How every number on this site is produced
The prediction engine is deterministic maths. The language layer only describes values that were already computed — it never creates, rounds up or embellishes a probability.
The ten-step pipeline
- 1 · Data intake
Fixtures, line-ups, injuries, suspensions, travel distance, rest days, congestion, weather and market prices are collected per fixture and scored for completeness.
- 2 · Attack and defence strength
Each team's scoring and conceding rates are normalised against the strength of opposition faced, so a strong record in a weak league is not mistaken for elite quality.
- 3 · Form weighting
Recent matches are weighted more heavily than old ones, but form never overrides underlying quality on its own.
- 4 · Expected goals
xG for and against replaces raw goals where the sample supports it, correcting for teams that are finishing unusually well or badly.
- 5 · Availability
Absences are weighted by importance and position. A missing first-choice goalkeeper or a 25-goal striker moves the numbers far more than a rotation squad player.
- 6 · Travel, rest and congestion
Distance travelled, days of recovery, midweek European involvement and matches in the last fortnight all adjust expected output.
- 7 · Context
Home advantage per competition, motivation, managerial change and derby effects are applied as bounded adjustments — never as a free-form override.
- 8 · Scoreline model
Expected goals for each side feed a Poisson scoreline grid with a Dixon-Coles correction for low scores, producing home, draw, away, over/under and both-teams-to-score probabilities that sum correctly.
- 9 · Market comparison
Bookmaker odds are de-vigged to a true implied probability so model and market are compared on equal terms.
- 10 · Value and risk
Expected value, edge, data quality, league reliability and volatility produce a final rating, a confidence tier and a stake recommendation.
Never bet to a schedule
If no fixture clears the threshold, the honest output is an empty page. Every section is allowed to return nothing.
Fractional Kelly only
Default quarter Kelly, capped at 3% of bankroll per single and 1.5% per accumulator. Full Kelly is displayed for reference and never used.
No chasing
A losing day never increases the next stake. Stake size is a function of edge and bankroll, not of mood or recent results.
Correlation matters
Accumulator legs from the same match, or heavily correlated legs, are penalised or rejected. The true combined probability is shown, not the naive product.
Closing line value is the scoreboard
Beating the closing price consistently is stronger long-run evidence than a good week of results.
Losses stay published
No selection is removed after the fact and no historical probability is edited once the fixture kicks off.
What this platform will never claim
- • That a selection is guaranteed, fixed, or "banker" certain.
- • That an 82% pick wins 82% of the time without publishing the measured calibration.
- • That the £10 → £1,000,000 challenge is likely to succeed. It is published with its real, very low probability.
- • That past performance predicts future returns.
All figures are estimated model probabilities, not guarantees. No outcome is certain. Historical performance is measured separately from model confidence. 18+ only — gamble responsibly.