A Fundamental Analysis of Stock Returns using Machine Learning Algorithms

Predicting stocks has always been an attractive prospect for researchers. This study aims to create a more replicable model that combines machine learning and fundamental financial features to predict long-term stock movement better. We test a support vector machine (SVM) and random forest (RF) model via a regression and classification task of 432 stocks selected from the NYSE and NASDAQ. The results show that both models are best at predicting stock movement 1 quarter in the future, with the best root mean square error (RMSE) of 0.119. The models perform the best at predicting stock directionality when predicting 5 quarters in the future as the SVM peaks at 77.3% accuracy. We also find key features by using a permutation feature importance algorithm. Our models show significant improvement over baseline decision-making (50% accuracy classification) considerably far into the future (>1 year), helping provide better indicators for long-term stock returns.

I. Introduction

Financial markets are at the core of the economy, having a significant impact on all sectors [1]. Thus, predicting stock prices has always been an attractive topic for researchers and investors because of their use as an indicator of the economy and an option for investment and returns [2].