How to Analyse Correct Scores With Probability

Define the outcome before looking for a pattern

Let H represent home goals and A represent away goals. A correct-score model estimates P(H=h, A=a) for each possible pair of non-negative goal counts h and a. Taken together, these cells form a joint probability distribution: outcomes are mutually exclusive and the probabilities across the complete, unbounded score space sum to one.

This definition distinguishes several questions that are often conflated:

For another angle on the same subject, see football methodology.

• Exact score: What is the probability of one cell, such as 1-0?
• Result class: What is the combined probability of a home win, draw or away win?
• Goal pattern: What is the probability of both teams scoring, a particular total, a clean sheet or a winning margin?
• Observed frequency: How often did a score occur in a selected historical sample?

Observed frequency is evidence about the selected past sample, not automatically a probability for the next match. Three consecutive 1-0 results may be consistent with stable defensive strength, but they may also reflect opponent quality, venue mix, finishing variation, dismissals, penalties or ordinary sampling noise. The sequence does not identify its own cause.

The first methodological safeguard is to state the prediction target and the information cutoff. If the target is the next match's exact score, the model should estimate scoring rates using only information that was available before that match. Counting historical scorelines without controlling for opposition, venue, changing team quality or selection effects addresses a different and generally weaker question.

Build a baseline from expected scoring rates

A common baseline models each team's goal count with a Poisson distribution. Let λH be the expected home-goal rate and λA the expected away-goal rate. Under the simplest specification, the two goal counts are conditionally independent given those rates:

P(H=h, A=a) = P(H=h) × P(A=a), where P(H=h) = exp(-λH) × λH^h / h!, with the away-goal probability calculated equivalently.

For a related perspective, see How to Analyse Banker Picks Using Probability and Value.

The parameters are expected scoring rates, not predicted final scores. A home rate of 1.60 does not mean the home team will score exactly 1.6 goals. It defines probabilities for zero, one, two and more goals.

Consider a fictional illustration with λH = 1.60 and λA = 0.90. Under the independent Poisson assumption, the joint probability of a scoreline is obtained by multiplying the relevant marginal probabilities. The resulting approximate probabilities are:

Here, 1-0 is the modal scoreline, meaning it has the largest individual probability. It is nevertheless expected in only about 13 matches out of 100 with equivalent inputs. A modal score should therefore not be interpreted as a high-certainty forecast.

A displayed score matrix may stop at 5-5 or 6-6 for readability, but the underlying distribution should retain the remaining tail probability. Otherwise the visible cells do not sum to one and derived markets can be distorted. The omitted mass can be reported as a tail category, or the displayed range can be extended until the remaining probability is immaterial for the stated purpose.

Derive patterns from the same probability matrix

Once a coherent score matrix exists, broader outcomes can be calculated by adding the relevant cells. Home-win probability is the sum of all cells where h is greater than a. Draw probability is the sum along the diagonal where h equals a. The probability of over 2.5 total goals is the sum of all cells where h+a is at least three.

Both teams to score can be derived from the same joint distribution:

P(BTTS) = 1 - P(H=0) - P(A=0) + P(H=0, A=0).

The final term is added back because a 0-0 score is removed twice by the two preceding zero-goal terms. Other summaries can include clean sheets, winning margins, one-goal games, scoreless draws and high-scoring draws. These aggregations show why an exact score should not be interpreted in isolation. A model can rank 1-1 as its largest individual cell while still assigning a greater combined probability to all home-win scorelines than to all draw scorelines.

Aggregation also provides a coherence check. If a separately generated totals model materially conflicts with the total-goals probabilities implied by the correct-score matrix, then the models are using different assumptions, are calibrated differently, or one of them is misspecified. Separate models need not produce identical estimates, but unexplained differences should not be presented as independent confirmation.

Pattern labels should remain operationally defined. “Low scoring” could mean under 2.5 goals, a total expectation below a stated threshold, neither team projected above one goal, or some other rule. If the definition changes after the result is known, it cannot be evaluated consistently.

Estimate goal rates without confusing signal and noise

The score matrix is only as credible as λH and λA. These rates can be estimated from historical goals, shot-based expected-goals data, attack and defence ratings, venue effects, rest periods, squad information or combinations of such variables. A common structure estimates a team's attacking strength, an opponent's defensive strength and a competition-level home advantage.

