Clustering Techniques

Abstract

This chapter discusses how to leverage clustering techniques in the context of demand prediction for retail applications. Specifically, our goal is to aggregate the data across different SKUs to improve the prediction accuracy. On the one hand, aggregating sales data across several SKUs will help reduce the noise and would allow the model to rely on a larger number of observations. On the other hand, this will overlook the fact that each SKU bears specific characteristics. Clustering techniques can be used to identify several groups (or clusters) of similar SKUs. Then, one can estimate a demand prediction model for each SKU by relying on the historical data from all the SKUs in the same cluster. We consider two common clustering techniques: k-means and DBSCAN and implement them using the accompanying dataset.