Earticle

초록

영어

This study evaluates whether market-state-conditioned scenarios generated by a Conditional Wasserstein GAN with gradient penalty (cWGAN-GP) can provide portfolio-relevant covariance inputs for minimum-variance optimization. Using 55 sector-balanced S&P 500 stocks with weekly returns from 2011 to 2025, we generate 5,000 forward-looking scenarios conditioned on seven macro-financial variables and convert them into Ledoit-Wolf shrinkage covariance matrices. Historical, raw GAN-derived, and variance-rescaled GAN-derived covariance inputs are compared in long-only minimum-variance portfolios across four event-based regimes: PreCOVID2020, InflationShock2022, HighRate2023, and Recent2025. Variance-rescaled GAN covariance improves realized 13-week return and Sharpe ratio in all four selected regimes. However, GAN-based portfolios become sharply concentrated, with Effective N falling from about 29 under historical covariance to about 4. The results suggest that cWGAN-GP covariance is best interpreted as a regime-dependent allocation signal, not as a standalone calibrated risk model.

목차

Abstract
1. Introduction
2. Literature Review and Theoretical Background
2.1 Covariance Estimation and Minimum-Variance Portfolios
2.2 Generative Models for Financial Scenario Generation
2.3 Regime Detection and Regime-Aware Portfolio Construction
2.4 Research Gap: From Scenario Fit to Portfolio-Level Evaluation
3. Data and Research Design
3.1 Asset Universe and Return Data
3.2 Market-State Conditioning Variables
3.3 Event-Based Expanding-Window Design
3.4 Validation Strategy and Anti-Leakage Considerations
4. Methodology
4.1 Conditional WGAN-GP Scenario Generation
4.2 Checkpoint Selection and Scenario Reconstruction
4.3 Variance Rescaling
4.4 Covariance Matrix Construction
4.5 Portfolio Optimization and Evaluation Metrics
5. Empirical Results
5.1 Scenario Calibration: Recent2025 Diagnostic
5.2 Portfolio Allocation and Concentration
5.3 Realized OOS Performance Across Regimes
5.4 Risk-Return Metrics by Regime
5.5 Downside-Risk Diagnostics
5.6 Portfolio Dynamics within the 13-Week OOS Window
6. Discussion
6.1 Interpretation of Main Findings
6.2 The Two-Axis Framing of Regime Dependence
6.3 Why Concentration Matters
6.4 Calibration and Implementation Limits
6.5 Practical Implications and Hybrid Covariance
6.6 Statistical Interpretation under Small-Sample Conditions
6.7 Limitations and Future Work
7. Conclusion
References

키워드

저자

• Kharitonova Irina [ Dong-A University, Department of Management Information Systems ]
• Sumin Lee [ Dong-A University, Department of Management Information Systems ]
• Yong-Bok Cho [ Dong-A University, Department of Management Information Systems ]

참고문헌

발행기관

간행물

표지보기 관심저널 등록

• 간행물명

  한국경영정보학회 정기 학술대회 [KMIS Conference]

• 간기

  반년간

• 수록기간

  1990~2026

• 십진분류

  KDC 325 DDC 658