Deep Neural Networks on Nonparametric Regression

Abstract

This thesis investigates how deep learning, specifically Deep Neural Networks (DNNs), can be theoretically understood and effectively applied to predict stock returns in the context of nonparametric regression. Empirically, the study shows that DNNs outperform traditional methods in forecasting U.S. stock returns, identifying key predictive stock characteristics. Theoretically, it develops a rigorous mathematical framework to explain why DNNs are effective at modeling complex relationships. By combining empirical success with theoretical insights, the thesis advances both machine learning and econometrics, offering a deeper understanding of DNNs for future research and real-world data analysis.