Large-Scale Portfolio Optimization Problem Under Cardinality Constraint With Enhanced Multi-Objective Evolutionary Algorithms
Source: arXiv:2607.09566 · Published 2026-07-10 · By Danial Ramezani, Mostafa Abouei Ardakan
TL;DR
This paper addresses the large-scale portfolio optimization problem with cardinality constraints, which is known to be NP-hard due to restrictions on the number of assets in the portfolio and allocation bounds. Traditional exact methods become inefficient for markets with many assets, so the authors propose enhancements to multi-objective evolutionary algorithms (MOEAs), specifically NSGA-II, to improve convergence speed and solution quality. The key innovations include a novel two-row chromosome encoding representing asset selections and weights, new constraint handling repair mechanisms, a new 'explorer' operator, and adaptive mating strategies focused on exploration early and exploitation late. The approach was tested on multiple benchmark financial indices (S&P 100, DAX 100, Nikkei 225, Tehran Stock Exchange) and demonstrated faster convergence trends and better approximations to the Pareto front compared to conventional MOEAs, without loss of performance as the number of assets increased in scale.
Key findings
- New chromosome representation with fixed size equals Kmax encodes asset selections and weights simultaneously, avoiding inefficient add/remove steps in feasibility repair.
- Knee-based tournament selection focusing on solutions near the utopia point accelerates convergence in early iterations by guiding exploration towards balanced trade-offs.
- Similarity-based tournament selection in late iterations promotes mating of highly similar portfolios to fine-tune allocation weights, improving local exploitation.
- The proposed multi-stage mutation operator balances exploration and exploitation by dynamically favoring asset swaps (global moves) early and weight perturbations (local moves) later.
- Experimental results on S&P 100, DAX 100, Nikkei 225, and Tehran Stock Exchange data sets show faster convergence and improved Pareto fronts over baseline NSGA-II and variants.
- Repair mechanisms guarantee cardinality and boundary constraints are satisfied throughout evolution, maintaining solution feasibility without degrading diversity.
- The proposed strategy maintains performance scalability even as the number of available assets increases, preserving quality approximations and convergence speed.
Threat model
n/a (Optimization problem setting rather than adversarial threat)
Methodology — deep read
The authors target the cardinality constrained portfolio optimization problem (CCPOP), defined as a multi-objective task to minimize portfolio risk (variance) and maximize return subject to real-world constraints: cardinality constraint limiting the number of assets selected between Kmin and Kmax, boundary constraints on weights, budget normalization, no short selling, and integer inclusion variables per asset. The threat model is an optimization problem, so no adversarial assumptions are needed.
Dataset provenance includes four benchmark indices from the OR-library — S&P 100, German DAX 100, Japanese Nikkei 225, and Tehran Stock Exchange — covering portfolios of sizes up to 225 assets. Exact dataset sizes and train-test splits are not specified but the indices are standard.
The main algorithmic base is NSGA-II, a popular multi-objective evolutionary algorithm using non-dominated sorting and crowding distance for selection. The authors propose a novel chromosome structure with two fixed-length vectors of size Kmax: one encoding asset indices (with possible duplicates to enforce cardinality) and the other encoding weights associated with each asset entry. This avoids changes to string length and facilitates feasibility in cardinality constraints.
Novel mating strategies comprise two tournament selection methods: a knee-based selection active during the first α% of iterations that selects solutions closest to the utopia point in objective space to accelerate convergence; and a similarity-based selection active in final iterations mating highly similar portfolios to refine allocations. Mutation operators are also dynamically scheduled: early iterations favor swapping assets to explore broadly; later iterations favor weight changes for local fine-tuning.
The authors introduce an 'explorer' operator to boost exploration (though details are limited in the excerpt). Repair mechanisms ensure compliance with cardinality and boundary constraints by consolidating weight allocations for duplicated assets and substituting or adding assets to maintain feasibility.
