New College Basketball model – 56.7% on 4 test seasons!
When the 2020 college basketball season abruptly ended we set out to program a college basketball model to improve upon the methods I’ve been using for years, which were less sophisticated in terms of priors and in-season blending with priors, while automating the player adjustment factors that I’ve been calculating by hand for years.
The results on 4 test seasons not used in the derivation of the model were even better than we expected, going 832-636-29 (56.7%) on side plays, while having a play on 11.7% of the games in the sample (see the full results below).
The key to having a good college basketball model is having good priors, which are the expectation of each team’s performance level entering the season. Our college basketball priors are derived from a blend of program, coach, and player ratings. The program and coach ratings are based on year-to-year correlations across a variety of metrics. The player ratings are derived from recruiting rankings and aging curves. We project every player’s contribution per 100 possessions and then predict playing time for each player to come up with all of the player value priors. The combination of all three ratings (program, coach, player) measure the changes of the offseason, which mainly come from transfers, new players (freshmen and JUCO), and coaching moves.
Once the priors were perfected the model was built using the same model building principles used in building the current NFL model, which has produced great results (41-22 on Best Bets this season and 58.3% over 5 seasons, through week 13).
The college basketball model was initially built using only data from the 2003-04 season through the 2015-16 season and then was tested on the next 4 seasons, which were not included in the derivation of the test model to avoid back-fitting.
The results of the model on the 2016 to 2020 seasons were incredibly good based on significant differences between our model prediction and the line (i.e. 3 ½ points or more difference).
2016-17: 217-149-5 (59.3%)
2017-18: 170-139-11 (55.0%)
2018-19: 244-18-7 (57.0%)
2019-20: 201-164-6 (55.1%)
Total: 832-636-29 (56.7%)
Those results exceeded expectations, especially considering that the model was tested on closing lines and we will be releasing most games in the morning when the lines are soft. The model test also did not include any adjustments for player personnel changes (injuries, lineup changes, etc), which is something that will be adjusted for and has been a big part of the success of the NFL model.
It’s certainly possible that the model will be as good going forward as it was in the 4 test years (56.7%) but it’s best for money management decisions not to overestimate your edge – so we’ll expect 55% going forward, which we believe is achievable. After all, I’ve been 53.6% in College Basketball over the years (for +353 Stars over 22 years) using far less sophisticated priors and in-season blending methods.
The model parameter weightings have been updated to include the last 4 seasons that were initially excluded, so the model is fine-tuned for the upcoming season. I’m really excited about the model and expect to have great results in college basketball going forward.
The season starts on November 25th and we’ll have plays starting opening night based on priors which are certainly better than the market (early season model plays based solely on our priors were 174-121-4, 59%, in our 4 test seasons).