Methodology

How the predictions work

Every prediction on this site comes from a statistical model, published before tip-off and graded against the final score. Nothing is edited after the fact — the accuracy page is the full ledger, losses included.

What the model looks at

For each game we compute, from data available strictly before tip-off: each team's last-ten-games form (win rate and average margin), season-to-date offensive and defensive efficiency, rest days and back-to-back situations, home court, and the recent head-to-head record.

The model itself

A blend of two models — a gradient-boosted decision tree model and a regularized logistic regression — trained on roughly ten seasons of NBA results and validated on held-out seasons they never saw during training. The blend's raw outputs are calibrated so that a stated 65% means the pick should win about 65% of the time — you can check that promise yourself on the accuracy page, where results are broken down by stated confidence.

What it doesn't know

The current version does not use injury reports, lineup news, or betting lines. When a star sits out on short notice, the model won't know — some losses come from exactly that. We prefer publishing an honest, simple model over quietly patching predictions after news breaks.

Grading rules

A pick is correct if the predicted team wins the game, full stop. Every published prediction is stored, graded, and counted — there is no "push," no vetoing bad picks, and no retroactive edits.

Nothing here is betting advice. It's a public experiment in whether a disciplined, transparent model can call NBA games well.