Machine learning for sports betting: Should model selection be based on accuracy or calibration?
Conor Walsh, Alok Joshi · Machine Learning with Applications
Main finding
Selecting models by calibration produced better simulated betting outcomes than selecting by accuracy. After the 2025 corrigendum, average returns were negative for both selection rules (about −9.8% for calibration versus −26.8% for accuracy), while the best single models reached about +23.1% and +10.9%.
Limitations
One test season, simulated bets without limits, and implementation errors that required a corrigendum which changed the headline numbers.
Study details
- Authors
- Conor Walsh, Alok Joshi
- Published
- 2024
- Venue
- Machine Learning with Applications 16, 100539; corrigendum 19, 100627 (2025)
- Status
- Published
- Evidence type
- Peer-reviewed empirical
- Evidence strength
- Limited (our assessment of how much weight the finding can bear)
- Sport
- Basketball (NBA)
- Markets
- Bookmaker moneyline odds
- Method
- Trains machine-learning models on earlier NBA seasons, selects them either by accuracy or by calibration, then simulates betting on a held-out season.
- Sample
- Several NBA seasons for training and one held-out season for testing.
- Replication
- The authors found errors while releasing their code and published a corrigendum; the corrected figures supersede the original ones.
Read the original
doi:10.1016/j.mlwa.2024.100539 (opens in a new tab) · Free version (opens in a new tab)
Our summary is an original paraphrase for education. Verification: Bibliographic record checked against publisher, index or author listings. Checked 2026-10-25.