The rapidly evolving field of machine learning, fostered by the eruption of data availability and computing power, is auspicious for investment management. The ability of machine learning to uncover patterns in data holds particular promise for return prediction and thus, for the enhancement of trading strategies. One of the most pervasive trading strategies in the financial literature is the price momentum strategy. With its occasional but severe profit punctuations, the momentum strategy poses an interesting case for reformation through machine learning. Motivated by the unique potential of coupling “momentum and machines”, we investigate the performance of three trading strategies based on machine learning models with momentumrelated input variables. We find that such machine learning-based trading strategies outperform the conventional price momentum strategy and mitigate its profit punctuations considerably. Further, we document a large economic potential from an ensemble strategy that equally weighs the three machine learning-based trading strategies and the conventional price momentum strategy. The ensemble strategy outshines all individual strategies and approximately doubles the Sharpe ratio of the conventional momentum strategy. Thus, our findings offer supportive evidence for the benefits of applying machine learning to return prediction. However, the complex nature of machine learning models makes it difficult to draw meaningful interpretations of what drives their superior performance. Hence, while the machine learning models engender profitable trading strategies, their complexity induces limited interpretability. Consequently, the enigma of machine learning hinders the practical implementation of our findings.
|Educations||MSc in Finance and Strategic Management, (Graduate Programme) Final Thesis|
|Number of pages||124|
|Supervisors||Lasse Heje Pedersen|