Inter-dependent LSTM: Baseball Game Prediction with Starting and Finishing Lineups

Abstract

With the wide availability of historical data from baseball games, one of the most popular sports, high accurate winner prediction has become a significant target of statistical analysis and machine learning. However, existing techniques for a pre-game prediction yield poor accuracies due to the incomplete player lists given in starting lineups and substitutions occurring during the game. We exploit the capability of Long Short-Term Memory (LSTM) in identifying hidden patterns of time series data to propose inter-dependent LSTM baseball game prediction with only the starting lineup information. Particularly, we preprocess historical data to generate a pair of pre-game and post-game records for each baseball game. The pre-game record indicates the incomplete player lists given in starting lineups, and the post-game one contains the list of all players who participated in the game. The inter-dependent LSTM model exploits the dependencies of the pairs to predict a game result with only pre-game input. Our experiment results show that the proposed model achieves up to 12% higher accuracy than the existing ones.