Empirical asset pricing via machine learning in the context of Nasdaq Helsinki

Abstract

This master’s thesis investigated the predictive power of different machine learning models in the context of Helsinki stock exchange. The study focused on eight machine learning models and compared their predictive performance using monthly out-of-sample R^2 as a measure of predictive accuracy and Diebold-Mariano test for pairwise comparison between the models. Factor importance and portfolio management benefits were also analyzed for each model.

The results showed that more complex models tend to produce better predictions, with the four-layered neural network being the best performing model. The study suggests that more complex machine learning models can improve the accuracy of stock return predictions, but the shallower models were also producing promising accuracy in the predictive task. With regard to factor importance, the market dividend-price ratio appeared to be the most important, with higher importance scores across all models. Other features that appear to be relatively important include market book-to-market ratio and 30-day relative strength index.

The study also answered the research question on the benefits of machine learning for portfolio management applications in the Finnish stock market with cautiously promising results. The cumulative returns for machine learning portfolios were higher than the benchmark portfolio's, with smaller drawdowns and better risk-return profiles. However, the study has some limitations, including the small sample size of only 46 stocks traded in the Helsinki Stock Exchange and the limited set of factors considered. Further research is needed to confirm these results and explore the potential of other predictive factors and higher-frequency predictions.