How PredictIQ Pro's Football Predictions Work

PredictIQ Pro produces win, draw and loss probabilities for football matches across 80+ leagues. Behind every number is not a single model but a stack of five components, each doing a different job. This article explains what they are and why the design works.

The five components

• A Poisson expected-goals model. Estimates how many goals each team is likely to score, then converts those two rates into probabilities for every possible scoreline. See how Poisson scorelines work.
• CatBoost. A gradient-boosted classifier trained on inputs such as recent form, rest days and head-to-head history.
• LightGBM. A second gradient-boosted classifier. It belongs to the same family as CatBoost but builds its trees differently, so it tends to make different mistakes.
• Devigged market odds. Bookmaker odds with the bookmaker's margin removed, treated as the market's opinion of the match. See what devigging means.
• A meta-learner. A final model that learns how much weight to give each of the other four.

Why stack instead of picking one model

Every model has blind spots. A goals model understands scoring rates but not schedule fatigue; a classifier can pick up patterns in form but has no view of what the betting market knows. When you combine models whose errors are only partly correlated, the combination is usually more reliable than any single member. The meta-learner's job is to work out, from past results, when each component deserves more trust.

One scoreline model behind every market

The 1X2, over/under and both-teams-to-score markets are all derived from the same scoreline model. That keeps them consistent with one another: a match can't be shown as a likely low-scoring game in one market and a goal-fest in another.

A retrained model has to earn its place

Models are retrained on a schedule, but a retrained model only replaces the live one if it demonstrably beats it. A scheduled retrain therefore can't silently make the predictions worse.

Checking the work

Claims about a model are only useful if you can check them. PredictIQ Pro publishes its measured performance on the accuracy page, including the live model version, log loss, Brier score and calibration. To understand what those measures mean, read prediction calibration explained. To see the model's current output, browse the live predictions.

PredictIQ Pro is an informational and statistical service. It does not accept wagers or process bets, and its probabilities are not guarantees of any outcome.