How to Read a Score Probability Map (Top-1, Top-3, Top-5 Coverage)

A football score probability map shows the estimated chance of each final score. Find the home and away goal labels, read the percentage at their intersection, then compare that cell with the rest of the distribution. The highest-probability score can still leave most of the probability on other results.

This guide explains the grid, the probability held by the top one, three or five scorelines, and how to compare those estimates with results across a series of matches.

Build a score probability map →

Each cell represents one scoreline, written home goals first and away goals second. Check the axis labels before reading the grid. In the example below, home goals run down the rows and away goals across the columns, so row 2 and column 1 give the chance of a 2–1 home win.

Brighter cells in this example have higher probabilities. The printed percentages give the actual values; colour alone may not be comparable across different maps.

The main ImpliedScore calculator fits goal rates to market targets after removing bookmaker margin. A Poisson or Dixon-Coles model then distributes probability across scorelines. The correct score calculator guide explains that process.

This illustrative map uses independent Poisson goals with expected goals of 1.60 for the home team and 1.10 for the away team. It is a model example, not a forecast for a named fixture. Only scores from 0 to 3 goals per team are displayed.

Rows: home goals. Columns: away goals.

For example, 2–1 is 9.46%, while 1–2 is 6.51%. Reversing the score changes the event. The largest cell is 1–1 at 11.83%, leaving 88.17% on all other scorelines.

The displayed cells contain 89.75% of the probability. The remaining 10.25% belongs to scores where at least one team reaches four goals. Those outcomes are outside this small illustration.

The example uses P(h, a) = e^−2.7 × 1.6^h/h! × 1.1^a/a!. Percentages are rounded after calculation.

The complete theoretical distribution sums to 100%. A displayed heatmap usually shows only a limited range of scores, so it can omit some probability. Rounding each cell can create a further small difference.

ImpliedScore uses a larger computational matrix than the visible heatmap and reports probability outside the visible grid. Its finite computational grid also leaves a small tail. Do not rescale the visible cells to 100% and then read them as the original match probabilities: that changes them into probabilities conditional on the result being inside the displayed range. The methodology documents these limits.

Rank all scorelines by their pre-match probability. Top-1 contains the largest cell, top-3 contains the three largest, and top-5 the five largest. Add the probabilities of the selected cells to get the model's chance that the result will fall in that set.

Here, top-1 is 11.83%, top-3 is 32.04%, and top-5 is 48.04%. Top-3 includes top-1; top-5 includes both. They are nested sets, so adding these three coverage figures together would double-count scorelines.

For this example, the model still assigns 51.96% to results outside the five highest-ranked scores. There is no universal 10–16% limit for the largest cell; the percentages depend on the model and the goal rates.

Rank scores using unrounded probabilities and keep Top-k to exactly k scorelines. Resolve equal probabilities with a fixed rule chosen before results are known, for example fewer home goals first, then fewer away goals. Use the same selected cells for expected and observed coverage. The example above has no tie at its Top-1, Top-3 or Top-5 cutoffs.

For a tie example, independent Poisson with both goal rates equal to 1 gives 0–0, 0–1, 1–0 and 1–1 the same probability of 13.53% each. A three-cell Top-3 covers 40.60%; all four tied modes together cover 54.13%. They are different prediction sets.

Covering several possible scores in one match differs from predicting scores across several matches. See the probability of multiple correct football scores for combinations across matches and the independence assumption.

A single map provides a probability for each top-k set. Observed coverage is measured across settled matches: how often did the final score land in the set selected for that match?

• Expected top-k coverage: the average of each match's pre-match top-k probability.
• Observed top-k coverage: the number of matches whose final score was in their pre-match top-k set, divided by the number of evaluated matches.

Use the same settled matches for both calculations. Freeze probabilities and ranks before kickoff, exclude pending results, and match the scoring period to the input market. Regulation-time odds should be evaluated against the regulation-time score, including stoppage time, rather than a score after extra time or a penalty shootout.

The published World Cup 2026 tracker reports the following historical summary for 104 settled matches. Its Top-1, Top-3 and Top-5 hit counts are 18, 41 and 56. These aggregate figures describe that sample; they are separate from the illustrative map above and are not a measured accuracy rate for every future match.

Top-1 observed coverage is 3.6 percentage points above its mean expectation. Top-5 differs by 0.3 percentage points. Neither comparison proves that the model is well calibrated or miscalibrated: with 104 matches, a single additional hit changes observed coverage by about 0.96 percentage points.

The tracker's 1X2 and over/under diagnostics use probabilities assigned to realised outcomes and corresponding model benchmarks. They are different metrics from the top-k hit rates in this table.

Coverage is a useful diagnostic. If top-5 probabilities are systematically too high, results will tend to land outside those sets more often than forecast. If they are too low, observed coverage will tend to exceed the forecasts. Small samples can show either pattern by chance.

Matching the overall average is not enough. A model could overstate coverage for one group of matches and understate it for another, with the errors cancelling in the total. Compare forecast probability ranges, account for sample uncertainty and examine further matches. The probability calibration guide explains reliability checks, Brier score and log loss.

A higher top-5 probability means the model puts more mass into five scores. It is not, by itself, evidence of greater accuracy or better betting value.

1. Check the labels and inputs. Confirm home/away order, market prices and the settlement period.
2. Read the largest cells together. Compare their probabilities and the mass outside the top five. A sharper peak means a more concentrated model distribution, not guaranteed predictability.
3. Account for scores outside the visible grid. Use the calculator's derived markets instead of summing only the cells that fit on screen.
4. Check Model Fit. Derived 1X2 probabilities can differ from the margin-removed market inputs because fitting is approximate. Residuals describe that numerical mismatch, not future forecast accuracy.

Home-win probability adds all cells where home goals exceed away goals; draws are the diagonal; away wins use the opposite cells. Totals and BTTS select different groups. See how over/under and BTTS come from the score matrix for those calculations.

Understand most-likely-score accuracy →