As financial markets grow increasingly complex, there is a need for efficient machine learning methods capable of producing reliable stock price predictions. However, implementing machine learning models can be challenging as it requires significant domain expertise and computational resources.
This study aims to contribute to overcoming these challenges by exploring the effectiveness of tree-based ensemble techniques and the PyCaret automated machine learning framework for forecasting S&P 500 index price movements.
Using a historical dataset spanning from 2015 to 2024, the research focuses on an univariate approach with one-step ahead and multi-step ahead forecasting for the Close index feature. Five tree-based ensemble models—Random Forest, Extreme Gradient Boosting, Light Gradient Boosting Machine, Categorical Boosting, and Adaptive Boosting—are evaluated using walk-forward cross-validation with a sliding window splitter technique.
Highlighting the superior performance of Light Gradient Boosting Machine and Categorical Boosting, with Categorical Boosting emerging as the best-performing model, achieving the lowest Root Mean Squared Error (181.5) and Mean Absolute Error (152.9). Interestingly, combining these top models into a hybrid approach resulted in diminished performance compared to Categorical Boosting’s performance. Recognizing the challenges of financial time series data, the study leverages PyCaret to streamline model development, cross-validation, and hyperparameter tuning, demonstrating its ability to simplify processes and conserve computational resources. The findings contribute to financial forecasting by providing a scalable and efficient framework for investment strategies, catering to both machine learning experts and non-experts.
Please check the presentation file for more details.