Stock Portfolio Optimization Using Metaheuristic and Deep Learning Algorithms: Evidence from Boursa Kuwait

There has been considerable interest in portfolio optimization, also known as POP, in the literature, which has led to the development of various theories and methodologies designed to enhance its accuracy and alignment with real-world conditions. This research delves into the resolution of POP by leveraging higher moments derived from expected return distributions, with the objective of precisely pinpointing a portfolio that generates positive returns while maintaining a lower risk profile favored by investors. The proposed model addresses this challenge by incorporating four key objectives: expected returns, skewness, conditional value at risk and kurtosis. Through converting the optimization problem into a non-linear programming system, the study leverages the daily stock prices of all stocks listed on Boursa Kuwait from January 2010 to December 2023 to identify the top 10 performing stocks through the utilization of the k-means clustering algorithm. Subsequently, the model employs long short-term memory to forecast the returns of the selected stocks and estimate their moments, which are then utilized by the genetic-algorithm and particle-swarm hybrid technique to determine the best investment allocations. Moreover, the algorithm integrates an automatic hyperparameter optimizer to fine-tune its settings and enhance performance. A comparative analysis with three other popular models validates the performance superiority of the study’s model using expected portfolio returns, risk levels, and risk-adjusted return ratios.