Time series prediction for algorithmic rescaling in the cloud

Abstract: The main goal of this thesis is to predict the number of players in some instance of DICE’s gaming platforms, e.g. Battlefield 3 for PC, 15 minutes in the future. This prediction may be used by the company when buying on-demand cloud servers to tell them how much they need to buy. Several different prediction models are examined and evaluated on historical data from 2011 to 2015. Another focus point is how to detect and handle outliers. The thesis also tries to create a general understanding of player behaviour by using different data separations and statistical methods. From the company’s perspective an auto regressive model of order 100 produces the best result. It is shown that with this model it would be beneficial for the company to use cloud servers instead of physical servers if their mean abundance in server capacity with physical servers is greater than 4134 players. The work was conducted primarily at DICE (EA Digital Illusions CE AB ) in Stockholm from 2015-09-07 to 2016-01-29.