Bayesian Ensemble of Quantile Trees for Sales Prediction

Retail stores invest much effort in strategic level decisions, such as the type of store, its location and the best product proposal, to maximise their income. Sales product prediction in a particular store, is therefore one of the most challenging problem in retail commerce, either for commercialising new products or opening new stores. This paper proposes a new nonparametric method for predicting sales quantiles at different confidence levels, conditional on several explanatory variables such as the type of store, the location of the store, the type of product and its the price, etc, thereby providing a complete picture of the relation between the response and the covariates. Moreover, predicting extreme sales quantiles provide valuable information for building automatic stock management systems and for strategic level decisions of retail stores. As concerns the methodology, we propose to approximate the conditional quantile at level $\tau\in\left(0,1\right)$ of the response variable using an ensemble of nonparametric regression trees. Decision trees and their ensemble counterparts are promising alternatives to linear regression methods because of their superior ability to characterise nonlinear relationships and interactions among explanatory variables that is of fundamental relevance to get accurate predictions.