Can machine learning beat the Norwegian stock market?

Abstract

Multiple studies on the performance of machine-learning stock portfolios have shown the efficacy of machine-learning portfolios on large stock exchanges, especially the American- and Chinese market. Fewer studies have been conducted on smaller cap markets, which consists of smaller, less-liquid investment options. The purpose of this thesis is therefore to explore the possibilities to beat the Norwegian stock market using machine-learning modalities. Eight different machine-learning portfolios have been constructed based on probability outputs of support vector machines, random forests and logistic regression created using the R software and packages “e1071”, “randomForest”, “gbm” and “caret”. Portfolios are tested from the end of 2013 to the end of 2022. Results of the thesis are in line with previous research that apply machine learning on the Oslo stock exchange for early periods in the sample, but find different results for the extended period. Machine-learning portfolios with monthly holding periods perform well before 2020, particularly the random forest portfolio. They do however lose their predictive power after this period and generate negative return beginning in 2021. Returns from daily portfolios are eaten up by transaction costs in multiple periods before 2020 and thus fail to consistently outperform the market. Some daily portfolios so show promise in the later period where the monthly portfolios underperform. The thesis therefore concludes that while machine-learning does show some promise on the Norwegian stock market, they cannot be relied upon to generate consistent outperformance over the benchmark index.