Price prediction and forecasting for items in the online game Path of Exile using Machine and Deep Learning methods

Abstract

Making the most profit in an online game with the least time investment is perhaps one of the most asked questions new players have. It requires a vast amount of knowledge especially for complex games with in-game economies. Using Path of Exile as our game of choice, we have built a dataset of in-game item transactions using its built-in API and using machine learning we analyse the dataset to find ways to reduce the knowledge needed for new players to generate currency faster and more efficiently through buy and sell transactions. After downloading 6 months‘ worth of data from the API feed, we apply a pipeline in which we read the data, convert it to a readable format using a document based database - since natively the feed schema is highly nested- and use statistics and domain knowledge to engineer features which will later help our machine learning models perform better. We finally produce two datasets: - A dataset comprising of items that present a firm set of attributes through time. We use Recurrent Neural Networks in the form of LSTM to build models to try and forecast the future prices of these items. These models can later be used to help players decide where to buy or sell the items according to our forecast model. - A dataset comprising of items that present a variable number of features. We use different machine learning algorithms as well as Artificial Neural Network models to predict the price of an item or classify it in range price clusters meaningful to a player. A player can later use these estimators, to get information on the price of items and decide if an item is valuable and how much.