Machine learning, stock market forecasting, and market efficiency: a comparative study

Analyses of the accuracy of machine learning algorithms for predicting stock market indices typically employ only one or a small set of stock indices and a small number of time periods. Moreover, these analyses generally ignore the influence of market efficiency on algorithm accuracy, despite theoretical evidence of the potential for such an effect. This paper aims to fill this gap by proposing and applying a method focused on the analysis of the accuracy of stock market prediction models in comparison with market efficiency. This method is applied to a dataset comprising 55 markets and 65 periods to compare the most accurate and frequently used algorithms for stock index prediction: support vector machines, artificial neural networks, gradient-boosted trees, Naive Bayes, random forest, logistic regression, and long-short term memory. These algorithms were trained to predict stock market indices using technical indicators as input. After analyzing the algorithms’ results, we present a detailed comparative analysis of their accuracy supported by a set of nonparametric measures. In addition, dynamic panel data analysis is used to determine whether there is a relationship between stock market’s efficiency and algorithmic precision. We highlight the algorithms that tend to show the best performance in terms of accuracy and obtain interesting results regarding market efficiency.