What Are the Odds of Predicting Multiple Correct Football Scores?
To calculate the probability of predicting multiple correct football scores across different matches, multiply the selected score probabilities when the matches are independent. For example, three picks at 15%, 12% and 10% have a combined probability of 0.18%, about 1 in 556. The calculator below also gives the chance of exactly or at least k correct picks. If you select several alternative scores for one match, add their probabilities instead.
- If independent match probabilities are p1, …, pn, the chance all are correct is Πpi.
- If every pick has the same probability p, the number correct follows Binomial(n,p).
- If probabilities differ, use the Poisson-binomial distribution.
- Expected correct picks equal Σpi; expectation is not a guaranteed result.
- At an illustrative 12% per match, six correct scores in a row have probability 0.000299%, about 1 in 334,898.
- Top-3 score coverage can be analysed with the same formulas by replacing each Top-1 probability with the sum of its top three scores.
- Multiplication assumes conditional independence; common teams, tournaments and model errors can create dependence.
First define what counts as a correct score prediction
“Correct score” can describe several different tasks. Their probabilities are not interchangeable.
| Task | Success condition | Per-match probability to use |
|---|---|---|
| One exact Top-1 score | The single selected score occurs | Probability of that exact score |
| Top-3 score coverage | The result is one of three selected scores | Sum of the three mutually exclusive score probabilities |
| Top-5 score coverage | The result is one of five selected scores | Sum of the five score probabilities |
| Correct 1X2 outcome | Home, draw or away category is correct | 1X2 probability, not exact-score probability |
| Correct score after extra time | Score under a specified extra-time rule | A separate distribution including extra time |
Use the settlement period specified by the market. For a regulation-time market, include stoppage time and exclude extra time and penalty shootouts.
Do not use one universal “chance of predicting a football score.” A 0-0 pick in a low-total match can have more than 20% probability, while a high-scoring match’s modal score can be below 9%.
The difference between rank and probability is explained in how often the most likely football score is correct.
Probability that all selected scores are correct
Let event Ci mean that the score pick for match i is correct. If the match outcomes are independent conditional on the information and model state used to produce the probabilities:
For three picks with probabilities 15%, 12% and 10%:
The corresponding one-in-X figure is:
The product becomes small rapidly because every additional match must also succeed.
Logarithms for long sequences
Products of many small probabilities can underflow in software. Calculate in log space:
Then exponentiate the sum if a decimal probability is required.
The product uses the probability of the score actually selected. If a model’s most likely score is 1-1 at 12%, using a generic historical average such as “correct scores happen 10% of the time” discards match-level information.
Multiple correct-score probability calculator
Enter one success probability per match, as a percentage from 0 to 100. Use the probability of your selected exact score, or a Top-k cumulative probability when each pick is a score list. These are probabilities, not decimal betting odds. The guide to the correct score calculator explains how a match score matrix supplies these inputs.
Separate values with commas, spaces or line breaks. Use decimal points in a comma-separated list, for example 12.5, 10, 9.5. For decimal commas, include a semicolon or tab as a list separator, for example 12,5; 10; 9,5. A value of 0.12 means 0.12%, not 12%.
Equal per-match probabilities: the binomial distribution
If all n picks have the same success probability p and are independent, the number of correct scores X follows:
The probability of exactly k correct scores is:
The probability of at least k is:
Example: ten picks at 12% each
| Outcome | Probability |
|---|---|
| No correct scores | 27.85% |
| Exactly one correct | 37.98% |
| Exactly two correct | 23.30% |
| At least three correct | 10.87% |
| All ten correct | 0.0000000619% |
Getting one or two correct is common in this setup; getting all ten is extraordinarily rare.
Different probabilities: the Poisson-binomial distribution
Real match probabilities differ. One top score might be 15%, another 9%, and another 12%. If the Bernoulli success indicators remain independent, their sum follows a Poisson-binomial distribution.
Let Xi equal 1 when pick i is correct and 0 otherwise:
The probability-generating polynomial is:
The coefficient of zk is P(X = k).
Dynamic-programming calculation
Start with P(0 correct) = 1. For each new probability p, update the distribution:
This calculates the full distribution in O(n²) time and is stable for ordinary interactive use.
Poisson-binomial is not the ordinary Poisson distribution. The similar name reflects its history as a sum of unequal Bernoulli trials, not football goal counts.
