To what extent can machine learning algorithms predict long-term stock price directions on Oslo Børs and Nasdaq Stockholm?

Abstract

This study aims to examine what value Machine Learning algorithms give when trading from a long-term perspective. Historically, it has been hard to consistently gain an excess return from investing in the stock market. The high complexity and number of factors affecting the markets are complicating this task. In our research, the performance of Machine Learning algorithms such as Random Forest and Support Vector Machine were analyzed both with and without feature selection methods. The models’ predictions and our constructed portfolios were compared to two benchmarks (Dummy Classifier and OSEBX/OMX30 index). From our analyses, the Random Forest model with SVM-RFE feature selection was found to give the most promising prediction results, and the performance was analyzed both during the whole backtesting period and through times of crisis. When implementing a simple trading strategy utilizing the predictions, we found the same model with a portfolio construction of 30 companies to outperform both the benchmarks and other algorithms from late 2006 until the first quarter of 2021. During times of crisis, our reference Machine Learning model did not significantly outperform the benchmarks. However, it showed uplifting results in economic rebounds. Thus, the highest potential for the Machine Learning model might be its ability to identify the best-performing stocks in periods after financial recessions.