A prescriptive machine learning framework to the price-setting newsvendor problem Pavithra Harsha IBM T. J. Watson Research Center, Yorktown Heights, NY 10598,
[email protected] Ramesh Natarajan IBM T. J. Watson Research Center (past affiliation when this research was done) Amazon, 515 Westlake Avenue N, Seattle, WA 98109 ,
[email protected] Dharmashankar Subramanian IBM T. J. Watson Research Center, Yorktown Heights, NY 10598,
[email protected] We develop a practical framework for modeling the price-setting newsvendor problem, which includes statistical estimation and price optimization methods for estimating the optimal solutions and associated confidence intervals. We present a novel and exact reformulation of the problem that leads to the framework which requires as input the estimates of only three distinct aspects of the demand distribution: the mean, quantile and superquantile (in contrast to the full-demand distribution), and it provides asymptotically optimal solutions under mild conditions, if these estimates are consistent. To estimate these quantities in a data-driven, distribution-free fashion potentially with multi-dimensional observational datasets, we investigate statistical estimators based on generalized linear regression (GLR), mixed-quantile regression (MQR), and superquantile regression (SQR). We propose a novel and exact large-scale decomposition method that is computationally efficient for SQR, and extend the MQR estimation method by relaxing its implicit assumptions of homoskedasticity (these two extensions are of independent interest). Our computational experiments, first, indicate the importance of flexible statistical estimation methods that inherently account for heteroskedasticity, and second, suggest that quantile-based methods such as MQR and SQR provide better solutions for a wide range of demand distributions, although for certain location-scale demand distributions similar to the Normal distribution, GLR may be preferable.