How the robot works, end to end

Candidate fixtures come from a market-activity feed covering football competitions worldwide, filtered to matches with genuine traded volume. Fixtures without an identifiable competition, kick-off time or team pairing are discarded before any modelling happens.

Every observation of a price is stored as a new timestamped snapshot; previous snapshots are never overwritten. That is what allows a drop to be measured rather than guessed, and it is what makes closing line value calculable after the fact.

Refresh cadence tightens as kick-off approaches. A selection whose most recent price snapshot is stale is not eligible — the robot would rather skip a match than act on an out-of-date number.

Each candidate carries a data-quality assessment covering how recent the team data is, how many matches of the current season are available, and whether the market looks internally consistent. Low-quality candidates are excluded rather than downweighted.

Three components contribute. A Poisson goal model converts venue-split attacking and defensive strength into a scoreline distribution. An Elo rating tracks relative team strength over time, computed strictly from matches that had already finished at the point of prediction, so no future information can leak backwards. A recent-form component captures short-term shifts the slower models miss.

The three are blended into an ensemble probability. No single model is trusted on its own, because each has a known blind spot.

Bookmaker prices include a margin, so raw implied probabilities sum to more than 100%. Normalising them removes the margin and produces a fair market probability, which is what the model is compared against.

implied = 1 / odds · fair = implied / Σ implied · edge = p_model − fair

EV per €1 = p_model × (odds − 1) − (1 − p_model)

A candidate must clear a market-volume threshold, show a meaningful and well-supported price movement, carry a fresh price snapshot, pass the data-quality checks and produce a positive edge against the no-vig fair probability. Edges that look implausibly large are flagged for sanity review rather than accepted at face value.

Most matches on any given day qualify for nothing. That is the intended behaviour.

Qualified selections are staked in parallel by four independent strategies on identical bets. Each strategy keeps its own bankroll and runs strictly in sequence: select, lock, wait for the match, settle, update history and statistics, then and only then select the next bet.

Final scores are read from public scoreboards through several independent providers, so settlement does not depend on any single paid source. A bet is graded only when a match is confirmed final. Grading writes the result, the profit or loss and the resulting bankroll into the permanent ledger, which is the single source of truth behind every public statistic on this site.

The public pages read that ledger. They never settle bets, never modify history and never trigger provider calls.

It does not place real bets, connect to a bookmaker account or handle stakes. It does not predict outcomes with certainty — it estimates probabilities, and probability estimates are wrong a predictable share of the time by design. It does not promise profit, and no strategy here is described as profitable while the settled sample remains small.