time weights so recent matches contribute more than older ones.
Numerical fitting should stay strictly inside the λ-dependent ρ interval so every τ multiplier remains positive. The feasible interval changes with the expected goal rates.
Bivariate Poisson
Each joint probability contains a sum from zero to min(h,a). Maximum-likelihood estimation is therefore more expensive than multiplying two marginal probabilities. Direct optimization, expectation–maximization and Bayesian methods are all possible.
The shared parameter can be:
- constant across all matches;
- league- or competition-specific;
- team-specific;
- a log-linear function of match covariates;
- dynamic over time.
Each added degree of freedom requires enough data. A match-specific λ3 generated by a flexible model can fit historical scores closely while generalizing poorly.
| Task | Independent Poisson | Dixon-Coles | Bivariate Poisson |
|---|---|---|---|
| Single matrix calculation | Outer product | Outer product + four multipliers | Finite sum for every cell |
| Likelihood per match | Two simple Poisson terms | Poisson terms + log τ | Joint bivariate probability |
| Nested baseline | — | ρ = 0 | λ3 = 0 |
| Parameter constraint | Positive means | Positive means + strict score-dependent ρ bounds | λ1, λ2, λ3 ≥ 0 |
Which football score model should you use?
Use independent Poisson when:
- you need a transparent baseline;
- the goal-rate model is the main source of uncertainty;
- sample size is limited;
- residual dependence tests show little improvement from extra parameters;
- speed and reproducibility matter.
Use Dixon-Coles when:
- errors are concentrated around 0-0, 1-0, 0-1 and 1-1;
- draw probabilities are systematically misfitted;
- you want to retain Poisson marginals and simple totals behaviour;
- one low-score parameter is supported out of sample;
- the application resembles the low-scoring data for which the correction was designed.
Use bivariate Poisson when:
- dependence appears across more than four cells;
- positive covariance is empirically supported;
- higher-score draws or shared match conditions matter;
- you have enough data to estimate λ3 reliably;
- changes to the total-goals distribution are desirable rather than accidental.
Use none of them without modification when:
- marginal goals are strongly overdispersed or underdispersed;
- negative dependence is substantial across the whole matrix;
- team strength changes rapidly and the rate model is static;
- red cards, game state or live time evolution are central to the use case;
- the target sport or competition has a score pattern unlike the calibration data.
Alternatives include negative binomial marginals, Conway–Maxwell–Poisson, bivariate Weibull count models, copulas, Sarmanov extensions, state-space team-strength models and event-time hazard models. The Poisson model limitations guide maps diagnostic symptoms to the first correction worth testing.
Prefer the simplest model that survives a paired out-of-sample comparison. Extra parameters should earn their place by improving predictive scores, not by improving in-sample fit alone.
How to compare the models properly
- Use identical pre-match information. Every model must receive rates or covariates available at the same cutoff time.
- Fit on the same training period. Do not give one model more recent results or a larger sample.
- Evaluate chronologically. Test on later matches that were not used for parameter estimation or model selection.
- Score the full exact-score distribution. Use exact-score log loss or another proper score, not only the percentage of top-score hits.
- Score derived markets. Report multiclass Brier and log loss for 1X2, plus calibration for totals and BTTS.
- Use paired differences. Compare match-level losses on the same fixtures and quantify uncertainty with suitable resampling.
- Check score-specific residuals. Inspect 0-0, 1-0, 0-1, 1-1, higher draws and tail scores separately.
- Penalize complexity. Use out-of-sample scoring first; AIC, BIC or regularization can supplement it during development.
- Test stability. Refit by league, season and rolling window to see whether ρ or λ3 is transferable.
Metrics answer different questions
| Metric | What it evaluates | What it can miss |
|---|---|---|
| Exact-score log loss | Probability assigned to the realized score | Where calibration errors occur |
| 1X2 Brier score | Full home/draw/away probability vector | Exact-score structure within each result class |
| Reliability diagram | Calibration by probability range | Sharpness and overall likelihood |
| Top-k score coverage | Whether outcomes fall inside ranked score groups | Probability quality outside the cutoff |
| ROI | Economic result at offered prices | Pure probability quality; highly sample-dependent |
See football probability calibration for Brier score, log loss, reliability and out-of-sample design.
How the comparison applies to ImpliedScore
ImpliedScore currently decodes margin-free 1X2 and Over/Under odds into home and away goal rates. The public calculator then offers independent Poisson and Dixon-Coles score matrices.
| Question | Independent / Dixon-Coles implementation | Potential bivariate implementation |
|---|---|---|
| Match-level parameters | Two goal rates; optional preset ρ | Two unique rates plus shared λ3 |
| Independent market constraints | Three when both markets are supplied: two from 1X2 and one from O/U; two with 1X2 alone | The same available constraints: three with O/U, two without it |
| Fit geometry | With both markets and fixed ρ: three independent targets for two rates | Three parameters can reduce residuals, subject to positivity and model shape |
| Over/Under interpretation | Mainly anchors μhome + μaway | Depends on both marginal means and λ3 |
| Risk | Residual draw error | Using the extra parameter to absorb inconsistent market inputs without predictive validation |
A three-parameter bivariate model may reproduce 1X2 and totals more closely than a two-rate model. That is not enough to justify adding it. Because the same market probabilities are used as fitting targets, lower residual error is partly mechanical.
