Leveraging ensemble learning-based stock preselection with multiobjective investment optimization for stepwise decision-supported portfolio management

Abstract Volatile market conditions, noisy price signals, and the need for coordinated decision-making make portfolio management more complicated. As a result, comprehensive approaches remain underexplored, primarily due to challenges in integrating information-driven methods for portfolio preselection and optimization, designing effective rebalancing strategies, and incorporating investor preferences. This study introduces the Automatic Investment Decision (AID)-Multiobjective Forensic-based Investigation (MOFBI)-eXtreme Gradient Boosting (XGB) model to construct high-return, low-risk portfolios based on investor preferences. The AID-MOFBI-XGB model combines the MOFBI optimization algorithm, XGB machine learning technique, and an expanded mean–variance strategy. The framework has two phases: stock preselection and portfolio allocation. Data from the Taiwan Stock Exchange Corporation, Taipei Exchange, and the Market Observation Post System were used for model training and testing. In the first phase, the FBI-XGB model predicts company profitability, selecting candidates with higher expected returns. In the second phase, these preselected assets are input into an expanded mean–variance model integrated with MOFBI to determine capital allocation. This stage leverages a newly developed algorithm tailored to address the expanded mean–variance problem, effectively decomposing and optimizing it while balancing multiple objectives. The Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) selects the optimal portfolio among equally optimal solutions. Backtesting with quarterly rebalancing strategies validated the model's performance, demonstrating that it outperformed the four benchmark models in terms of annualized returns and Sharpe ratios, indicating higher investment quality.