RL-NSGA-II-GRC boosts portfolio optimization with 5.8% better convergence on NASDAQ data

A team of researchers from multiple institutions has introduced RL-NSGA-II-GRC, a hybrid algorithm that integrates reinforcement learning (RL) with the Non-dominated Sorting Genetic Algorithm II (NSGA-II) and Gray Relational Coefficients (GRC). The method addresses the classic multi-objective optimization challenge of minimizing risk while maximizing return in portfolio selection. The RL agent dynamically adjusts NSGA-II parameters—such as mutation and crossover rates—based on real-time metrics like hypervolume, feasibility, and diversity. Meanwhile, the GRC-enhanced binary tournament operator ranks parent solutions using a unified score that considers dominance rank, crowding distance, and proximity to an ideal reference point, improving both convergence and diversity of the Pareto front.

Tested on standard benchmark functions (Kursawe and CONSTR), RL-NSGA-II-GRC achieved convergence improvements of approximately 5.8% and 4.4% over the base NSGA-II, while maintaining well-distributed solutions. The real-world application on NASDAQ portfolio optimization produced a smooth, densely populated efficient frontier with a maximum annualized Sharpe ratio of 1.92—a key metric for risk-adjusted returns. The algorithm also identified utility-optimal portfolios for different risk-aversion levels. Published in Mathematics (14(2), 296, 2026), the paper’s contributions include a novel RL-driven adaptation mechanism, a comprehensive selection operator, and empirical validation that the method yields actionable insights for quantitative finance.