How Often Is the Most Likely Football Score Actually Correct?
The most likely football score is a scoreline with the highest probability in the model's complete score distribution. Rank alone does not indicate confidence: a leading score at 12% leaves 88% probability on other results. Its probability depends on the fixture and model. To assess accuracy across matches, compare observed Top-k hits with the probabilities assigned to the same pre-match score sets.
Final historical case-study snapshot. The ImpliedScore World Cup 2026 tracker contains all 104 settled tournament matches; the published Top-k summary was checked on 10 September 2026.
- The most likely score is the mode of the score probability distribution.
- The expected score is a pair of means such as 1.70–1.10; it is not necessarily a possible final score.
- Rounding expected goals does not reliably produce the modal score.
- Top-1 accuracy must be compared with the average probability assigned to each match’s top score.
- Top-3 and Top-5 include more scores; they still need probability benchmarks and uncertainty intervals. Absolute sampling variance is not determined by k alone.
- One correct or incorrect score pick says almost nothing about model quality.
What does “most likely football score” mean?
A score model assigns a probability to every possible result: 0-0, 1-0, 0-1, 1-1, 2-0, 2-1 and so on. The most likely score is the cell with the largest probability:
In statistics, this is a mode of the joint score distribution. Several scores can share that maximum. Whether one is more likely than all other results combined depends on whether its own probability exceeds 50%.
If 1-1 has 12%, 1-0 has 11%, 2-1 has 9% and every other score has less than 9%, then 1-1 is the most likely score. But the combined probability of every score other than 1-1 is still 88%.
This distinction exists because exact scores form a large categorical outcome space. Even when a model is confident about the match direction or total goals, that probability is spread across several scorelines.
Ranking and probability are different outputs. Rank #1 answers “which score is largest?” Probability answers “how large is it?” A ranking without the associated percentage is incomplete.
Most likely score versus expected score
If the home and away goal counts have expected values λhome and λaway, the mathematical expected score is:
This pair can contain decimals, such as 1.70–1.10. It describes the centre of the two marginal goal distributions, not a final score that can appear on the scoreboard.
| Concept | Example with λ = 1.70 and 1.10 | Meaning |
|---|---|---|
| Expected score | 1.70–1.10 | Mean goals for each team across repeated comparable matches |
| Rounded expected score | 2-1 | A display convention, not a probability theorem |
| Most likely score | 1-1 | The individual score with the largest joint probability |
The rounded mean differs from the mode because a Poisson distribution is right-skewed, particularly at low goal rates. For λ = 1.70, two goals are closer to the mean, but one goal is the most probable single count.
Do not label rounded λ values as “the predicted score.” A display such as 2-1 can look more certain than the model is. The matrix probability of 2-1 may be lower than 1-1, 1-0 or another score.
The most likely score under independent Poisson
Under independent Poisson scoring:
For a one-dimensional Poisson distribution with rate λ:
- if λ is not an integer, the unique mode is floor(λ);
- if λ is a positive integer m, both m − 1 and m are modes;
- if λ = 0, the only mode is 0.
Because the independent joint probability is a product, the joint mode is the Cartesian product of the two marginal mode sets.
Why floor(λ), not round(λ)?
The ratio of consecutive Poisson probabilities is:
Probabilities rise while k + 1 is below λ and fall after it passes λ. This places the mode at floor(λ), with a tie when λ is an integer.
The basic matrix construction is explained in Poisson Football Calculator and model guide.
Why is the most likely football score usually low probability?
The probability mass is divided across many plausible combinations. A balanced match may assign meaningful probability to 0-0, 1-0, 0-1, 1-1, 2-0, 0-2, 2-1, 1-2, 2-2 and several tail scores.
Higher scoring usually spreads the distribution
For large λ, the modal probability of one Poisson count is approximately:
When both goal rates are large and the teams' goal counts are independent, the leading joint probability is approximately:
As both goal rates rise, the number of plausible scores expands and the largest individual cell generally becomes smaller.
There is no universal Top-1 probability
The goal-rate examples below produce leading probabilities from 8.72% to 24.66%. They illustrate different scoring environments, not a measured range for every league. Very low-scoring fixtures can have an even higher 0-0 probability.
Team imbalance can also increase concentration. If one side has a very low scoring rate, several leading outcomes may share zero goals for that team.
