PORTFOLIO OPTIMIZATION VIA HYPER-HEURISTICS: EXPLORING A CUSTOMHYS ENHANCED FRAMEWORK

This study investigates the CUSTOMHyS framework, a Simulated Annealing-based hyper-heuristic, for portfolio optimization with cardinality and bound constraints. The framework is extended with nine population initialization strategies based on low-discrepancy sequences, stratified sampling, and probability distributions. This extension mitigates the limited diversity and clustering effects observed with the default pseudo-random initialization. Experiments on the Hang Seng, DAX, and Nikkei instances show that the Exponential and Weibull strategies systematically outperform the random baseline across interpolation error metrics, with larger gains on higher-dimensional cases. The best-performing configuration attains interpolation errors comparable to those of state-of-the-art metaheuristics, while relying on an automatically tailored metaheuristic rather than a problem-specific design. These results identify population initialization as a key design component in hyper-heuristic search. They also show that improving this stage can yield substantial performance gains without altering the underlying search operators or objective function.