Exploring Machine Learning for Stock Prediction: An Investigation Into the Potential and Limitations of Predictive Models

Abstract

This thesis explores the predictive capabilities of advanced machine learning models in identifying outperforming stocks within an evolving market landscape, which is shifting from active to passive investment strategies. The study analyses a dataset spanning from 2006 to 2023, including macroeconomic variables, market indicators, financial fundamentals, and relative valuation metrics, focusing on 243 stocks classified under the Standard & Poor's industry categories excluding the financial sector. Several machine learning models, such as Decision Tree, Adaptive Boosting, Random Forest, Gradient Boosted Machine, and Multilayer Perceptron, are evaluated for their effectiveness in predicting stock performance. The findings reveal that while tree-based models, particularly Decision Trees, demonstrate notable success with an annualised Sharpe ratio of 0.96 and an alpha return of 7.9%, the overall effectiveness of machine learning models varies significantly under different market conditions and among different models. Moreover, the tree-based machine learning models show greater proficiency in capturing the complex dynamics of financial markets compared to simpler linear models and Multilayer Perceptron. However, the variability in performance and the modest improvement over traditional buy-and-hold strategies highlights the limitations with the obtained machine learning models. This thesis contributes to the empirical research on machine learning in financial asset pricing and suggest areas for further exploration to enhance the robustness and applicability of these models in real-world scenarios.