Machine learning methods for predicting stock returns from financial and microeconomic variables

Abstract

This report analyses the ability of machine learning algorithms to predict next quarter stock returns based on macroeconomic features and financial, company-specific variables, by benchmarking achieved results in terms of root mean squared error against the expected return baseline that outputs as a prediction of period t return, the average of returns until t-1. The deployed study compares the performance of Random Forest, Support Vector Regression, Lasso Regression and Multi-layer perceptron regressor in modelling the predictive issue considered and analysis whether specific industries are better suited for prediction purposes. The achieved results show that there are no major differences in performance across models in terms of root mean squared error, but that Lasso regression fails to properly model the problem as it deploys the same prediction regardless of the values in the set of predictive features. Furthermore, on an industry level, the analysis shows that some industries are more prone to prediction, with the Health Care and Semiconductors sector displaying the worst results. What is more, considering the selected features, the models showed better performance levels either when the entire set of features was used, suggesting that reducing the number of features was not helpful for the model to deploy its predictions, or when principal component analysis was used. The results from the RF show that macroeconomic variables are, overall, more important than company-specific ones to predict next quarter’s returns, which may be at the basis of the similar performance across some of the industries and of the models.

Date of Award	20 Oct 2021
Original language	English
Awarding Institution	Universidade Católica Portuguesa
Supervisor	Viktor Pekar (Supervisor)