How λ changes Top-1, Top-3 and Top-5 probability
The examples below use independent Poisson and rank individual score cells by probability. They are mathematical illustrations, not typical values for every league.
| λ home–away | Rounded means | Modal score | Top-1 probability | Top-3 probability | Top-5 probability |
|---|---|---|---|---|---|
| 0.70–0.70 | 1-1 | 0-0 | 24.66% | 59.18% | 77.31% |
| 1.35–1.35 | 1-1 | 1-1 | 12.25% | 30.39% | 46.93% |
| 1.70–1.10 | 2-1 | 1-1 | 11.37% | 31.37% | 46.85% |
| 2.10–1.60 | 2-2 | 2-1 | 8.72% | 24.01% | 36.76% |
| 2.80–0.60 | 3-1 | 2-0 | 13.08% | 34.64% | 51.03% |
Example: λ = 1.70–1.10
| Rank | Score | Probability | Cumulative coverage |
|---|---|---|---|
| 1 | 1-1 | 11.37% | 11.37% |
| 2 | 1-0 | 10.34% | 21.71% |
| 3 | 2-1 | 9.67% | 31.37% |
| 4 | 2-0 | 8.79% | 40.16% |
| 5 | 0-1 | 6.69% | 46.85% |
Even the five highest scores cover less than half of the distribution. A final score outside the displayed top five is therefore not surprising.
How Top-1, Top-3 and Top-5 accuracy should be measured
Let pi,(1), pi,(2), … be the score probabilities for match i, sorted from highest to lowest.
The probability that the final score falls inside the top k is:
Define yi,k as 1 when the actual score is inside the top k, and 0 otherwise. Across N matches:
A calibrated model should produce observed coverage close to expected coverage over a large enough sample.
Why “18 correct out of 100” is incomplete
Eighteen Top-1 hits can be excellent, ordinary or poor depending on the probabilities.
| Average Top-1 probability | Expected hits in 100 matches | Interpretation of 18 observed hits |
|---|---|---|
| 8% | 8 | Far above the probability benchmark |
| 14% | 14 | Moderately above the benchmark |
| 20% | 20 | Slightly below the benchmark |
Top-k coverage is a marginal consistency diagnostic for ranked prediction sets. Aggregate agreement is necessary, but it does not establish calibration across probability ranges or match types, and it is not a proper scoring rule for the entire distribution. Two models can have identical Top-5 coverage while assigning very different probabilities to the actual score.
What if two or more scores are equally likely?
Exact ties occur naturally under independent Poisson when one or both λ values are integers. Near-ties are common even when the values are not integers.
| λ home–away | Joint modal scores | Probability of each mode | Combined modal-set probability |
|---|---|---|---|
| 1.00–1.00 | 0-0, 0-1, 1-0, 1-1 | 13.53% | 54.13% |
| 2.00–1.30 | 1-1, 2-1 | 9.59% | 19.18% |
| 3.00–3.00 | 2-2, 2-3, 3-2, 3-3 | 5.02% | 20.08% |
A software interface that displays only one of four equal modes creates an arbitrary “winner” determined by sorting order.
Recommended tie policy
- Use probability-aware ranks. For display, define a shared probability rank as one plus the number of scores with strictly greater probability. This shared rank is not a unique position in a fixed-size list.
- Show ties explicitly. Label mathematically equal modes as sharing rank #1. Use a documented numerical tolerance for computation artifacts; do not merge genuinely different probabilities just because the printed percentages round to the same value.
- Separate display from evaluation. A fixed five-row table can use a stable tie-breaker; evaluation must then apply the same tie-breaker, so that expected and observed coverage refer to identical cells.
- Keep Top-k as exactly k cells. Break exact ties with a predeclared stable rule (for example, fewer home goals first, then fewer away goals), apply the same rule to the expected and observed sides, and report tied modal-set coverage separately when a tie touches first place or a Top-k boundary.
The boundary policy is not cosmetic. At λ = 1.00–1.00, a strict three-cell Top-3 covers 40.60%, while including all four tied modes covers 54.13% — the gap comes entirely from the tie rule, so the same rule must feed both the expected and the observed side of the comparison.
Floating-point equality needs tolerance. Scores that should be mathematically equal can differ by tiny numerical amounts. Ranking code should not turn a rounding artifact into a meaningful ordering.
