Selecting characteristics using the Adaptive Group Lasso on U.S. industries

Abstract

Throughout the years, hundreds of factors have been proposed to forecast stock returns. Cochrane (2011) referred to these factors as the "zoo of new factors." In this thesis, we consider 62 of these factors and analyze which of them provide incremental value when forecasting stock return in 12 U.S industries. We apply the Adaptive Group Lasso (AGL) method for model selection described by Freyberger, Neuhierl, and Weber (2018), and use the Classical Linear Regression Model (CLRM) as a benchmark. The AGL selects, on average, approximately three characteristics, while the linear approach selects 24. The results indicate that the AGL approach generates more accurate predictions when the sample size increases compared to the CLRM. Our analysis indicates that there is no superior method for model selection in our samples.