Discrete level-bundle method for mean-CVaR portfolio optimization with cardinality constraint

Liping Pang,Shuangyu Li

DOI: https://doi.org/10.1007/s11075-024-01947-0

IF: 2.37

2024-09-28

Numerical Algorithms

Abstract:In this paper, we propose a novel bundle method, called discrete level-bundle (DLB) method , for solving mean conditional value-at-risk (mean-CVaR) portfolio optimization problem with cardinality constraint. Due to the potential large number of asset return scenarios and the discrete nature of the cardinality constraint, dealing with the cardinality-constrained mean-CVaR portfolio model is challenging. The DLB method reformulates the model as a two-stage optimization problem by separating the discrete variable, and then treats the second-stage problem as an embedded procedure to provide first-order information for the objective function of the first-stage problem, distinguishing it from conventional bundle methods used in continuous optimization. More importantly, despite tackling discrete problem, we still achieve the theoretical iteration complexity of the DLB method. Preliminary numerical experiments demonstrate that, compared with existing methods, the DLB method is highly effective and exhibits significant superiority, especially in handling problems with a large number of scenarios. This is the first time that bundle methods have been extended to discrete optimization problems.

mathematics, applied