Forecasting promotional sales within the neighbourhood

Abstract

Promotions are a widely used strategy to engage consumers and as such, retailers dedicate immense effort and resources to their planning and forecasting. This paper introduces a novel interpretable machine learning method specifically tailored to the automatic prediction of promotional sales in real-market applications. Particularly, we present a fully-automated weighted k-nearest neighbours where the distances are calculated based on a feature selection process that focuses on the similarity of promotional sales. The method learns online, thereby avoiding the model being retrained and redeployed. It is robust and able to infer the mechanisms leading to sales as demonstrated on detailed surrogate models. Also, to validate this method, real market data provided by a worldwide retailer have been used, covering numerous categories from three different countries and several types of stores. The algorithm is benchmarked against an ensemble of regression trees and the forecast provided by the retailer and it outperforms both on a merit figure composed, not only by the mean absolute error, but also by the error deviations used in retail business. The proposed method significantly improves the accuracy of the forecast in many diverse categories and geographical locations, yielding significant and operative benefits for supply chains. Additionally, we briefly discuss in the Appendix how to deploy our method as a RESTful service in a production environment.