How to Estimate 0-0 Probability in Football

The standard starting point for estimating a 0-0 probability is an independent Poisson score model. Estimate the home and away teams’ expected goal rates, add them together, and calculate the probability that the match produces zero goals. The arithmetic is simple; the modelling choices behind the two scoring rates determine whether the answer is useful.

A defensible estimate must account for competition scoring levels, home advantage, team attack and defence, opponent strength, recency and relevant pre-match information. It should also acknowledge that football scores are not always perfectly independent. The objective is not to manufacture certainty, but to produce a probability that can be tested, calibrated and updated.

For a related perspective, see match prediction framework.

The direct Poisson estimate

Let λH represent the home team’s expected number of goals and λA the away team’s expected number of goals. These are scoring intensities, not probabilities. A home rate of 0.90 means that the model expects an average of 0.90 home goals across many matches played under comparable conditions.

Under a Poisson model, the probability that a team with scoring rate λ records zero goals is e^−λ. If the two goal counts are conditionally independent, the joint 0-0 probability is:

P(0-0) = e^−λH × e^−λA = e^{−(λH + λA)}

This produces an important simplification: in the basic model, the 0-0 estimate depends only on the total expected-goal rate. Rates of 0.80 and 0.80 give the same 0-0 probability as rates of 1.30 and 0.30 because both pairs sum to 1.60. Their probabilities for 1-0, 0-1 and other scores differ, but their basic 0-0 probabilities are identical.

For a related perspective, see Correct Score Prediction: A Practical Probability Framework.

The independence assumption is conditional. It does not claim that the clubs or their historical scores are unrelated. It says that, after the model has accounted for team strength, venue and other available information, the remaining home and away goal counts are treated as independent. Omitted match conditions or tactical interaction can make that assumption too strong, which is why dependence adjustments may be required later.

Estimating the home and away scoring rates

The largest source of error is usually not the Poisson calculation. It is the estimation of λH and λA. Raw recent goal averages are unstable because they mix opponent quality, finishing variance, red cards, penalties and changing line-ups.

A typical model estimates the rates on a logarithmic scale:

log λH = baseline + home advantage + home attack + away defence + match adjustments

For a related perspective, see Expected Goals (xG) for Correct Score Analysis.

log λA = baseline + away attack + home defence + match adjustments

The signs assigned to defensive parameters depend on the model’s convention. The essential point is that each rate combines a competition-level baseline with the relevant attacking and defensive strengths.

Competition and venue baseline

Start with the scoring environment in which the match is played. Different competitions, seasons and home-away settings can have different goal rates. A model trained in one environment should not be transferred mechanically to another without recalibration.

Attack and defence estimates

Team effects should be adjusted for opponent quality. Scoring twice against a weak defence is not equivalent evidence to scoring twice against a strong one. Hierarchical or regularised models are useful because they shrink noisy team estimates towards the competition average. This matters especially early in a season, after promotion or when only a small sample is available.

Goals, expected goals and recency

Goals are the target outcome but contain substantial finishing and goalkeeping variance. Expected-goal data can provide a less outcome-dependent measure of chance creation and concession, although values from different expected-goal providers are not necessarily interchangeable. A pre-match model forecasts future chance quality; it should not simply reuse post-match expected goals from recent fixtures as its forecast.

Recent matches can receive greater weight, but aggressive time decay makes the estimate react to noise. Line-ups, tactical changes, rest and weather can be included when reliable information exists. Such adjustments should be estimated from data where possible rather than imposed as arbitrary goal deductions. The same information should not be counted twice through both recent performance and a separate manual adjustment.

A repeatable pre-match calculation

A practical estimation process can be applied in five steps:

1. Set the information cut-off. Use only information genuinely known before kick-off.
2. Estimate λH and λA. Generate opponent-adjusted scoring rates from a fitted model, or from a transparent baseline-and-strength procedure.
3. Add the rates. Define the total scoring intensity as S = λH + λA.
4. Calculate e^−S. This is the independent-Poisson probability of 0-0.
5. Apply and validate any correction. If using a low-score dependence model, estimate its parameters from training data and test the corrected probabilities out of sample.

The output should be retained as a probability rather than rounded immediately into a verbal label. An estimate of 0.19 means that the model assigns approximately 19 chances in 100 to 0-0 under comparable conditions. It simultaneously assigns about 81 chances in 100 to every other score combined.

Decimal fair odds can be calculated as 1 divided by the probability, but this is only a probability representation. It does not include a bookmaker margin, transaction costs, model uncertainty or compensation for estimation error.

When independent Poisson is too simple

Football matches contain interactions that a pair of independent Poisson distributions cannot fully represent. A team’s approach can depend on the score, and both teams may respond to the same pitch, weather, officiating or tactical conditions. Cautious teams can preserve a low-event state, while an early goal can change both teams’ subsequent scoring hazards.