Worked example: six score picks with different probabilities
Suppose the selected exact scores have probabilities:
All six correct
Full probability distribution
| Correct scores | Probability |
|---|---|
| 0 | 47.4345% |
| 1 | 37.7516% |
| 2 | 12.4264% |
| 3 | 2.1656% |
| 4 | 0.2108% |
| 5 | 0.01086% |
| 6 | 0.0002317% |
Derived cumulative results:
- at least one correct: 52.5655%;
- at least two correct: 14.8139%;
- at least three correct: 2.3875%;
- expected correct picks: 0.70.
An expectation of 0.70 does not mean “usually one correct.” In this example, zero correct is the single most likely count at 47.43%, while exactly one occurs 37.75% of the time.
Odds of predicting 3, 6, 10 or 18 correct scores in a row
For illustration, assume every selected score has probability 12% and matches are independent. This is a hypothetical comparison, not an observed accuracy rate for this tool.
| Scores required | Calculation | Probability | Approximate one in X |
|---|---|---|---|
| 3 | 0.12³ | 0.1728% | 1 in 579 |
| 6 | 0.12⁶ | 0.0002986% | 1 in 334,898 |
| 10 | 0.12¹⁰ | 0.00000006192% | 1 in 1.615 billion |
| 18 | 0.12¹⁸ | 2.662 × 10−15% | 1 in 3.756 × 1016 |
These calculations show why a long sequence of exact scores can look extraordinary even when each individual pick is the most likely score in its match.
One-in-X is not expected waiting time in every setting
If identical independent six-match attempts are repeated, the expected number of attempts until one complete success is 1/P. Real attempts are not identical: probabilities, match selection and dependence change.
Probability of at least one correct score
It is easier to calculate the complement: none of the picks is correct.
Six matches at 12% each
There is a better-than-even chance of at least one correct score, while the probability all six are correct is only 0.000299%. The two questions describe opposite ends of the count distribution.
Do not add probabilities to calculate “at least one.” Six 12% probabilities sum to 72%, but events can overlap across different matches. The correct result is 53.56% under independence.
Expected number of correct scores
For any set of Bernoulli indicators, expectation is additive:
This result does not require independence. If 100 picks each carry 12% probability, the expected number correct is 12.
Variance under independence
If picks are dependent, covariance terms must be added:
Expected hits do not validate probabilities
A model can produce the correct total number of hits while assigning probability badly across matches. For example, predicted 10% and 20% groups can be miscalibrated in opposite directions and cancel in aggregate.
Multiple Top-3 or Top-5 score predictions
Suppose a match’s three selected scores have probabilities 12%, 10% and 9%. Since exact scores are mutually exclusive:
For several matches, use each match’s cumulative Top-k probability in the same product or Poisson-binomial formulas. Enter one combined value per match: 31 in this example, not three separate inputs of 12, 10 and 9.
Example: six Top-3 lists at 35% each
This is about 1 in 544. It is the probability of list coverage across all six matches. Covering three scores in each of six matches creates 36 = 729 distinct exact-score combinations. Buying every combination is a different staking and payout problem; higher coverage alone does not establish a profitable bet.
Top-k boundary ties
Top-k should contain exactly k score cells. Rank with unrounded probabilities and use a predeclared stable tie-breaker when equal probabilities cross the cutoff—for example, fewer home goals first and then fewer away goals. Apply the same rule when calculating expected coverage and when checking the observed result.
If an exact tie includes the modal score or crosses a Top-k boundary, the combined probability of the full tied set can be reported separately. That supplementary modal-set figure should not silently change the definition of Top-k used for evaluation.
Top-k probability is a set probability. It should equal the sum of the k included score cells, not k multiplied by the Top-1 probability.
Combined fair odds and bookmaker correct-score accumulators
For a no-push combined event with probability P, model-implied fair decimal odds are:
In the equal-probability example, six independent picks at 12% give P(all six) = 0.000002985984:
Why multiplying bookmaker odds is not the same calculation
Decimal accumulator odds multiply the quoted prices:
Each correct-score price can contain margin. Multiplying several prices compounds whatever pricing bias and margin exist in the individual legs.
Expected return from model probabilities
Here EV is expected net profit per unit staked, assuming one all-or-nothing accumulator, no voids or pushes, no fees, and the full quoted payout. Payout caps or different settlement rules require a different payoff calculation. The probability estimate also depends on the individual inputs and the independence assumption.
A huge quoted payout does not imply positive value. Long correct-score combinations naturally require very large odds because their probability is extremely small.
If bookmaker correct-score prices are used, their margin must be handled explicitly before comparing them with model probabilities. The broader de-vig methods are explained in the No-Vig Calculator and margin-removal guide.
When is multiplying match probabilities invalid?
