Factor-based Stock Selection and Portfolio Construction Utilizing Machine Learning Methods

Quantitative investment becomes the future direction of the financial market. More machine learning tools are used for stock price forecasting. Therefore, this paper combines traditional multi-factor stock selection models with machine learning. It filters fundamental and technical factors to construct a new factor-based stock selection model. It predicts stock price trends using LightGBM and Random Forest models, with parameter optimization performed using Bayesian optimization. In constituents of the S&P500 index, stocks with investment value are identified according to data over the past three years, and two effective investment portfolios are constructed. The study finds that in terms of prediction, LightGBM is faster in computation, but it is less accurate in trend forecasting compared to Random Forest. Random Fores exhibits a lag in predicting sudden changes in stock prices. Portfolios constructed using both machine learning models can generate excess returns, with the portfolio built using Random Forest offering higher returns and risks, making it suitable for more aggressive investors. LightGBM, on the other hand, provides better risk management. This study proposes some ideas and approaches for investors to predict stock prices, providing individual investors with a convenient method for forecasting.