Work in Progress

Econometrics and Machine Learning

Robust GANs for Downstream Decisions

Joint with F. Palomba

Generative adversarial networks (GANs) are widely used to generate realistic-looking synthetic data, increasingly employed to inform downstream decisions. We challenge the common assumption that the decision-relevant data-generating process coincides with the training distribution. We study decision regret when these distributions differ by at most (\delta), considering both population and finite-sample settings.

A Projection-Based Approach to Adaptive Bandwidth Selection

Joint with M. D. Cattaneo

We develop a novel adaptive method for bandwidth selection in kernel density estimation. Our approach exploits the geometry of Sobolev spaces to construct a data-driven assessment of smoothing bias, allowing bandwidth selection without prior knowledge of the density’s smoothness.

Economic Theory

Cautious Completions of Mutual Information

Joint with S. Dotta

We characterize convex and dilution-linear completions of the mutual information order and investigate the observable patterns of information acquisition induced by the resulting information costs.