GAN-MP hybrid heuristic algorithm for non-convex portfolio optimization problem

Abstract During recent decades, the traditional Markowitz model has been extended for asset cardinality, active share, and tracking-error constraints, which were introduced to overcome the drawbacks of the original Markowitz model. The resulting optimization problems, however, are often very difficult to solve, whereas those of the original Markowitz model are easily solvable. In order to resolve the portfolio optimization problem for the new extensions, we developed a novel heuristic algorithm that combines GAN (Generative Adversarial Networks) with mathematical programming: the GAN-MP hybrid heuristic algorithm. To the best of our knowledge, this is the first attempt to bridge neural networks (NN) and mathematical programming to tackle a real-world portfolio optimization problem. Computational experiments with real-life stock data show that our algorithm significantly outperforms the existing non-linear optimization solvers.