Portfolio Optimization Using a Hybrid Machine Learning Stock Selection Model

Abstract

Portfolio Optimization can be challenging due to the uncertainty about the value of a future asset. With recent developments in machine learning, there are significant prediction tools available that can be applied to portfolio selection. Financial markets are known to be dynamic and complex, but algorithms are designed to capture patterns in the data. In this paper, seven machine learning techniques are used for stock price prediction: Linear Regression, Support Vector Machine, Random Forest, Recurrent Neural Network, Long Short-Term Memory, Bidirectional Long Short-Term Memory, and LightGBM. Additionally, two hybrid machine learning methods are used for prediction: CNN-LSTM and BiLSTM-BO-LightGBM. After training the models, the algorithm creates an optimal portfolio of assets over a simulated year of trading. The symmetric mean absolute percentage error of the algorithms on unseen data evaluates the prediction power. The generated alpha and Sharpe ratio evaluate the quality of the constructed optimal portfolios under Mean-Variance Optimization. Using data on the 50 largest United States companies from January 2, 2019 to December 29, 2023, the results demonstrate that the hybrid models perform better than the individual models, and the CNN-LSTM model outperforms benchmark market indices.