How the model works

PredictXI isn't a tipster's gut feel. Every prediction comes from a statistical pipeline you can inspect.

Two models, two lenses

A Dixon-Coles goals model estimates each team's attacking and defensive strength from years of results, weighted toward recent form, and turns that into a full distribution of possible scorelines. Alongside it, an Elo rating model tracks team strength as a single moving number. Two independent views catch each other's mistakes.

Ensemble & blend

We combine both models into one set of win/draw/loss probabilities, then blend with bookmaker odds where available. Where our probability meaningfully beats the price the market implies, we flag a value signal.

Every market from one engine

From that scoreline distribution we read off every goals market consistently — Over/Under, Both Teams To Score, team goals, correct score, first-half lines — so they can never disagree with each other. Corners and cards have their own models built from each team's corner- and card-rates (cards use a negative-binomial because bookings are "spikier" than goals). We only ship a market once we can grade it against real results; see markets explained.

Calibration

Raw model probabilities are re-tuned against our own out-of-sample record so the numbers mean what they say — when we show 70%, outcomes at that confidence really land about 70% of the time. See what a probability really means.

Sealed & graded in the open

Each day's predictions are sealed into a tamper-evident, hash-chained ledger before kickoff, then graded against the real result — win or lose. Nothing is quietly edited after the fact, and the full track record is public.