Bootstrap robust prescriptive analytics

We address the problem of prescribing an optimal decision in a framework where cost function depends on uncertain parameters that need to be learned from data. Earlier work proposed prescriptive formulations based supervised machine learning methods. These methods can factor contextual information potentially large number covariates take context specific actions which are superior any static decision. When working with noisy or corrupt data, however, such nominal prone adverse overfitting phenomena and fail generalize out-of-sample In this paper we combine ideas robust optimization statistical bootstrap propose novel safeguard against overfitting. show indeed particular entropic counterpart guarantees good performance synthetic As data is often sensible proxy actual our interpreted directly encourage performance. The associated furthermore reduce convenient tractable convex problems local as nearest neighbors Nadaraya–Watson learning. illustrate data-driven decision-making robustness notion small newsvendor problem.