How Our Predictions Are Made
Every pick on CourtFrame comes from our own statistical model. It is retrained every night on every game we track, and judged on games it has never seen.
A prediction is a probability, not a certainty, and not betting advice. Past accuracy does not guarantee future results.
What Goes In
Team strength rating
An Elo-style rating updated after every game, with home court measured separately for each league. Three quarters of a team’s rating carries into the new season.
Recent form
Winning percentage, average margin, and points scored and allowed over the last 10 games.
Efficiency
Net rating over the last 10 games and the season, plus shooting (effective field-goal percentage), turnovers, offensive rebounding and free-throw rate — each measured against the opponents played.
Strength of schedule
A simple rating system that adjusts every margin for the quality of the opponent.
Rest
Days since each team’s last game, and whether either side is on the second night of a back-to-back.
Each game's inputs are built only from games played before it, so the model never learns from a result it is meant to predict. A logistic regression trained on that history turns the inputs into a win probability for each side and a projected margin.
How It Is Tested
The model is trained on older games and scored on newer ones it has not seen. We measure two things: how often the favourite wins, and whether the probabilities hold up — of all the games called at 70%, about 70% should go that way. It is also compared every night with the bookmakers' closing line.
Every pick we publish is graded against the final score, and misses stay on the record. The full history is on the track record page.
Reading a Probability
Clear favourite
The model expects this side to win about seven or more of every ten games like it. Upsets still happen.
Lean
One side has the better numbers, but the game is competitive.
Toss-up
Close to a coin flip on the evidence available.
When We Publish No Pick
- •Either team has played fewer than five games in that league — too little to rate it.
- •Preseason games: rotations and minutes say little about the real thing.
- •Leagues where predictions are switched off while their data is checked.
CourtFrame Power Index (CPI)
CPI is our team ranking for the power rankings — a separate measure, not an input of the prediction model. It combines eight dimensions into a 0–100 score, updated daily, from games of the current season. Teams that have not played yet show no score.
Known Limitations
- •The published model does not yet account for which individual players are available.
- •News that breaks shortly before tip-off is not reflected in the probability.
- •Early in a season the ratings lean on last season, so they can lag a team that has changed.
- •Leagues with few games per season give the model less to learn from, so their probabilities are less certain.