Machine Learning for NBA Betting: An Analysis and Exploitation

Abstract

The recent legalization of sports betting in the United States has led to rapid industry growth across the nation, with significant implications for economics, entertainment, and individual well-beings. This paper investigates the potential profitability of systematic betting strategies on NBA games, leveraging historical data, player statistics, and odds records. We first present analyses of odds records to corroborate arguments from existing literature regarding these markets' characteristics. We then apply machine learning techniques, namely neural networks, to these data with the goal of predicting game outcomes. We finally translate these predictions into viable betting strategies, both directly and sequentially. The best of these models achieves an accuracy of 62.34% on game outcomes, and the best strategy achieves an annualized 7.26% return over a betting history of 17 years; we thus demonstrate the feasibility of a consistently profitable, systematic betting strategy. Beyond achieving simple profitability, our analysis contributes to a growing understanding of pricing mechanisms and the consistent, likely intentional inefficiencies in these markets. As these markets continue to evolve, ongoing analysis will be essential for both profit-seekers and for those concerned with fairness and regulation.