Predicting user activity in video games

Abstract

This paper describes the results of the implementation of a full cycle of solving the problem of predicting the number of online players in video games in a month based on his-torical and current data.
Within the framework of data mining, several forecasting models were tested, including Gradient Boosting (XGBoost, LightGBM), LinearSVR, and the SARIMA and SARIMAX time series models. According to the results of preliminary testing, it was found that the SARIMA and SARIMAX models provide the lowest average error rates (about 20–22%), while other methods had a significantly higher error (about 40%), which does not allow them to be effec-tively used to solve the problem.
The final testing of the SARIMA and SARIMAX models on the example of three games with different levels of popularity allowed us to conclude that, despite the slight lag of SARIMAX in terms of the average MAPE value, this model provides more stable forecasting results, better adapts to the characteristics of different types of games, and more accurately approaches real online values. Thus, the most optimal approach to predicting the future of online video games was the SARIMAX model, which demonstrates high accuracy, stability and flexibility.
The results obtained may be useful for further application in the gaming industry for the purpose of planning server loads, marketing campaigns and general monitoring of the dy-namics of interest in the product.