Training parameters like population size, crossover and mutation rates, and stopping criteria are not explicitly detailed in the summary but typical evolutionary algorithm setups are implied. Evaluations are performed by comparing the enhanced NSGA-II variants to classical versions on well-known financial indices, assessing convergence speed and approximation quality of Pareto fronts.
One concrete example is a portfolio with Kmin = 3 and Kmax = 5 assets, where the two-row chromosome encodes three unique assets (with possible duplicates) and their corresponding weights summing to one, illustrating the representation and repair process. Though no explicit seed or hardware details are given, reproducibility would require access to their code and datasets, which are based on public financial indices but implementation code status is unclear.
Technical innovations
- Two-row fixed-length chromosome encoding separating asset indices and weights to naturally impose cardinality constraints without expensive add/remove repair operations.
- Dual tournament selection strategy combining knee-based (utopia proximity) and similarity-based (portfolio asset similarity) selection to balance exploration and exploitation adaptively over iterations.
- Dynamic mutation operator scheduling swapping assets early and fine-tuning weights late to improve search efficiency and convergence.
- Novel repair mechanisms designed for the dual-row encoding enabling fast fixes to feasibility violations in cardinality and boundary constraints.
Datasets
- S&P 100 — ~100 assets — OR-library
- German DAX 100 — ~100 assets — OR-library
- Japanese Nikkei 225 — ~225 assets — OR-library
- Tehran Stock Exchange — size unspecified — public/local market data
Baselines vs proposed
- Standard NSGA-II: convergence rate slower and Pareto front less accurate vs proposed enhanced NSGA-II with new encoding and mating strategy: faster convergence and better Pareto approximations (Fig 3 and 4 discussed).
- Classical repair mechanisms: higher overhead and slower convergence vs proposed repair using two-row encoding: faster repair and improved population diversity.
- Single tournament selection strategies: lower exploration-exploitation balance vs combined knee-based and similarity-based selection: improved convergence speed and solution quality.
Limitations
- The paper does not provide explicit hyperparameter settings like population size, crossover/mutation probabilities, or iteration counts, limiting reproducibility.
- No adversarial or robustness analysis against noisy or non-stationary financial data is included.
- Generalization to other objective functions (beyond mean-variance) or additional constraints (e.g., transaction costs) is not demonstrated.
- The experimental results focus on well-known market indices but variety in market conditions or unseen asset pools is unclear.
- Code release and exact replication details are not mentioned, raising barriers to independent verification.
Open questions / follow-ons
- How would the proposed encoding and algorithms perform under additional real-world constraints such as transaction costs, liquidity limits, and turnover constraints?
- Could the dual tournament selection strategy be generalized or adapted to other multi-objective combinatorial optimization problems beyond portfolio selection?
- What is the sensitivity of the proposed method to hyperparameters like population size, mutation rates, and cycle length for switching tournament strategies?
- How robust is the approach to market regime changes, noisy return estimates, or stress scenarios common in financial markets?
Why it matters for bot defense
While this paper does not directly address bot or CAPTCHA defense, the methodological advances in multi-objective evolutionary algorithms under complex constraints may inspire techniques in security domains that require balancing multiple objectives (e.g., detection accuracy vs user friction) under combinatorial restrictions. The novel encoding and constraint-handling strategies could be adapted to optimization challenges in CAPTCHA design or adaptive bot detection systems where resource allocation or response selection must satisfy cardinality and feasibility constraints efficiently. Additionally, the dual tournament strategy balancing exploration and exploitation over algorithm runtime might inform adaptive model training or threshold tuning in bot defense pipelines. However, direct applicability would require significant domain-specific tailoring. Practitioners in bot defense may find value in the multi-objective optimization perspectives, especially when designing challenges that must optimize security efficacy against usability constraints.
Cite
@article{arxiv2607_09566,
title={ Large-Scale Portfolio Optimization Problem Under Cardinality Constraint With Enhanced Multi-Objective Evolutionary Algorithms },
author={ Danial Ramezani and Mostafa Abouei Ardakan },
journal={arXiv preprint arXiv:2607.09566},
year={ 2026 },
url={https://arxiv.org/abs/2607.09566}
}