Out-Of-Sample Validation and Distributional Robustness
Thursday, October 04, 2018 at 4:15pm to 5:15pm
Bart Van Parys
Abstract: This talk deals with the problem of overfitting in data-driven decision-making. Decisions based on one particular dataset indeed often have poor out-of-sample performance; a phenomenon commonly denoted as the "curse of optimization''. Distributional robust optimization has quite recently been identified as one particular method enjoying good out-of-sample performance. In this talk we argue that the reverse is true as well. Any data-driven decision method enjoying good out-of-sample performance must necesarrily be dominated by a distributional robust decision formulation. Distributional robustness for out-of-sample performance is hence a natural choice.
Bio: Bart is an assistant professor at the MIT Sloan School of Management. His current research interests are on the interface between optimization and machine learning. Bart did obtained his Ph.D. at ETH Zurich and was a SNSF Postdoctoral fellow at the MIT Operations Research Center.