Methodology · Week of July 13 · 6 min read

How AI Predicts Sports Outcomes Without Pretending to Know the Score

The model never tries to predict a final score. It estimates a probability, compares it to a price, and acts only when the two disagree by enough.

Prediction is a pricing problem

The common misunderstanding about AI sports prediction is that the goal is forecasting outcomes. It is not. The goal is estimating probabilities more accurately than the market has priced them, then betting only the disagreement.

That reframing changes everything about how the system is built. Accuracy against reality matters less than calibration against price.

Seven stages, in order

Ingest raw logs, rotations, injuries, travel, weather and multi-book line movement. Standardise every feature to a z-score against its own league-season distribution. Weight the eight feature families with fitted weights summing to one. Calibrate the composite signal through logistic regression into a probability.

Then price: de-vig the market odds to get the market's implied probability. Size: apply quarter-Kelly staking, clamped between half a unit and three units. Publish: post the probability, the edge and the stake before the event starts, and never revise it afterwards.

The part that is hardest to fake

Anyone can publish picks. What is difficult to fake over time is publishing the probability and stake in advance, then reporting settled results against that published record without editing.

That is the operating standard here, and it is the reason the audit page exists as a separate, permanently linked artefact rather than a screenshot in a marketing section.

Go deeper

Get the daily card — $12.28 upfront →

More from AI Picks & Insights

Browse the full AI Picks & Insights archive →
Published weekly

The weekly Insights digest

One email a week: the new report, what the model got right and wrong, and the numbers behind it. No touts, no upsells between issues.