Equity Risk Premium Prediction using Machine Learning Techniques and its Application in Portfolio Optimisation

Abstract

This research project delves into the domain of finance and investment management, with a focus on accurately predicting equity risk premiums and their subsequent application in portfolio optimisation - in the South African Context. The primary objective is to establish a robust framework that employs machine learning techniques to predict equity premiums effectively and utilise these predictions for portfolio optimisation. The study aims to implement two efficient variable selection techniques from the field of machine learning—specifically, Elastic Net Regression and Random Forests—to identify the most influential variables capturing variations in equity returns. These selected variables will then be integrated into Support Vector Regression and eXtreme Gradient Boosting models to predict equity returns and assess their predictive efficacy. Spanning a comprehensive dataset from January 1990 to December 2022, the study sources data from various repositories, including the IRESS Expert database, Federal Reserve Economic Data, and Yahoo Finance. This dataset encompasses monthly closing values for 10 companies listed on the Johannesburg Stock Exchange, 31 economic variables, and 10 international indices, facilitating a robust and multifaceted analysis. Methodologically, the research integrates machine learning techniques within an Arbitrage Pricing Theory framework. It rigorously assesses models using subsets of variables against those utilising the entire set, thoroughly examining their performance metrics and robustness. Key findings highlight the effectiveness of Random Forests in selecting pivotal variables for constructing both linear and non-linear equity premium prediction models. Notably, models trained with selected subsets exhibit superior performance, underscoring the critical importance of variable selection for heightened predictive accuracy. Significant recommendations stemming from this study emphasise the importance of diversification strategies in mitigating risks associated with concentrated positions. Furthermore, it stresses the necessity for a balanced approach in portfolio strategies to optimise outcomes. The research suggests a potential avenue for further exploration: a deeper investigation into sophisticated ensemble methods or advanced neural network architectures. This exploration aims to enhance model accuracy, generalisability, and interpretability, essential for establishing trust in complex models within the domain of finance and investment management.