Variable Division and Optimization for Constrained Multiobjective Portfolio Problems

Yi Chen, Aimin ZhouarXiv 2021 · arXiv

Abstract

Variable Division and Optimization (D&O) is a strategy for solving high-dimensional multiobjective problems by partitioning decision variables into discrete and continuous components. In portfolio optimization, this corresponds to asset selection ($x_I$) and capital allocation ($x_{II}$). This paper presents D&O-MOEA/D, a hybrid framework that integrates the Multi-Objective Evolutionary Algorithm based on Decomposition (MOEA/D) with exact mathematical programming (CPLEX). By using scalarizing functions to evaluate the fitness of partial solutions, the proposed method effectively resolves the "dimensionality mismatch" between discrete choices and their associated local Pareto fronts. Experimental results on benchmark and real-world datasets demonstrate that D&O-MOEA/D significantly outperforms traditional heuristics and exact methods in large-scale, constrained scenarios.

The Challenge of Dimensionality Mismatch

In complex optimization, "Divide and Conquer" is a standard strategy. In Evolutionary Algorithms (EAs), this often takes the form of Variable Division and Optimization (D&O): splitting a high-dimensional decision variable $x$ into two parts, $x_I$ and $x_{II}$, and optimizing them separately.

A classic example is portfolio optimization:

  1. $x_I$: Which assets should I pick? (Combinatorial/Discrete)
  2. $x_{II}$: How much capital should I allocate to each? (Continuous)

The core challenge is elitist selection. To decide if a specific asset selection ($x_I$) is "good," you need to know its fitness. However, risk and return cannot be calculated without the weights ($x_{II}$). In multiobjective scenarios, a single $x_I$ doesn't have one "best" fitness; it has a whole local Pareto Front (PF) of possible weight distributions.

Local Pareto fronts for different asset selections.

As shown in the figure above, comparing different $x_I$ candidates is difficult because each corresponds to a set of points rather than a single point in the objective space. The authors address how to perform selection when comparing these partial variables (local PFs) against each other.

Methods: D&O-MOEA/D

The authors propose D&O-MOEA/D, a hybrid algorithm that bridges the gap between evolutionary search and exact mathematical programming.

  • Decomposition-Based Selection: They utilize the framework of MOEA/D, which decomposes a multiobjective problem into many scalar subproblems using weight vectors. This allows them to "reduce" a local Pareto Front into a single scalar value for a given weight vector, solving the dimensionality mismatch.
  • The Hybrid Loop:
    • Outer Loop (EA): Uses Differential Evolution (DE) and mutation to search the space of asset combinations ($x_I$).
  • Inner Loop (Math Programming): For every new asset combination generated by the EA, the algorithm calls an exact quadratic programming solver (CPLEX) to find the optimal weights ($x_{II}$) for that specific selection.
Selection of partial solutions using weight vectors.

By assigning each individual in the population to a specific direction in the objective space (defined by weight vectors $\lambda$), the "best" $x_I$ for that direction becomes clearly defined as the one whose local PF contains the best scalarized value.

Experimental Evaluation

The algorithm was evaluated on 20 constrained portfolio instances, including the OR-Library, NGINX (Yahoo Finance data), and a custom dataset of 4,510 funds from the Shanghai and Shenzhen stock markets. The problems include realistic constraints: cardinality (exactly $K$ assets), floor/ceiling bounds, pre-assignment, and "round lot" constraints.

Key Results

  • Scalability: While exact methods like AUGMECON2 are superior for small problems, they fail as the number of assets grows. D&O-MOEA/D acts as a bridge, using exact methods where they are fast (the continuous subtask) and heuristics where the search space is too complex for solvers.
  • Large-Scale Dominance: On large-scale instances (D13–D20), D&O-MOEA/D significantly outperformed other state-of-the-art MOEAs in both Hypervolume (HV) and Generational Distance (GD).
  • Real-World Performance: In tests on the Chinese stock market, portfolios generated by the algorithm achieved an expected return of ~13%, consistently outperforming the CSI300 index.

Insights and Future Work

The "Aha" moment of this research is the realization that MOEA/D's scalar subproblems provide a natural way to compare partial solutions. While exact mathematical methods hit a "complexity wall," this hybrid approach maintains efficiency; the inner CPLEX call takes only about 30 milliseconds.

However, the method's dependency on a fast inner solver is a limitation. If the subtask ($x_{II}$) were NP-hard or non-convex, efficiency would degrade. Future work will explore extending this framework to multi-period portfolio optimization and deepening the theoretical understanding of selection dynamics in partial variable spaces.

Summary updated

BibTeX
@misc{chen2021variable,
  author = {Yi Chen and Aimin Zhou},
  title = {{Variable Division and Optimization for Constrained Multiobjective Portfolio Problems}},
  year = {2021},
  eprint = {2101.08552},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2101.08552},
}
Figure