Predicting purchases and churn for paying customers using player behavioral data

Abstract

Mobile games have large amount of data available which contain very detailed information into player behavior.A common problem within free-to-play mobile games is players churning, that is players decide to leave games for various reasons and since they receive the game for free, they might never spend any money towards it and can stop playing it without notifying anybody, once they uninstall the game they might never return back to it again, this is considered as a well defined and definitive state of player churn. Companies are much more likely to be successful in encouraging active players to continue playing, rather then trying to convince players that have already churned to start playing again. This is why player retention and churn prevention is extremely important.For our implementation, we created a time window of seven days (moving week), starting from players first date of activity to his last date of activity. Each moving week starts the day after previous moving week starts and ends on the day after previous week ends. For a full year this would give us a total of 372 moving weeks instead of the typical 52 weeks (each day was thus considered seven times instead of once). Each moving week contains summarized data about events and actions made during each day in weeks period, this gives us much better information into exactly when player behavior changes then we would have had if we only considered 52 weeks.Using the moving weeks, we extracted features such as matches, in-game currency, streaks and social network.We then created two models, the first to predict players about to make a purchase and the second to predict for players about to churn from the game.