No magic, no promises. A weekly model that turns match data into calibrated probabilities, and shows its own track record.
Every week the model reads five seasons across five leagues: who played whom, where, and how it ended.
Each team carries an Elo rating that updates after every match. A win against a strong side moves it more than a win against a weak one.
Only information available before kickoff is used. Nothing from the match itself leaks into its own prediction.
From the ratings and recent form, the model estimates how many goals each side is likely to score.
A double-Poisson model turns those two numbers into the full grid of possible scores, and from there into 1X2, over/under, and both-teams-to-score probabilities.
The output is a probability, never a certainty. A 60% home win still loses four times out of ten.
Once a match is played, the model compares what it predicted to what happened, and records the result.
It is measured against the bookmaker on the exact same matches, using log-loss, the honest metric for probabilities.
You can see the running comparison, week by week, on the performance page. Good weeks and bad weeks, nothing hidden.
The full method, code, and deployment are taught step by step in the course.
See the courseAI-generated predictions for informational and educational purposes. Not betting advice.