Abstract
Due to its ability to learn complex patterns in data, deep learning has become a very popular machine learning technique, lately, for various applications and business, sciences, and engineering. In this paper, we compare the effectiveness of deep neural networks (DNNs) with simple (or shallow) neural networks (SNNs) in making investment decisions in stocks. We use publicly available financial data on thousands of firms, to train both deep and simple neural networks to rank companies and apply various investment strategies to create paper portfolios of stocks using each technique, from the same pool of companies. We then compute the return on investment on the portfolio over five years. We find statistically significant difference in the returns on investment for portfolio of stocks created by DNNs vs. those created by SNNs. Portfolios with DNNs outperformed those with SNNs by 16 % over five years. We also looked at how DNNs and SNNs compared with the traditional logistic regression approach and found that DNNs also outperformed logistic regression approach statistically significantly. There was no statistically significant difference in the returns of portfolios created by SNNs and logistic regression. All three approaches (Deep NN, Simple NN and logistic regression) outperformed the returns from four major indices, viz., Dow Jones Industrial Average, Nasdaq Index, S&P 500 and Russell 2000 for the same five-year time period.