Title: The bet for similarity: Adaptive discrete smoothing with application in finance
Authors: Ye Luo - University of Florida (United States) [presenting]
Martin Spindler - University of Hamburg (Germany)
Victor Chernozhukov - MIT (United States)
Xi Chen - New York University (United States)
Abstract: The traditional linear panel data model does not allow individuals to have different predictive models except for different intercepts. This restriction is too strong to analyze modern data with heterogeneity in individual behaviors. We consider a scenario where individuals are allowed to have their own models. In practice, such a setting leads to undesirable prediction because the longitude of the data is usually short. We consider a set of algorithm called Adaptive Discrete Smoothing (or the ADS algorithm). The ADS algorithm tries to cluster similar individuals and utilize the data from the clustered group for prediction. This method is beneficial when the cross-sectional size of the panel data is large. The ADS algorithm can be applied to non-linear panel data, high-dimensional panel data and etc. We show that the ADS algorithm can substantially improve the theoretical convergence rate as well as the practical prediction performance in panel data. We apply the ADS algorithm to the Fama-French regression framework with monthly return data in the U.S. stock market.