Research notes
Notebook
A working log of changes to the model and the questions behind them, including the ideas that were tested and dropped.
-
How often does a Questionable player actually play?
About half the time. Players listed Questionable on the 5:30 PM ET injury report played in 49% of cases, yet the model currently counts them as fully available.
-
NBA Cup games: are they higher scoring?
Point differential is a Cup tiebreaker, so teams might keep starters in. Across 180 group-stage games, scoring and late-game behavior looked like regular games.
-
Retrained on five seasons of history
Elo and training now start in October 2021 instead of 2023. On held-out seasons, the longer history matched or beat the shorter one.
-
A walk-forward backtest of every season since 2022-23
After filling data gaps, every game from 2022-23 on was predicted by a model that had never seen that season, using that day's injury report and no betting lines.
-
Does home court matter less when the visitor is clearly better?
Across the full backtest, mostly no. Road favorites ended close to their forecast margin at every size of gap. The one exception was 2022-23.
-
Elo plus a season-margin blend, in a lean 21-feature model
Replacing a 42-feature set with Elo, a blended season margin and a short list of form, rest and availability inputs improved every headline metric out of sample.
-
Removed betting lines from the forecast
The model used to lean toward the Vegas spread and total. It now stands on its own, so it can be compared with the market honestly.
-
Schedule load, travel and per-team home court
Games in the last four or six nights, travel distance, time zones and team-specific home-court strength were each tested and did not improve accuracy.
-
Team model only; the player model is kept for display
Blending summed player predictions into the game score cost correct winners in every slice tested, so official forecasts now use the team model alone.