The relevant test is whether the resulting exact-score probabilities perform better on completed matches—particularly higher draws, BTTS, tails and exact-score log loss—than the existing Poisson and Dixon-Coles matrices.
Minimum evidence before adding bivariate Poisson
- frozen historical market snapshots with final scores;
- one documented λ3 estimation rule;
- paired out-of-sample exact-score log loss;
- calibration by score family and probability range;
- stability tests across leagues and seasons;
- evidence that improvements are not confined to the fitted 1X2 and O/U inputs.
Compare the current independent Poisson and Dixon-Coles outputs using the same market odds and inspect both fitted rates and residuals. Refitting can change λ between modes. For the fixed-rate comparison above, use the xG calculator linked in the example.
Open the football score calculatorThe inverse fitting process is explained in Market-Implied Goals from Betting Odds.
Common comparison mistakes
Calling Dixon-Coles “not bivariate”
It is a joint two-variable model. The meaningful distinction is between its local τ dependence and the shared-component bivariate Poisson construction.
Holding the wrong parameters constant
In bivariate Poisson, λ1 and λ2 are unique components, not the marginal goal means. For fixed marginal means, set λ1 = μhome − λ3 and λ2 = μaway − λ3. Subtract the shared rate from each marginal mean, not from an already adjusted unique component.
Adding λ3 without recalculating totals
The total-goals distribution is no longer Poisson. Reusing the independent Over/Under formula after introducing a shared component is inconsistent.
Assuming Dixon-Coles changes every score
At fixed λ, the standard τ correction directly changes only 0-0, 1-0, 0-1 and 1-1.
Assuming bivariate Poisson can model negative covariance
The standard shared-component parameter is non-negative. More flexible constructions are required for negative dependence.
Comparing in-sample likelihood only
A model with an extra parameter should usually fit the training data at least as well. The question is whether the improvement survives on future matches.
Changing the rate model and dependence model simultaneously
If one system uses better xG inputs and a different joint distribution, the source of any improvement cannot be identified. Change one layer at a time.
Judging models by the single most likely score
A top-ranked score is not necessarily likely in absolute terms: in the independent example above, 1-1 ranks first at 11.37%. Its probability varies with the goal rates and dependence model. Proper scoring rules evaluate the probability distribution rather than only whether one categorical pick wins.
Cover the full score space
For exact-score log loss, evaluate −log P(H = observed home goals, A = observed away goals) under each model, including observations outside the displayed grid. A display cutoff does not make those scores impossible. A single “all other scores” category defines a different, coarsened categorical loss; report it separately and use identical categories for all models. Increasing the computational grid can control truncation error, but does not turn the tail category’s probability into the probability of one exact score.
What about diagonal-inflated bivariate Poisson?
Karlis and Ntzoufras also describe a diagonal-inflated extension that adds probability mass directly to drawn scores through a normalized mixture. It is a candidate when draw residuals remain after simpler corrections, but it adds parameters and changes the marginal distributions: they are generally mixtures rather than Poisson. If the inflation weight is w and its diagonal distribution D has mean mD, the final home mean is (1 − w)(λ1 + λ3) + wmD, with the analogous away expression. Keeping the base Poisson means unchanged therefore does not generally keep the final means unchanged. Recalibrate the full mixture when comparing equal marginal means, and evaluate its entire probability distribution out of sample.
Practical conclusion
- Start with independent Poisson as the reproducible baseline.
- Add Dixon-Coles when low-score residuals justify a local correction.
- Test bivariate Poisson when dependence extends across the matrix and is predominantly positive.
- Keep marginal-goal modelling separate from dependence modelling.
- Preserve equal marginal means when making mechanical model comparisons.
- Recalculate totals and BTTS under the selected joint distribution.
- Select the final model by paired chronological out-of-sample scores.
References
- Maher, M. J. (1982). Modelling Association Football Scores . Statistica Neerlandica, 36(3), 109–118.
- Dixon, M. J. & Coles, S. G. (1997). Modelling Association Football Scores and Inefficiencies in the Football Betting Market . Journal of the Royal Statistical Society: Series C, 46(2), 265–280.
- Karlis, D. & Ntzoufras, I. (2003). Analysis of Sports Data by Using Bivariate Poisson Models . The Statistician, 52(3), 381–393.
- Groll, A., Kneib, T., Mayr, A. & Schauberger, G. (2018). On the Dependency of Soccer Scores: A Sparse Bivariate Poisson Model for the UEFA European Football Championship 2016 . Journal of Quantitative Analysis in Sports, 14(2), 65–79.
- Karlis, D. & Ntzoufras, I. (2005). Bivariate Poisson and Diagonal Inflated Bivariate Poisson Regression Models in R . Journal of Statistical Software, 14(10), 1–36.
- Michels, R., Ötting, M. & Karlis, D. (2025). Extending the Dixon and Coles Model: An Application to Women’s Football Data . Journal of the Royal Statistical Society: Series C, 74(1), 167–186.
- Gneiting, T. & Raftery, A. E. (2007). Strictly Proper Scoring Rules, Prediction, and Estimation . Journal of the American Statistical Association, 102(477), 359–378.