ESTIMATION AND OPTIMIZATION OF LOGIT DEMAND MODEL WITH COVARIATES, CENSORED DATA, AND AUXILIARY INFORMATION

Abstract

We formulate a newsvendor model where demand depends on a set of stochastically changing covariates, and show that Bayesian estimation of this model, given a history of covariates, censored sales data, and auxiliary information such as customer traffic flow, can be implemented efficiently using a Metropolis-within-Gibbs Markov Chain Monte Carlo algorithm. We benchmark the impact of demand censoring, ignoring covariates, and incorporating traffic flow information on the out-of-sample performance of an optimal inventory decision. We show in an extended example that improvements in profit can be substantial, even if an incomplete set of covariates is being used, and that initiatives to improve demand forecasting by collecting relevant side information and combining multiple data sets can have a substantial impact on operational efficiency.