This thesis investigates whether a data-driven sports-betting strategy can capture sustainable, risk-adjusted profits in the Hong Kong Jockey Club’s pari-mutuel pools. A gradient-boosted model was trained on market odds, horse-specific attributes, race context, and indicators that measure bettor sentiment and bias to generate win-probability forecasts. Capital was then allocated using a Kelly-based position-sizing overlay that maximized return on investment, measured as internal rate of return (IRR) and multiple of money (MoM).

Across 100 Monte-Carlo back tests, the strategy produced positive expected profits, albeit with elevated variance. Although the model’s precision for predicting the winner exceeded the crowd-favorite’s predictions only marginally, profitability arose from systematically exploiting non-favorite opportunities that the crowd typically underestimates. The best parameter set delivered a mean IRR of 4.46% and a mean MoM of 0.09x.

These findings underscore three contributions. First, they show that modest predictive gains, when coupled with disciplined bet sizing, compound into economic value. Second, they demonstrate that behavioral-bias variables enhance model inference by revealing pockets of neglected value. Third, they provide quantitative bettors with a framework for mitigating losses caused by cognitive bias. The approach is readily extensible to other pari-mutuel domains such as greyhound racing, trotters, and exotic pools, offering a holistic template for scholars, professional bettors, and operators alike.