World Cup 2026: observed versus expected score coverage
The ImpliedScore tracker froze a pre-match score matrix for each tournament match, then recorded the final regulation-time score, its probability and its matrix rank. The completed historical sample contains all 104 matches.
17.3% / 13.7%
Top-1 observed / expected
39.4% / 36.4%
Top-3 observed / expected
53.8% / 53.5%
Top-5 observed / expected
104
settled matches
Complete
tournament sample
| Metric | Observed | Expected | Gap |
|---|---|---|---|
| Top-1 | 17.3% | 13.7% | +3.6pp |
| Top-3 | 39.4% | 36.4% | +3.0pp |
| Top-5 | 53.8% | 53.5% | +0.3pp |
Does 17.3% prove that the top score is unusually accurate?
No. The completed sample contains 18 Top-1 hits in 104 matches. If every match had a common Top-1 probability of 13.7%, a simple binomial calculation gives a rough one-sided reference probability of about 17.5% for observing at least 18 hits. Because the Top-1 probability differs by match, the formal predictive interval and tail probability should be computed from the full Poisson-binomial distribution built from the per-match probabilities — E[X] = Σpi, Var(X) = Σpi(1 − pi) under independent match-success events. Dependence requires a joint model and covariance terms. The same distinction applies to Top-3 and Top-5.
As a descriptive common-rate approximation, the 95% Wilson interval around 18 hits in 104 matches is roughly 11.2% to 25.7%, and the expected 13.7% rate lies inside it. One additional hit in a 104-match sample changes observed coverage by about 0.96 percentage points, so small observed-versus-expected gaps remain noisy.
For the related question of how likely several correct exact-score picks are across a series of matches, see the multiple correct scores analysis.
Final sample reading: Top-1, Top-3 and Top-5 all finished above their mean expected coverage, but the gaps are small relative to the sampling uncertainty of a 104-match tournament. This sample alone is not enough to establish persistent score-model bias.
Examples from the tracker
| Match | Top score before kickoff | Final score | Actual rank | Actual probability |
|---|---|---|---|---|
| France v Iraq | 3-0 at 13.1% | 3-0 | #1 | 13.1% |
| Mexico v South Africa | 1-0 at 16.4% | 2-0 | #2 | 15.0% |
| Republic of Korea v Czechia | 1-1 at 14.5% | 2-1 | #5 | 8.0% |
| England v Croatia | 1-0 at 13.7% | 4-2 | #18 | 0.9% |
All four examples are compatible with the same probabilistic system. A first-ranked score can occur, the second-ranked score can occur, or a low-probability tail result can occur.
Inspect every match’s pre-match top scores, final score, actual rank and expected coverage.
Open the completed World Cup score trackerTop-score accuracy is not enough to validate a model
Top-1 evaluates only whether the largest category occurred. It discards the remaining probabilities.
| Forecast | Top score | Actual score probability | Top-1 result |
|---|---|---|---|
| Model A | 1-1 at 12% | 2-1 at 11% | Miss |
| Model B | 1-1 at 25% | 2-1 at 1% | Miss |
Both models miss Top-1, but Model A assigned substantial probability to the actual score while Model B was strongly wrong.
Use probability assigned to the actual score
For match i, record pi,actual. If outcomes are generated by that model's own distribution, the expected probability assigned to the realised score is its collision benchmark:
Mean probability assigned to the realised score is a descriptive concentration diagnostic, not a proper scoring rule and not a standalone model-ranking criterion. Its observed average and collision benchmark must be calculated on the same frozen match sample; the final Top-k summary above does not by itself determine this metric.
Use exact-score log loss
Log loss rewards probability assigned to the score that occurs and penalizes assigning it very small probability. For exact-score log loss, evaluate the probability of the actual score under the frozen model even if that score lies outside the displayed grid. A finite display cutoff must not turn a possible score into a zero-probability event.
If the stored forecast instead combines all omitted scores into one residual category, use that category for the outcome and label the result as log loss on the coarsened categories. Its probability is the sum of many scores, so this is a different metric from exact-score log loss. Keep the same category boundaries when comparing models.
Use calibration by probability range
Scores assigned 10–15% should occur about 10–15% of the time. Scores assigned 1–2% should still occur occasionally in a large sample. Grouping exact scores by their forecast probability tests the distribution more directly than a fixed rank cutoff.
