Glossary

Est. win chance

The model's estimated chance that a team wins a game, shown for both teams. Our pick is the team above 50%.

Pick of record

The last prediction stored before tip-off. It has a time stamp, is never edited, and is the one that gets graded.

Accuracy

The share of picks of record that picked the winner. Useful, but it says nothing about whether the chances themselves were right.

Brier score

The squared gap between the stated chance and the result (1 for a win, 0 for a loss), averaged over games. Lower is better; always saying 50% scores 0.25.

Log loss

Another score for probability predictions that punishes confident misses even harder than the Brier score. Lower is better.

Calibration

Whether chances mean what they say: over many games, teams given an est. 70% should win about 70% of the time.

Walk-forward test

A backtest where each day's predictions use only games already played, the way the model works live. It's the honest way to measure a model on past seasons.

Home-team baseline

How often the home team wins. Any NBA model has to beat it to be worth anything.

Team rating

A number for how strong a team is, from its results adjusted for the strength of the opponents it played.

Lineup rating

Our estimate of a player's impact: how much the score moves per 100 possessions while that player is on the floor, built from play-by-play since 2021. It's an estimate, not a talent ranking.

Per 100 possessions

A way to compare teams and players at the same pace: points scored or allowed for every 100 trips down the floor.

Rest

Days off since a team's last game. Back-to-backs (no day off) are one of the factors in a pick.

Season projection

Est. wins, playoff and title chances from simulating the rest of the schedule 5,000 times.

Play-in tournament

The NBA's mini-tournament for the teams that finish 7th to 10th in each conference; it decides the last two playoff seeds in each conference.

More: NBA FAQ · How accurate? · niftiIQ glossary