About the model

NCAABPredict estimates men's NCAA Division I basketball win probabilities from public historical schedules and scores. It uses a chronological team Elo model with a tuned home-court adjustment, season-to-season carryover, and no betting-market inputs.

Data

The current pipeline ingests the public sportsdataverse/hoopR-data men's basketball schedule master, which is ESPN-derived. The normalized snapshot contains 122,716 completed games across seasons 2002-2023. Raw play-by-play is not used.

Ratings

Games are processed in date order. Before each game, the model records a home-win probability from the two teams' current Elo ratings plus a home-court bonus (removed at neutral sites). After the game, ratings are updated by result. At season boundaries, ratings carry over partially toward the initial rating to reflect roster turnover.

Current tuned parameters: K=26, home edge=75 Elo points, carryover=0.9, probability temperature=1.

Performance

The performance page reports chronological pre-game log loss, Brier score, accuracy, and reliability buckets on completed games. These metrics are historical validation only; no claims are made about beating sportsbooks.

These are model probabilities for entertainment and analysis, not betting advice.