Enhancing Fund Selection Using Supervised Machine Learning : Evidence From the Nordic Mutual Fund Market
MetadataShow full item record
- Master Thesis 
In this research we aim to extend the literature on the performance predictability in actively managed mutual funds. We use the Nordic mutual fund market as our laboratory. We develop a performance-enhancing system to assist retail investors in selecting mutual funds by utilizing gradient boosting, random forest, and deep neural networks. Furthermore, we seek to obtain positive abnormal returns from our predicted quintile portfolios. We thus retrieve data free of survivorship bias for 2748 Nordic mutual funds from Morningstar Direct. First, we run the algorithms to test the possibility of classifying alphas. Secondly, we create a ranking system that categorizes funds based on predicted alpha, enabling us to separate the best from the worst-performing mutual funds. At last, we benchmark our findings against Morningstar’s acknowledged rating platform to examine whether our top quintile portfolios manage to outperform Morningstar’s top quintile portfolio. We find that our models can classify the sign of the alpha coefficient, whereas gradient boosting and random forest does this exceptionally well. Further, we manage to create a categorization system significantly outperforming both an equally weighted and asset weighted benchmark on risk-adjusted returns. Finally, our best performing portfolios generate risk-adjusted returns in excess of Morningstar, although only significantly for gradient boosting. Results are further robust to changes in risk-adjustment models for both equity funds and fixed income funds. The findings are consistent with the current machine learning literature and enable us to state that machine learning algorithms can be used to select successful mutual funds.