Predicting And Analyzing The Outcome Of NBA Games Based On Maximum Entropy Principle

National Basketball Association (NBA) is the world's most influential basket-ball league, attracting basketball fans all over the world and is becoming more popular. Usually, competing teams are almost equally strong and yet several ob-scure factors affect the outcome of the game which makes forecasting the outcome very difficult. In this study, we make use of several independent features to build a predictive model based on maximum entropy principle to forecast the outcome of NBA games, with improved accuracy.Initially, to cater for the effect of missing players, we modified previous game scores taking into consideration the effect of the missing player. Then we construct-ed a voting matrix between the teams using the modified scores by making use of an algorithm akin to the PageRank algorithm. Then by using the power of the ma-trix method, we solved the voting matrix and got the real relative strengths of the teams. We then we used K-means clustering algorithm to discretize the selected feature data and constructed an NBAME model according to maximum entropy principle. Finally, we trained the NBAME model's parameters by Generalized Iter-ative Scaling (GIS) algorithm using the training set and tested the model's ability to predict the NBA games using a separate testing set. We calculate the probability home team winning a game and this probability informs our decision to forecast a win or loss. By varying the probability threshold above which we forecast a win for the home team, we are able to increase the model accuracy with a drawback of making predictions for fewer games.Our results reveal that when we set the threshold at 0.5 we make prediction for all games and the model prediction accuracy goes as high as 75.6%. However when we increase the threshold to 0.7, prediction accuracy increases to 84.8% but with fewer predictions made. The features selected for the games and NBAME model based on maximum entropy model can effectively predict the outcome of the NBA game, and the prediction accuracy of the NBAME model is higher than most traditional machine learning algorithms.