Each input has limitations. Raw goals are volatile because finishing and goalkeeping outcomes contain substantial variation over short periods. Shot-based expected-goals measures may reduce some finishing noise, but their usefulness depends on the event-data provider, the shot model, the consistency of definitions and whether the data capture relevant aspects of chance creation. Neither observed goals nor expected goals should be assumed to be a complete measure of future scoring ability.

Time weighting creates a trade-off. Long historical windows increase sample size but may describe a team that no longer exists in the same form because of managerial changes, tactical shifts, transfers, injuries or promotion and relegation. Short windows respond more rapidly but can overreact to random finishing runs and uneven fixtures. The weighting rule should be specified before evaluating results rather than selected because it explains a particular past sequence.

Opponent adjustment is essential. Scoring twice per match against weak defences is not equivalent evidence to scoring twice against strong defences. Venue, competition context and schedule imbalance can also confound simple recent-form measures. A team that has faced several unusually weak opponents may appear to have improved even if its underlying attacking rate has not changed.

Pre-match information should be included only when it can be recorded consistently and was genuinely available at prediction time. Injury reports and expected line-ups may be informative, but they introduce uncertainty, subjective judgement and potential hindsight bias. If historical line-up variables are reconstructed after confirmed team news while future forecasts rely on uncertain expected line-ups, the backtest has access to information that would not have been available in live use.

A defensible model documents data timestamps, source definitions, weighting rules, opponent adjustment, missing-data treatment and parameter-update rules. The proposition that an additional variable improves correct-score forecasting should be tested against a simpler baseline on later matches. Plausibility alone is not evidence that a variable improves probability estimates.

Challenge independence, dispersion and match-state assumptions

The independent Poisson model is useful because it is transparent and provides a reproducible baseline. Its assumptions should not be mistaken for facts. In particular, conditional independence means that once λH and λA are known, one team's goal count supplies no further information about the other's. That can be restrictive in football.

Match state can connect the two scoring processes. A goal changes incentives, tactics and risk tolerance: a leading team may protect its advantage, while a trailing team may increase attacking risk. Red cards, penalties, weather, refereeing decisions and early injuries can also shift both teams' scoring prospects during a match. A static pre-match model does not directly represent these paths.

Observed goal counts may also show more or less variability than the Poisson distribution permits. Under a Poisson distribution, the variance equals the mean. Overdispersion occurs when the observed variance is greater than the mean, potentially because matches arise from different latent game environments or because team strength is imperfectly measured. Underdispersion is the opposite case. These properties should be assessed empirically rather than presumed from theory.

Possible extensions include:

• Low-score corrections: Adjust selected cells, commonly 0-0, 1-0, 0-1 and 1-1, when a baseline displays systematic errors in those outcomes.
• Bivariate goal models: Introduce a shared latent component or another dependence structure between the teams' goal counts.
• Negative-binomial models: Allow greater variance than a standard Poisson model when overdispersion is supported by the data.
• Mixture models: Represent uncertainty over distinct match regimes, such as lower-tempo and higher-tempo environments.
• Dynamic-strength models: Allow attacking and defensive parameters to change over time rather than treating team quality as fixed.

These are not automatic improvements. A flexible specification can fit prior scorelines more closely while making less accurate forecasts on new matches. A low-score adjustment may be specific to a competition, period, data source or model design. A dependence mechanism that improves draw calibration may worsen estimates in higher-scoring cells. The relevant comparison is forward predictive performance, including calibration, not in-sample fit alone.

Validate probabilities, not just winning selections

Exact-score hit rate is a weak primary evaluation measure. It gives full credit when a selected score occurs and no credit otherwise, even if a model assigned sensible probability across nearby and alternative outcomes. It also rewards concentrating attention on the first-ranked cell while ignoring the quality of the remaining distribution.

Proper scoring rules assess the full probability forecast. Log loss penalises forecasts that assign very low probability to the score that occurs. A multiclass Brier score measures squared differences between the predicted probabilities and the one-hot realised outcome vector. These measures reward well-calibrated uncertainty rather than only the ranking of one selected score. Distance-aware measures can be useful if the analytical purpose treats a 2-1 forecast as more similar to 1-1 than to 5-0, but that choice introduces a judgement about the cost of different errors and should be specified before testing.

