How the predictions are made, and what they can and cannot do
Short version: the market is the best public forecaster of a football match. We turn bookmaker prices into fair probabilities, add a modest statistical model, publish the most likely outcome before kick-off with the odds we saw, and then grade it in public.
1. Where the probabilities come from
For every match we store the pre-match prices of 7 bookmakers (bet365, Pinnacle, Betfair, Marathonbet, BetVictor, William Hill, 1xBet) for these markets: match result, double chance, over/under, both teams to score, result & both teams to score, goals 2.5 & both teams to score, result & goals 2.5. Combination markets exist only where a bookmaker actually prices them; we never multiply probabilities to invent a price, because the outcomes are correlated and the record would not be verifiable.
Each bookmaker's prices include a margin. We remove it per bookmaker (each outcome's implied probability divided by the sum of the implied probabilities) and average the result across bookmakers. That gives the market consensus: a fair probability for each outcome. We also keep the best price, the average price and the price at a sharp reference bookmaker.
2. The model
Alongside the market we run an independent Poisson model with the Dixon-Coles low-score correction. Each team gets an attack and a defence strength from this season's results, weighted towards recent matches (half-life 90 days), shrunk towards the league average with a prior of 5 games, and blended with expected goals where xG is available (weight 50%). Expected goals for a match come from home-ground scoring rates times the two strengths. The scoreline grid on match pages is this model.
The published probability is 75% market consensus and 25% model. The model needs at least 5 games for both teams in the current season; otherwise the pick follows the market alone and the match page says so.
Why so little weight on the model? Because we tested it. On 46,800 out-of-sample matches a Poisson model with xG blending and Dixon-Coles correction was worse than the closing odds on both match result and total goals. A model is useful for explanation and scorelines; it is not an edge.
3. What a pick is
The pick in each market is simply the most likely outcome. Confidence labels are bands on the probability: for the match result, Strong is 60% and above, Lean is 45% and above, anything lower is a Toss-up. For the two-way markets the thresholds are 62% and 55%.
A high probability means short odds. Short odds do not mean profit. A 75% pick at 1.25 loses money if it wins less than 80% of the time.
3b. The best pick and the Strong segment
Each match also gets one best pick: the most probable of its match-result, total-goals and both-teams-to-score picks, provided it reaches 60%. It is chosen once, at publication, and a market priced later never replaces it. The headline numbers on the home page are the best picks; the full record shows all picks and the Strong 1X2 segment (match-result picks at 60% or more) side by side.
Why show three selections? Because a hit rate on its own is a choice of odds, not a measure of skill. Restricting to favourites raises the percentage and lowers the price in the same move. Each selection is therefore printed next to its break-even hit rate and its units, and you can switch between them on the record page.
4. Publication and immutability
Picks are generated for matches starting within the next 7 days once at least 2 bookmakers have priced them. At publication we record the timestamp, the pick, the probability and the odds available at that moment at a named bookmaker. A published pick is never edited or removed. If odds change later, the record still uses the price we published. When the match kicks off we also store the closing price, so you can see whether our timing beat or trailed the market (closing line value).
5. Grading
- Picks are graded on the 90-minute score, including stoppage time, excluding extra time and penalties.
- A pick is won or lost, nothing in between. Postponed, cancelled and abandoned matches are void and excluded from every percentage.
- Hit rate is shown next to the break-even hit rate (one divided by the average odds at publication) so a percentage is never read without its context. The CSV export carries the odds of every pick for anyone who wants to run their own numbers.
6. Live record versus backtest
The live record contains only picks published before kick-off. The backtest applies the identical method to past matches using the last odds available before kick-off, with team ratings computed only from matches played before each one. It exists so the method can be judged on a large sample from day one, and it is labelled as a backtest everywhere it appears. The two are never mixed in any number on this site.
Right now the backtest stands at 37310 picks across all markets, 57.6% hit rate against a break-even of 52.7%. The live record has 0 settled picks.
7. Free and Pro
Without Pro you get the 5 strongest best picks of the day (from the highest-tier competitions), every market probability, the complete track record and every pick on a finished match. Pro adds every pick in every market, the scoreline model, the odds comparison, the odds-band and competition breakdowns and CSV export. The switch in the header previews both; there are no accounts yet.
8. Limits, honestly
- We do not beat the closing line and we do not claim to. The product is information presented honestly, not an edge.
- Coverage follows our data provider. Some competitions have no prices or no xG, and then there is no pick or no model.
- Odds at publication are a snapshot. You may get better or worse prices.
- Betting carries risk. Follow the picks only with money you can afford to lose, and only if you are of legal age where you live.