The product formula requires conditional independence:
Separate unrelated league matches
Conditional independence is often a practical approximation when matches involve different teams, times and competitions and the probabilities already include known information.
Sources of cross-match dependence
- the same team appears in multiple matches;
- a tournament result changes later incentives or qualification states;
- several matches share weather, venue or scheduling conditions;
- matches are played simultaneously with strategic interaction;
- probabilities share uncertain league or team-strength parameters.
Sources of probability misspecification
- lineup or injury news appears after probabilities are frozen;
- league-wide scoring conditions are estimated incorrectly;
- the same model bias affects all selected scorelines.
Stale or biased probabilities are not, by themselves, dependence between matches: they make an individual pi wrong. Dependence arises when the same information or a shared hidden factor affects several picks at once — the shared-model-error case below can create predictive dependence when that common uncertainty is treated as a random quantity.
Shared parameter uncertainty can create predictive dependence
Even when final scores are conditionally independent given their true probabilities, estimated probabilities can share uncertainty. If a model underestimates league scoring, several high-score picks can fail together more often than the product calculation suggests.
Positive and negative dependence
Positive covariance between picks increases the variance of the total number correct and often produces a more dispersed count distribution. However, its effect on a particular tail probability must be calculated from a specified joint model. Negative covariance reduces the variance relative to the independent case.
Independence is an assumption to document, not a fact inferred from separate fixtures. Without a joint model or repeated data, the exact dependence correction is not identifiable.
How to validate multiple-score probabilities
- Freeze each score distribution before kickoff. Store every selected score and its probability.
- Test individual calibration first. Scores assigned around 12% should occur around 12% over a sufficient sample.
- Compare expected and observed counts. Across a batch, compare Σpi with the number of correct picks.
- Check the full count distribution. Compare frequencies of 0, 1, 2 and more hits with Poisson-binomial predictions.
- Use exact-score log loss. Top-1 hit count discards the probability assigned to other possible results.
- Test dependence. Group by league, team overlap and time to identify excess joint failures or successes.
- Separate selection from evaluation. Do not choose batches retrospectively because they happened to contain many correct scores.
- Report uncertainty. Small batches produce highly variable counts even under calibrated probabilities.
Individual and grouped probability evaluation is explained in football probability calibration.
Common mistakes in multiple correct-score calculations
Multiplying a generic 10% by itself
Every match has its own score probability. Use the actual selected cell from each score matrix.
Adding probabilities for “all correct”
Addition applies to mutually exclusive alternatives. Separate match successes can occur jointly, so independent joint probability uses multiplication.
Multiplying for “at least one”
Calculate one minus the probability that every pick fails.
Using the ordinary binomial with unequal probabilities
The binomial requires one common p. Unequal probabilities require Poisson-binomial or an approximation whose error is acknowledged.
Confusing Poisson-binomial with Poisson goals
One models the count of successful predictions; the other models goals in one match.
Ignoring dependence
Repeated teams, shared model errors and tournament structure can invalidate the simple product.
Treating fair odds as a forecast guarantee
Reciprocal probability is a break-even price under the model, not evidence that the event will occur.
Evaluating only spectacular all-correct sequences
A probability model should be judged across every batch and every match, not only rare successes selected after the fact.
Practical conclusion
- Multiply individual probabilities to calculate all-correct probability under independence.
- Use binomial only when every per-match probability is equal.
- Use Poisson-binomial for different match probabilities.
- Calculate at-least-one probability through the complement of no successes.
- Use Σp as the expected number correct, not as a guarantee.
- Replace Top-1 probabilities with cumulative Top-k probabilities for score-list coverage.
- Validate individual calibration and the batch count distribution before trusting long-sequence odds.
Generate fixture-specific exact-score probabilities before combining them, rather than applying one generic success rate to every match.
Open the correct score calculatorReferences
- Wang, Y. H. (1993). On the Number of Successes in Independent Trials . Statistica Sinica, 3(2), 295–312.
- Hong, Y. (2013). On Computing the Distribution Function for the Poisson Binomial Distribution . Computational Statistics & Data Analysis, 59, 41–51.
- Chen, S. X. & Liu, J. S. (1997). Statistical Applications of the Poisson-Binomial and Conditional Bernoulli Distributions . Statistica Sinica, 7, 875–892.
- Gneiting, T. & Raftery, A. E. (2007). Strictly Proper Scoring Rules, Prediction, and Estimation . Journal of the American Statistical Association, 102(477), 359–378.
- ImpliedScore. How Often Is the Most Likely Football Score Actually Correct? (linked above).
- ImpliedScore. Football Probability Calibration: Brier Score, Log Loss and Reliability (linked above).