How the model works

The pipeline, step by step

1. Prices in. Twice a day we pull live odds for every fixture across 18 leagues from dozens of bookmakers, and store a snapshot — that price history later powers the dropping-odds view and our closing-line measurements.

2. Fair odds. For every market we average the implied probabilities across bookmakers and remove the margin, producing the market's own "fair" price for each outcome — the baseline any bet must beat.

3. Our own signal. A Poisson goal model — attack and defence strengths fitted on two seasons of results per league, recency-weighted — produces independent probabilities for match outcomes and totals. Where the dataset is too thin (a promoted side, a new league), the model honestly abstains.

4. Blend and select. Model and market probabilities are blended, and a fixture only becomes an official pick when the best available price clears a real value threshold at sane odds. Roughly a quarter of fixtures qualify; the rest carry a clearly-labelled "model lean" for information. Official picks — and only those — are archived with price and timestamp and settled against final scores.

5. Receipts. Every number on the performance page links back to individual archived bets. Losing runs stay on record; nothing is edited after publication.

Version history

v1 (launch). Market-consensus only: fair odds plus best-price comparison, a bet on every fixture. Its expected return was, by construction, roughly zero minus the margin — and its record shows exactly that, including a bad run of longshot 1X2 picks. We keep it on the books because that's the point of the site.

v2 (29 Jul 2026, current). Added the Poisson goal model, probability floors, an odds cap for match-result bets and selectivity (~25% of fixtures). Judged on its own slice of the record and, faster, on closing-line value.

How we grade ourselves

ROI on small samples is noise, so the primary early metric is CLV — closing-line value: whether the price we took beats the last price we tracked before kick-off. Beating the close consistently is the standard evidence of a real edge; failing to is disqualifying no matter how a hot streak looks. The running numbers live in the public CLV experiment.

Honest limits

Our "closing" price is the last snapshot we took, not the true market close. The goal model is deliberately simple — no injuries, lineups or news, only results and prices — so it abstains rather than guesses. Snapshots are periodic, so opening prices are "first tracked", not the bookmaker's true opener. And no model removes variance: a positive-EV bet still loses often. If v2's CLV settles negative over a few hundred bets, the model has no edge and this site will say so on its own pages.