Journal of Resource Management and Decision Engineering In Press 2026 05 01 A Hybrid Approach Based on a Memory-Instance–Based Gated Transformer (MIGT) Algorithm and Metaheuristic Optimization for Portfolio Management with the Aim of Return Optimization and Risk Control A Hybrid Approach Based on a Memory-Instance–Based Gated Transformer (MIGT) Algorithm and Metaheuristic Optimization for Portfolio Management with the Aim of Return Optimization and Risk Control 1 20 EN Journal Article 2025 11 12 <p>The objective of this study is to propose an efficient hybrid framework for portfolio management that can simultaneously optimize investment returns and effectively control risk under both normal and turbulent market conditions. The primary focus is on improving the accuracy of asset return forecasting and transforming these forecasts into optimal portfolio weighting decisions. In this research, a hybrid approach is employed that combines a memory-instance–based gated transformer model for forecasting asset returns with a hybrid metaheuristic algorithm based on adaptive differential evolution and particle swarm optimization for portfolio optimization. Financial data from the Iranian capital market covering the period from 2016 to 2024 were collected and, after preprocessing, cleaning, and feature extraction, were entered into the modeling process. The mean absolute error in the test period was 0.0068, and the root mean square error was 0.0100; these values exhibited a standard deviation of less than 0.0004 in cross-validation, indicating prediction stability. The optimal portfolio obtained by integrating these forecasts with the hybrid metaheuristic algorithm achieved an annualized return of 0.325 and an annualized standard deviation of 0.238, resulting in a Sharpe ratio of 1.16 and a Sortino ratio of 1.61. Tail risk measures also remained at controlled levels, such that the value at risk (VaR) at the 0.95 confidence level was 0.021 and the conditional value at risk (CVaR) was calculated as 0.0316. The paired Wilcoxon test conducted to compare the proposed model with benchmark methods yielded statistics above the significance threshold (for the Sharpe ratio, z = 2.51 and p = 0.012; for the Sortino ratio, z = 2.78 and p = 0.005), indicating a statistically significant improvement in performance. Based on the findings, it can be concluded that the proposed hybrid framework is capable of establishing an appropriate balance between return and risk in portfolio management by leveraging deep learning and metaheuristic optimization. The stability of the results across different temporal subsamples and the maintenance of acceptable performance under turbulent market conditions indicate that this approach possesses strong generalizability and practical applicability.</p> https://journalrmde.com/index.php/jrmde/article/download/255/250