Title:
Portfolio Selection Using Multiple Factors: A Machine Learning Approach

Authors:
Bilal Elmsili, Benaceur Outtaj, Morocco

Abstract:

In this paper we present a novel framework for portfolio selection using deep neural networks and elastic net regularization: At the beginning of each month T, we follow a three-step methodology. First, for each stock, we use the previous seven years of data in order to compute over 36 firm-specific factors. Second, we perform features selection using elastic net regularization. Finally, we train a deep neural network in order to learn portfolio weights and hold this portfolio until the end of the month T. Compared with momentum, long-term reversal, and short-term reversal strategies, our approach demonstrates a superior performance in terms of the monthly rate of return (2% versus 1.22% for long-term reversal, 1.15% for momentum, and only 0.68% for short-term reversal), Sharpe ratio (21.67% versus 19.31% for momentum, 15.51% for long-term reversal, and 8.69% for short-term reversal), and the monthly risk-adjusted return (1.85% versus 0.74% for momentum, 0.72% for long-term reversal, and 0.31% for short-term reversal). The results of our approach are all statistically significant at 1% level.

DOI: https://doi.org/10.51505/IJEBMR.2022.61207