Predicting the Number of Video Game Players on the Steam Platform Using Machine Learning and Time Lagged Features

Abstract

Predicting player count can provide game developers with valuable insights into players’ behavior and trends on the game population, helping with strategic decision-making. Therefore, it is important for the prediction to be as accurate as possible. Using the game's metadata can help with predicting accuracy, but they stay the same most of the time and do not have enough temporal context. This study explores the use of machine learning with lagged features on top of using metadata and aims to improve accuracy in predicting daily player count, using data from top 100 games from Steam, one of the biggest game distribution platforms. Several combinations of feature selection methods and machine learning models were tested to find which one has the best performance. Experiments on a dataset from multiple games show that Random Forest model combined with Pearson's Correlation Feature Selection gives the best result, with R2 score of 0.9943, average R2 score above 0.9 across all combinations.