Machine Learning til Reward-Risk-Timing: En undersøgelse af muligheden for at anvende machine learning til at optimere sin handelsstrategi i et walk-forward optimization setup på det amerikanske aktiemarked

Abstract

This thesis explores the potential of applying machine learning techniques to optimal portfolio allocation with the aim of achieving higher risk-adjusted returns compared to traditional models. Modern portfolio theory provides the theoretical foundation for portfolio optimization, emphasizing the importance of accurate estimates of expected returns and volatilities. In light of the limitations associated with traditional estimation methods, this study investigates how machine learning can enhance the precision of such estimates by leveraging advanced techniques capable of handling large datasets and identifying complex patterns. Unlike most existing research in this field, this thesis focuses on predicting both expected returns and expected volatilities, inspired by the article Machine Learning Portfolio Allocation. In the empirical analysis, both simple methods based on the empirical mean, variance, and covariance, and the advanced machine learning models, including Random Forest and Recurrent Neural Networks are employed. The analysis uses data from all individual stocks in the S&P 500 index over the period from 2017 to 2025. A walk-forward optimization framework with rolling windows is implemented to simulate a realistic investment process. The results indicate that the machine learning models generally outperform the market, particularly when trained on the full dataset without a separate validation period due to more information. However, the predictive accuracy of stock returns remains limited, especially in comparison with existing studies, where both the length of the analysis period and the structure of the investment universe play crucial roles. The study concludes that machine learning techniques can improve the quality of financial forecasts and, thereby, support the development of more effective and robust trading strategies. Nevertheless, their performance depends heavily on the choice of input features and the availability of sufficient data to reflect market dynamics and various types of risk. This research contributes to the growing body of literature on machine learning in finance by highlighting key methodological considerations in the pursuit of accurate return predictions for stocks, portfolios, and indices.