Calibration asks whether stated probabilities correspond to observed frequencies over many forecasts. For example, events assigned roughly 10% probability should occur roughly 10% of the time within an adequately large and comparable evaluation set. Individual exact scores are sparse, so direct cell-level calibration is often imprecise. More stable diagnostics may group forecasts by probability band or examine broader events such as draws, clean sheets, both teams to score and total-goals thresholds.

Discrimination and calibration are distinct. A model can rank outcomes reasonably well while being systematically overconfident or underconfident. Conversely, a heavily shrunk model can appear calibrated on broad events while failing to distinguish meaningful match-to-match differences. Both properties matter when the objective is a usable probability distribution.

A credible backtest should use time-ordered evaluation:

1. Estimate parameters using only information available before the test match or test period.
2. Generate the full score distribution for each test match.
3. Advance through time and update the model only under a pre-declared update rule.
4. Compare performance with simple benchmarks, such as league-average rates or a basic attack-defence Poisson model.
5. Report aggregate results alongside errors across relevant subgroups and time periods.

Randomly mixing old and new matches can leak future information about team strength into the training set. Repeatedly changing the model after reviewing one test period turns that period into de facto training data. A final untouched holdout period, or a disciplined rolling-origin design with limited model selection, reduces but does not eliminate this risk.

Uncertainty remains after validation. Rare exact scores require large samples for stable assessment, and average performance can conceal systematic errors for promoted teams, late-season matches, matches with incomplete team news or competitions with limited data. Confidence intervals, bootstrap procedures and evaluation across multiple periods can help distinguish persistent improvement from ordinary sampling variation.

Separate score probability from price and decision

A probability estimate and a decision based on a quoted price are different objects. If a model assigns probability q to an exact score and the available decimal odds are d, the simplified expected net return per unit staked is q × d - 1. A positive value depends jointly on the estimated probability and the available price. Identifying the modal scoreline alone does not establish that any price is favourable.

This expression is simplified because it treats q as known. In practice, q is estimated with error. Correct-score probabilities are especially sensitive to small changes in goal-rate inputs, and a nominal advantage can disappear when λH, λA, line-up assumptions or model parameters move slightly. Point estimates can therefore exaggerate precision.

Quoted market odds usually include a margin. Converting each listed price into 1/d and treating the result as a fair probability will generally produce a total above 100%. Comparisons should account for the overround, the treatment of suspended or unavailable prices, and grouped categories such as “any other home score”, which combine multiple tail outcomes. A score matrix that omits these grouped outcomes is not directly comparable with a complete market book.

Market probabilities can serve as a demanding benchmark because prices may reflect widely available information. However, agreement with the market does not validate a model, and disagreement does not prove that the model has identified an error. Differences can arise from timing, margins, different score truncation rules, differing assumptions, stale inputs or model misspecification. They are signals to investigate rather than conclusions in themselves.

Use sensitivity analysis to identify fragile score calls

Sensitivity analysis varies plausible modelling assumptions and records how the forecast changes. Useful tests can include altering recency weights, comparing goals-based and shot-based inputs, changing the home-advantage treatment, excluding an uncertain line-up variable, varying a prior or shrinkage assumption, and comparing independent Poisson with an empirically justified low-score correction.

The purpose is not to search until a preferred scoreline appears. Assumptions should be varied within ranges that are substantively plausible and documented in advance where possible. If a forecast changes only under unrealistic inputs, that is different from a forecast that changes under small, defensible variations in key assumptions.

Ranking instability deserves particular attention. Suppose 1-0, 1-1 and 2-0 have similar estimated probabilities. A small change in λH or λA may reorder those scores without materially changing the broader interpretation that the match has relatively low expected scoring. Reporting only the top-ranked cell can make the model appear more decisive than the underlying probabilities justify.

A proportionate interpretation can therefore report the modal score, nearby alternatives, aggregate result probabilities and the sensitivity of those outputs to reasonable assumptions. If the leading score remains first across relevant specifications, it has some ranking robustness. If several cells exchange places, the appropriate inference is that exact-score resolution is weak, not that one specification should be selected retrospectively because it produces the preferred answer.

The workflow is iterative: define the target, establish the information cutoff, estimate scoring rates, construct the complete matrix, derive related outcomes, validate calibration and scoring performance, compare with benchmarks, inspect subgroup errors and stress-test the assumptions. Confidence should come from reproducible forward performance and transparent limitations rather than a short sequence of correct score calls.