Keep a frozen out-of-sample process
Probabilities must be stored before kickoff. Model choice, ρ values and λ-fitting rules must not be changed retroactively for completed matches.
The complete evaluation framework is described in football probability calibration, Brier score and log loss.
Can Dixon-Coles change the most likely score?
Yes. Dixon-Coles changes the probabilities of 0-0, 1-0, 0-1 and 1-1. If one of these scores is already near the top, the correction can change the ranking.
For fixed λ values and negative ρ:
- 0-0 and 1-1 increase;
- 1-0 and 0-1 decrease;
- all other score cells remain unchanged directly.
If independent Poisson gives 1-0 at 12.0% and 1-1 at 11.8%, a sufficiently negative ρ can lower 1-0 and raise 1-1, making 1-1 the new mode. The λ values have not changed; only the local dependence structure has.
When λ is re-fitted with Dixon-Coles active, more scores can shift indirectly because the optimized goal rates may also change.
See the Dixon-Coles model guide and the football score model comparison.
Common misinterpretations of the most likely score
“The model predicted 1-1, so every other score is a failure”
The model predicted a distribution. Rank #1 is a summary of that distribution, not the only result the model considered plausible.
“A 12% top score means the model is only 12% accurate”
It means that exact Top-1 coverage should be near 12% if the model is calibrated. Other tasks—1X2, totals or BTTS—can have much higher accuracy because they aggregate many score cells.
“Top-3 should be three times Top-1”
The second and third scores usually have lower probability than the first. Top-3 coverage is the sum of the three actual probabilities, not three times the leader.
“The expected score is the most likely score”
The expected score uses mean goals. The most likely score uses the mode. They often differ.
“A score outside the top five disproves the model”
If the top five carry 54%, the model explicitly assigns 46% to outcomes outside that list.
“The mode with the higher probability is always a useful bet”
Betting value depends on the offered price relative to the probability, not on rank alone. Correct-score markets also tend to contain substantial bookmaker margin.
“More Top-1 hits always mean a better probability model”
A distorted model can concentrate probability too aggressively and produce more hits in one sample while performing worse on log loss and calibration.
How to use a most-likely-score ranking correctly
- Read probability before rank. Distinguish a 24% mode from an 8% mode.
- Inspect cumulative probability. Top-3 and Top-5 show the mass in those score sets. For axes, colours and outside-grid probability, use the score probability map guide.
- Compare neighbouring scores. A 12.1% leader over an 11.9% runner-up has a small probability advantage. It is still a different probability, so keep the predeclared ordering when evaluating fixed-size sets.
- Check the expected goal rates. λ explains why particular scores occupy the top of the matrix.
- Check the model. Poisson and Dixon-Coles can rank low scores differently.
- Do not infer certainty from a single label. “Most likely” only identifies the maximum cell.
- Evaluate over many frozen matches. Compare observed and expected coverage, then use proper scoring rules.
Enter 1X2 and Over/Under odds to view the complete score matrix, the probability of each ranked score and cumulative market probabilities.
Open the football score probability calculatorPractical conclusion
- The most likely football score is a mode, not a confident point prediction.
- The mode's probability varies by fixture and model; use its percentage rather than a generic accuracy rate.
- The rounded expected score can differ from the modal score.
- Observed Top-k coverage should be compared with summed Top-k probability.
- Tied and near-tied scores need explicit ranking rules.
- Top-1 accuracy should be supplemented with actual-score probability, log loss and calibration.
References and data
- ImpliedScore. World Cup 2026 Market-Implied Score Tracker (linked above). Pre-match score rankings, actual-score overlays and Top-1/Top-3/Top-5 calibration.
- ImpliedScore. Methodology: From Odds to a Score Probability Matrix.
- 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.
- Gneiting, T. & Raftery, A. E. (2007). Strictly Proper Scoring Rules, Prediction, and Estimation . Journal of the American Statistical Association, 102(477), 359–378.
- Foulley, J.-L. (2021). More on Verification of Probability Forecasts for Football Outcomes . arXiv:2106.14345.
- Reade, J. J., Singleton, C. & Brown, A. (2021). Evaluating Strange Forecasts: The Curious Case of Football Match Scorelines . Scottish Journal of Political Economy, 68(2), 261–285.