Deep Reinforcement Learning & Feature Extraction For Constructing Alpha Generating Equity Portfolios

Abstract

The ambition of this paper is to catch hidden information inside the Securities and Exchange Commission’s (SEC)13F public holding data in order to construct an equity portfolio that maximizes returns. The 13F ling data give us the quarterly stock trading decisions of included funds, but we’re not given any insight on how they made their decisions or if information has been shared between funds. To remedy this lack of knowledge, this paper used feature extraction in order to lter out the best performing funds through several criteria. We propose a method employing powerful machine learning techniques (Deep Reinforcement Learning) to try to catch the missing pieces of information behind the decision process and use them as a prediction tool to construct quarterly equity portfolios. This approach reached an annualized return of 21% with a sharpe ratio of 1.8 outperforming the S&P 500 both in returns and stability through historical backtesting.