AI Predicts Baseball Wins and Betting Returns

Baseball has evolved from a game of instinct to one driven by data, with teams using advanced analytics to gain a competitive edge. A new study applies machine learning to predict Major League Baseball game outcomes, offering insights not just for sports enthusiasts but for anyone interested in how data can inform real-world decisions. This research demonstrates that AI models can accurately forecast wins and connect these predictions to actual game dynamics, such as score differences, while also revealing strategies to avoid financial losses in betting.

The researchers found that several machine learning models significantly outperform a simple baseline that always predicts the home team wins, which has a 53.15% accuracy due to home-field advantage. Key models, including logistic regression, support vector machines, and artificial neural networks, achieved accuracies around 62-63% when tested on games from the 2016-2019 seasons. For instance, logistic regression reached 62.94% accuracy, showing that these AI systems can reliably identify winning teams based on historical player and team statistics.

To build these models, the team collected a comprehensive dataset spanning the 2001-2019 MLB regular seasons, excluding playoffs to avoid performance variations. They used 83 input variables, such as on-base percentage, earned run average, and win-loss records, sourced from reputable databases like Baseball Reference and FanGraphs. The models were trained on data from 2001-2015 and evaluated on held-out test data from 2016-2019, ensuring no overlap. Techniques like cross-validation and grid search optimized parameters, with models ranging from simple logistic regression to complex neural networks, all designed to predict whether the home team would win.

The results show that higher predicted win probabilities correlate with larger score differentials. For example, logistic regression and neural networks had R-squared values of 0.140 and 0.120, respectively, indicating that as the predicted probability of a home win increases, the average margin of victory tends to rise. In toss-up games where predictions were near 50%, the average score differential was close to zero, but with high variability. Conversely, for games predicted as dominant wins (e.g., 75-100% probability), the average differential was around 2-4 runs, suggesting that models capture the likelihood of a decisive outcome. Ensembling multiple models could push accuracy toward 70%, though current methods show limited gains due to high agreement between top performers.

This research matters beyond baseball, illustrating how AI can translate probabilities into actionable insights for decision-making. In betting, for instance, a naive strategy of betting on every game led to a 51.39% loss, but optimizing cutoffs for when to bet based on model predictions turned losses into gains of over 10% in some cases. This approach reduces risk by focusing on high-confidence predictions, applicable to fields like finance or healthcare where probability-based choices are crucial. It also highlights the value of interpretable models, like logistic regression, for scenarios where understanding AI decisions is key.

Limitations include the models' inability to achieve perfect accuracy, with a theoretical maximum around 75% due to the unpredictable nature of sports. The study notes that while probabilities relate to score differentials, individual game outcomes remain volatile, and factors like player injuries or sudden performance changes are not fully captured. Future work could explore dynamic betting strategies or alternative modeling techniques to bridge this gap.