One score is not the whole prediction

Start with a complete outcome distribution

Consider an illustrative match, not a live SportXG prediction. The model assigns 49% to a home win, 25% to a draw and 26% to an away win. These three mutually exclusive outcomes account for 100% of the result distribution. A home win is the largest individual category, but the other two categories together still have 51%. Calling the home team the favourite does not make its win certain, or even more likely than all alternatives combined.

A score is one part of an outcome

Suppose the exact-score distribution includes 1–1 at 11%, 2–1 at 9% and 1–0 at 9%. The draw score is larger than either home-win score on its own. But the home-win category also contains 2–0, 3–1, 3–2 and every other score where the home team finishes ahead. Adding all those home-win scores can produce the 49% outcome probability.

The three displayed scores sum to 29%, not 100%, because the rest of the score distribution is not displayed. Renormalising only those three to 100% would overstate their actual chances. The remaining 71% must stay visible as uncertainty, rather than disappearing from the explanation.

Judge the original call, not a rewritten forecast

Before kickoff, the useful questions are what the model expects and which evidence supports that view. During play, the score, incidents and available statistics become the main story. After full time, a fair review compares the recorded pre-match call with the actual result. A correct outcome and an incorrect exact score are separate findings: neither should be hidden to make the model appear better.

Missing information is also meaningful. A lineup that has not been published is not an empty lineup, and missing possession data is not zero possession. A forecast page should distinguish unavailable coverage, a failed refresh and a genuine zero value. Always check the publication time, fixture status and data timestamp before drawing a conclusion.

A useful probability needs more than a winning streak

For a calibrated model, predictions around 60% should succeed roughly six times in ten over a sufficiently large, comparable sample. One win does not validate a 60% estimate, and one loss does not invalidate it. Small league samples and changing model versions make short-term comparisons particularly fragile.

Accuracy measures the fraction of outcome calls that were correct. Log loss also evaluates the probability assigned to the result that actually occurred, penalising confident mistakes. The multiclass Brier score sums the squared differences between the three predicted probabilities and the three observed outcome indicators; SportXG’s monitoring uses the unscaled 0–2 convention. Lower values of these two scoring metrics are better, but comparisons only make sense on the same evaluation cohort.

This is why a transparent archive needs sample counts and model versions as well as a hit rate. It is also why a high hit rate alone does not establish a positive financial return: returns depend on available prices and use a different eligible sample. Treat probabilities as descriptions of uncertainty, not promises or betting instructions.