How SportMetrics Predictions Work

1. The data we start from

For each team we pull together the measurable signals that actually move a result. Nothing here is subjective — every input is a number taken from real matches:

• League standing — rank, points, and games played in the current table.
• Attack & defence — goals (or points) scored and conceded per game.
• Recent form — how the side has performed over its most recent matches.
• Home / away splits — many teams are a different side at home than away.
• Head-to-head — the recent history between these two specific teams.
• Schedule strength — the calibre of opponents recently faced, and fixture congestion.

2. The football model — Poisson

Football is a low-scoring sport where goals arrive roughly independently, which is exactly what a Poisson distribution describes. It is the industry-standard method for modelling football scorelines.

From each side’s attacking and defensive strength we estimate an expected number of goals (λ) for that match, applying a home-field adjustment — the home side’s expected goals are nudged up, the away side’s down. The Poisson distribution then converts each λ into the probability of every possible scoreline: 0-0, 1-0, 2-1, and so on.

Summing those scoreline probabilities gives the win, draw and loss probabilities, and the single most likely scoreline becomes the projected score. Because it is pure maths over the inputs, the same data always produces the same prediction — it is fully reproducible.

3. The basketball model — normal margin

Basketball is high-scoring, and the margin between two teams is roughly normally distributed rather than Poisson. So for basketball we model the points margin with a normal distribution built from each team’s scoring and defensive averages, producing a win probability and a projected margin and score.

4. When the data is thin

Not every fixture comes with a full dataset, and we would rather adjust honestly than pretend otherwise:

• Newly promoted teams have no top-flight table, so their figures come from a lower division. Lower-division output rarely carries over one-to-one, so we discount it and flag the match as lower-confidence.
• Early in a season, before enough games have been played, we fall back to last season’s data and label the analysis as based on the previous season.
• Where two sides have genuinely no comparable data, we say so rather than manufacturing a number.

5. Turning numbers into words

Only after the statistical model has produced the probabilities and projected score does a language model write the preview. Its job is purely to phrase the numbers the model already computed — the two sides’ form, the key statistical factors, and what the figures collectively point to.

The model does not calculate probabilities and it does not invent statistics. Every article is checked against the underlying data before it is published. This is the single most important design choice on the site: it keeps predictions reproducible and rules out the “the model made up a stat” problem.

6. What predictions do not include

A prediction is a probability, not a certainty, and it is deliberately scoped:

• It does not account for injuries, suspensions or confirmed line-ups.
• It cannot know about last-minute tactical or motivational factors.
• It is for informational and analytical purposes only — it is not betting advice.

7. A record you can check

Transparency only counts if it is verifiable. Every prediction is stored and, once the match finishes, scored against the actual result. Accuracy accumulates in the open — historical results are never back-fitted to flatter the numbers. When enough predictions have settled to be statistically meaningful, the running accuracy is published for anyone to inspect.