Due to a colossal soccer market, soccer analysis has attracted considerable attention from industry and academia. In-game outcome prediction has great potential in various applications such as game broadcasting, tactical decision making, and betting. In some sports, the method of directly predicting in-game outcomes based on the ongoing game state is already being used as a statistical tool. However, soccer is a sport with low-scoring games and frequent draws, which makes in-game prediction challenging. Most existing studies focus on pre-game prediction instead. This paper, however, proposes a two-stage method for soccer in-game outcome prediction, namely in-game outcome prediction (IGSOP). When the full length of a soccer game is divided into sufficiently small time frames, the goal scored by each team in each time frame can be modeled as a random variable following the Bernoulli distribution. In the first stage, IGSOP adopts state-based machine learning to predict the probability of a scoring goal in each future time frame. In the second stage, IGSOP simulates the remainder of the game to estimate the outcome of a game. This two-stage approach effectively captures the dynamic situation after a goal and the uncertainty in the late phase of a game. Chinese Super League data have been used for algorithm training and evaluation, and the results demonstrate that IGSOP outperforms existing methods, especially in predicting draws and prediction during final moments of games. IGSOP provides a novel perspective to solve the problem of in-game outcome prediction in soccer, which has a potential ripple effect on related research.
Keywords: Bernoulli distribution; in-game outcome prediction; machine learning; probability prediction; regression coefficients; soccer.