The Dixon-Coles adjustment is a common extension. It retains Poisson-based scoring rates but modifies the probabilities of 0-0, 1-0, 0-1 and 1-1 using a fitted dependence parameter. Under the commonly used Dixon-Coles convention, the 0-0 cell is multiplied by 1 − λHλAρ, where ρ is the fitted low-score parameter. A negative ρ therefore increases the 0-0 probability under that convention; a positive ρ decreases it. Parameter conventions should always be documented, and the correction must be estimated rather than selected to produce a preferred result.

A bivariate Poisson model introduces a shared latent scoring component. This creates explicit positive covariance between the score counts and can alter the joint-zero probability even when the marginal goal expectations are similar. Its dependence structure remains restrictive, so greater mathematical complexity does not automatically imply better calibration.

A time-based simulation can model changing hazards, red cards, substitutions and game state more directly. It is useful when those mechanisms are central to the forecast, but it requires more parameters and substantially more data. For a pre-match 0-0 estimate, independent Poisson remains a useful benchmark. An extension is justified only if it improves held-out likelihood, probability calibration or another declared out-of-sample criterion.

Sensitivity to the total expected-goal rate

Because P(0-0) = e^−S, small changes in the total scoring rate can materially affect the forecast. Increasing S by 0.10 multiplies the 0-0 probability by e^−0.10, approximately 0.905. That is a relative reduction of about 9.5%, regardless of the starting rate.

This sensitivity explains why apparently minor modelling decisions matter. A line-up adjustment, a change in competition baseline or a different weighting of recent matches may move the total rate by only a few tenths, yet produce a noticeably different 0-0 probability.

In the independent model, changing the allocation of a fixed total between the teams does not change P(0-0). Allocation does matter for the rest of the score distribution and can also affect a Dixon-Coles correction. A full correct-score model therefore still needs credible separate home and away rates rather than only a total-goals forecast.

Calibrating the probability rather than trusting the formula

A mathematically correct transformation can still produce poorly calibrated forecasts if the input rates are biased. Calibration asks whether events assigned a given probability occur at approximately that frequency over a large set of genuine pre-match predictions.

Store each forecast before the match and define a binary outcome: one for 0-0 and zero otherwise. Group forecasts into reasonably broad probability bands, then compare the mean forecast with the observed 0-0 frequency in each band. The bands must not be so narrow that each contains only a handful of events.

Brier score and log loss can evaluate the individual probabilities. Brier score measures squared probability error, while log loss penalises confident mistakes more severely. Both should be compared with simple benchmarks, such as a competition-level base rate or an unadjusted Poisson model. A complex model that cannot outperform those baselines on later matches has not demonstrated useful improvement.

Evaluation should respect time order. Randomly mixing old and new fixtures can hide changes in team strength or scoring environment. A rolling or expanding-window test better represents the actual task: fit using past information, predict the next period, and repeat.

Parameter uncertainty should also be propagated. Bootstrap samples or posterior draws can generate multiple plausible values of λH and λA, with a 0-0 probability calculated for each draw. The resulting distribution shows how uncertain the estimate is. As an illustrative calculation, if the plausible total scoring rate ranges from 1.50 to 1.90, the corresponding independent-Poisson probability ranges from 22.3% down to 15.0%. Reporting only a central estimate would conceal that sensitivity.

Rare-event calibration requires patience. If an illustrative test set contains 200 forecasts averaging 10%, it contains only about 20 expected 0-0 outcomes. A difference of a few matches can then move the observed frequency substantially. Apparent overperformance or underperformance in a small sample may be ordinary variance rather than evidence of model quality.

Interpretation, cross-checks and common errors

The estimated probability is a model output conditional on assumptions. It is not a claim that 0-0 is the most likely outcome, and it is never evidence that a result is fixed or guaranteed. Even a relatively high 0-0 probability normally leaves most probability mass on other scores.

Useful consistency checks can expose mistakes. The 0-0 probability cannot exceed either team’s modelled clean-sheet probability. It is exactly the probability of under 0.5 total goals. Under 1.5 goals also includes 1-0 and 0-1, while both teams to score: no includes every scoreline in which at least one team scores zero. Probabilities across correct-score, over-under and both-teams-to-score outputs should be generated from the same score distribution if they are intended to be internally coherent.

Common estimation errors include:

• Using league position as a direct substitute for attacking and defensive rates.
• Averaging recent goals without adjusting for opponent quality, venue or red cards.
• Multiplying historical clean-sheet percentages as if they were stable and independent.
• Applying a subjective injury adjustment on top of a model that already incorporates the affected matches.
• Fitting a dependence correction to a small sample and treating the fitted parameter as permanent.
• Reporting excessive decimal precision despite substantial parameter uncertainty.
• Selecting a model because it explains past 0-0 results without testing later matches.

Market comparison is a separate interpretation stage. The reciprocal of offered decimal odds is a raw implied probability, but correct-score markets can contain substantial and uneven margins. A fair comparison requires an explicit method for removing that margin and should allow for uncertainty in the model estimate. A numerical difference between model and market is not automatically a reliable edge.

Nurul Adira

Best Bet & Sure Bet Analysis · 10 years experience

Read full author profile →