Deep Learning with Gated Recurrent Unit Recurrent Neural Networks and Multi-Objective Optimisation for Portfolio Management

A large body of literature to-date suggests Fama and French Five Factor Model (FFFFM) as the “best of regression” algorithms for financial stock return regression. In the present paper, we will study whether the nonlinear regressions like deep learning - Gated Recurrent Unit Recurrent Neural networks (GRU RNNs) and machine learning - Support Vector Regression (SVR) can improve the prediction using Fama and French Five Factors. Our experiments using financial data showed that deep learning - GRU RNNs performed the best among the three models, and the final returns of the Portfolio using optimal weights showed it achieved the best return prediction among them. We proposed a novel portfolio management framework based on stock prediction using GRU RNNs and multi-objective optimisation to help investors to maximize their profit with minimal risk.