Investment portfolio selection using machine learning techniques

Abstract

The stock market is a crucial component of the financial market, influenced by various factors and exhibiting intrinsic patterns of change. Effective stock selec- tion is a key focus for researchers and practitioners. Traditional stock selection methods include multiple regression analysis and multi-factor scoring, but these have limitations such as sensitivity to outliers and reliance on subjective weight assignment, respectively. This paper explores the application of several ensem- ble learning methods for stock selection and integrates these with the Markowitz portfolio theory to optimize investment portfolios, thereby enhancing investment returns. The main contributions of this work are: (1) Constructing a stock fea- ture indicator database, selecting key features using the RankIC method, and conducting effectiveness and correlation analyses; (2) Utilizing various machine learning models, including Decision Trees, Random Forests, AdaBoost, and XG- Boost, to evaluate their predictive power and generalization ability; (3) Optimiz- ing the Markowitz portfolio model based on model predictions and proposing an improved portfolio optimization method that maintains high investment re- turns while reducing overall risk. Empirical analysis validates the effectiveness of the proposed methods, offering new approaches and tools for stock selection and portfolio optimization.