Designing and Utilizing Machine Learning in Investment Strategies

Abstract

This thesis presents a comprehensive exploration of a machine learning (ML)-based trading strategy using the XGBoost regressor to predict daily stock price percent changes to be utilized in an investment strategy. The methodology involves data acquisition from Yahoo Finance and preprocessing to make it suitable for ML applications, focusing on predicting percentage changes rather than absolute prices. A critical aspect of the strategy is the selection of the top five stocks predicted to yield the highest returns, which are then used as daily investments.

The trading model is evaluated against the S&P 500 index, serving as a baseline, with results showing varying levels of success. While the model does not consistently predict price changes accurately, it demonstrates potential in identifying stocks that outperform the market average when focusing on the top five predictions. The strategy’s performance is notably variable, with profitability heavily dependent on low commission rates.

Backtesting across different periods and integrating transaction costs are used to validate the strategy’s robustness and identify its sensitivity to commission rates. Future work could enhance this strategy by incorporating more granular data and additional financial indicators and exploring other ML model architectures. The potential integration of real-time data could also provide further insights into the strategy’s performance in a live market environment.