How Our Predictions Work
Last updated: August 2026
What this page explains
Derbily calculates its own match predictions instead of showing bookmaker odds. This page explains, in plain language, how the model turns recent results into probabilities, what its confidence score actually measures, where the model is weakest, and how well it has performed so far. Nothing here is betting advice — see our Terms of Use.
Step 1 — expected goals from recent matches
For each team, the model looks at its recent official matches and works out how many goals it tends to score and concede compared to the league average. Friendly matches are excluded, because a club's warm-up results can be wildly out of step with its competitive level and would otherwise distort the estimate. More recent matches count more than older ones: a result from a week ago carries more weight than one from three months ago, and that weighting fades on a roughly four-month scale rather than dropping off sharply.
Step 2 — smoothing out small-sample noise
Five or six recent matches are not a lot of data, and reading too much into a hot or cold streak is a common way statistical models go wrong. To correct for this, each team's short-term numbers are pulled partway back toward its known level from the previous season when that baseline is available, and toward the league average when it is not. A team coming off one unusually high-scoring match is not assumed to keep playing at that pace.
Step 3 — one shared model, not separate guesses per market
Once both teams' expected goals are set, the model builds a single grid of scoreline probabilities (0–0, 1–0, 1–1, and so on, with a well-known statistical adjustment for how often low-scoring draws occur). Every market shown on Derbily — match result, over/under goals, both teams to score, half-time result, handicaps — is read off that same grid. This matters because it keeps the numbers internally consistent: the model cannot say a home win is likely while also implying a scoreline that contradicts it, because they come from the same source.
What the confidence score means — and does not mean
Each prediction carries a confidence label (low / medium / high) with a short list of reasons. This score reflects how solid the model's inputs are for that specific match — enough recent matches, recent-enough data, a known baseline for both teams — not how likely the prediction is to be correct. A high-confidence prediction can still lose, and a low-confidence one can still land; confidence describes the quality of what went into the calculation, not the outcome.
What the model does not account for
The model does not weigh who each team's recent opponents actually were — five wins against relegation-threatened sides and five wins against European contenders count the same in the raw numbers, though the previous-season baseline partly offsets this. Missing players due to injury or suspension are not factored into the goal estimate itself; a long injury list only lowers the confidence score, it does not change the predicted probabilities. When a team's first-half scoring split cannot be measured from too few goals, the model falls back to a league-wide average for that split. These are stated limitations, not hidden ones — where they apply to a specific match, the reasons list says so.
Our measured hit rate
We re-score every graded prediction every week and publish the result. In the week of 24–30 August 2026 (week 35), the model's main predictions hit 55.5% (283 of 510 graded picks across 85 matches). Past weeks are never deleted or hidden, whether the number went up or down. A hit rate describes past performance across many matches — it is not a probability for any single upcoming match, and it is not a guarantee of future results.
Not betting advice
These predictions are a statistical exercise, not a recommendation. We deliberately do not use bookmaker odds anywhere in the calculation, and none of this content is intended to encourage betting. Please see our Terms of Use for the full disclaimer.