Methodology

Each match gets one probability for home win, draw and away win. It comes from the PredictIQ Stack, an ensemble of four inputs: a CatBoost classifier, a LightGBM classifier, Poisson goal models, and the probabilities implied by bookmaker odds.

The Poisson goal models turn expected goals into a grid of scorelines. A Dixon-Coles adjustment corrects that grid for a known quirk of football: low scores such as 0-0 and 1-1 happen a little more often than independent goals would suggest.

A meta-model blends the four inputs into one probability, and a calibration step then adjusts it so that, over many matches, a stated 60% happens about 60% of the time. We check that calibration on the accuracy page. For goal-based markets, Monte Carlo simulation is used to estimate how uncertain the result is.

A newly trained model replaces the production one only if it measurably beats it on held-out matches, so a scheduled retrain cannot quietly make predictions worse.

The first forecast we make for a match, before it starts, is saved and never overwritten. Track records on this site are scored against those saved forecasts, not against predictions updated after the teams are known to be winning or losing. That is what stops a track record from being flattered by hindsight.

We publish three measures on the accuracy page:

• Accuracy: how often the most likely outcome happened.
• Log loss: penalises confident wrong calls harder than cautious ones. Lower is better.
• Brier score: the average squared gap between the probability given and what happened. Lower is better.

The Top Picks list shows only matches where the model is most confident. Its percentage is the share of settled picks that were right, shown together with the number of picks, so you can see how large the sample is.

• Small samples are noisy. A record built on a few dozen picks can move several points in a week without the model changing.
• A probability is not a promise. A 70% pick still loses three times in ten.
• Football has surprises the data cannot see: late injuries, weather, motivation, a red card in minute 8.
• Past accuracy does not guarantee future accuracy.

Fixtures, scores, line-ups, standings and odds come from a licensed sports data partner and from our own database of match results. Team and league statistics shown on this site are computed from those results. Where a value is not reported, we show it as missing rather than as zero.

If you find a wrong score, a mislabelled match or an error in this page, email support@predictiqpro.com. Factual errors are corrected and the correction is not hidden.

PredictIQ Pro is an information and statistics service. It does not accept wagers, process bets or guarantee outcomes. You must be of legal age to bet where you live. If betting is causing you harm, please contact a local support service.