Rossmann Store Sales

The objective of this project is to forecast sales in euros at 1115 stores owned by Rossmann, a European pharmaceutical company. The input to our algorithm is a feature vector (discussed in section 3) of a single day of data for that store. We tried using a number of algorithms, but mainly gradient boosting, to output the predicted total sales in euros for the given store that day. Rossmann provides a massive 1.1 million example data set as part of a sponsored (by Rossmann) contest to see who can best predict their sales. Over 3500 groups have entered the contest, with the best group achieving 8.9% root mean squared error. This project is also being used for CS 221, and use different parts of the project for that class that includes K-means to group the data and an Markov decision process (MDP) to model store output over a period of time. For this class, our algorithm focused around regressive algorithms and decision-tree based regression algorithms like gradient boosting. As a baseline algorithm for both classes we used support vector regression (SVR) over an unprocessed feature set. However, to improve performance we used an augmented feature set. We also experimented with linear regression and neural nets with less success.