Using machine learning to predict customer lifetime value of players in a freemium mobile game: Effect of seasonal features

Abstract

Freemium business model is currently largely used in the mobile gaming industry. The key idea of the model is that a game can be played for free, and revenue is generated through in-app purchases and advertising. However, the freemium model makes predicting the lifetime value of players, the amount of revenue they will generate, challenging as the revenue distribution is highly skewed and majority of revenue is generated by a relatively small group of spenders.

Predicting lifetime value of players (LTV) is one of the hottest topics in the freemium mobile games industry. Knowing how much revenue players brings games companies competitive advantage as it allows for better user acquisition optimization and financial planning, to name a few. Freemium games have several unique characteristics that set them apart from other similar fields such as online retail and traditional games such as high amount of behavioral data and high skewness of the data as only a very small share of players spend money.

This thesis has two objectives. First, different state-of-the-art machine learning models are compared to see which performs the best predicting lifetime values on a 360-day window. The models used haven been proven to be the most accurate by recent studies and include deep multilayer perceptron, random forest, gradient boosted trees as well as linear regression. The second goal of the thesis is to empirically test whether including seasonal features to the prediction dataset improves the model performance. Two different ways of using seasonal features is tested. The first approach is one-hot encoding and second applying sine and cosine transformations to make the seasonal features cyclical, representing better real-life situation. To the knowledge of the author, this is the first time these methods is used literature in freemium game setting.

Results show that deep multilayer perceptron performs the best, standing apart from the other models. This suggests that there are some complex relationships in the data that simpler models cannot capture. Against expectations, including seasonal features do not improve performance of most of the models.