AU students use AI to handicap Kentucky Derby

American University professor Matthew Bakowicz had 32 students use artificial intelligence tools and a Python interface to handicap the Kentucky Derby. The class modeled wagers with a mock $1,000 bankroll across six races at Churchill Downs, allocating $500 to the Derby. The algorithm’s Derby ticket recommended $100 on Commandment to win and $50 to place, with additional bets on Danon Bourbon, Chief Wallabee and Silent Tactic.

Bakowicz, a former DraftKings sportsbook and racebook operations manager, designed the project for his Introduction to Sport, Gaming, and Entertainment course to demonstrate how AI can inform business models. He described the exercise as ‘a cool classroom project’ that parallels models used in finance and real estate.

Students used Perplexity, Claude and ChatGPT to analyze data and built a Python interface to turn AI outputs into betting recommendations. The model’s sixth iteration identified Commandment as the most likely winner, flagged Silent Tactic as a value pick to place and listed Chief Wallabee as likely to show. Silent Tactic was scratched before the race.

Work for the project included compiling past performances, workout and race times, historical weather and public betting markets. Graduate student Camden Egan collected much of the raw data. Student coders led by Mikias Goshime created a weighting system that allowed variables to be adjusted; the team applied a 70% weight to historical performance and 30% to tendencies.

The group refined the model by inputting results from non-Derby races until outputs began to mirror actual outcomes, then tested it at Laurel Park in Maryland in late March, where the team recorded four winners. Bakowicz said the model was intended for classroom learning and not to exploit pari-mutuel wagering pools.

The project took place amid industry debate over computer-assisted wagering systems, which use predictive models and late large bets that can move pari-mutuel odds. Bakowicz and handicapper colleagues, including retired American University professor Kevin Boyle, planned a ‘man versus machine’ comparison after final scratches to compare human judgment with the algorithm.

Bakowicz noted factors that limit any model’s accuracy for the Kentucky Derby: the 1 1/4-mile distance, a 20-horse field and the race’s prominence. He said those variables make a perfect algorithm unlikely.