How it works
We never copy a third-party prediction. Here is exactly how a pick is produced.
1. Data collection
Fixtures, results, form, head-to-head, standings, team statistics, injuries and line-ups are pulled from API-Football through a throttled, cached pipeline and stored locally.
2. Feature engineering
A feature builder converts raw data into normalised 0–1 signals: recent form, home/away performance, attack and defence strength, H2H, league position, availability and context (rest days).
3. Expected goals
Attack and defence ratios relative to the league average produce an expected-goals figure for each side, with Bayesian shrinkage so small samples cannot distort the model.
4. Poisson goal model
Two independent Poisson distributions generate a full score matrix. From it we derive 1X2, double chance, over/under goals, BTTS and the most likely correct score.
5. Confidence calibration
Confidence is not the same as probability. It is the probability pulled toward 50 when key data is missing, so a thin-data pick can never look like a certainty.
6. Selection & settlement
The day’s free top-5 are chosen with market and team-concentration controls. When matches finish we settle every market and record the outcome for back-testing.
What we will never do
- • Claim a prediction is a “sure banker”, “fixed” or “guaranteed”.
- • Show a 100% probability or invent missing statistics.
- • Publish results we cannot substantiate from settled, timestamped predictions.