This study examines whether covariance matrices derived from Conditional WGAN-GP scenarios can serve as portfolio-relevant inputs for minimum-variance optimization, as a forward-looking alternative to backward-looking historical covariance estimation....
This study examines whether covariance matrices derived from Conditional WGAN-GP scenarios can serve as portfolio-relevant inputs for minimum-variance optimization, as a forward-looking alternative to backward-looking historical covariance estimation.
The empirical design uses a sector-balanced universe of 55 S&P 500 stocks and weekly log returns from 2011 to 2025. For each reference date, a Conditional WGAN-GP model is trained with an expanding window using only pre-reference information, generating 5,000 forward-looking 13-week scenarios conditioned on seven market-state variables. Historical, raw GAN, and variance-rescaled GAN covariance matrices are then estimated with the same Ledoit-Wolf shrinkage estimator and used in a long-only minimum-variance portfolio, evaluated over the 13-week out-of-sample window across four event-based regimes (PreCOVID2020, InflationShock2022, HighRate2023, Recent2025).
GAN-derived covariance materially changes allocation: in the Recent2025 diagnostic run, the effective number of assets falls from about 29 under historical covariance to about 4, indicating strong concentration. Across the four out-of-sample regimes, GAN-derived portfolios achieve higher endpoint returns and Sharpe ratios than the historical portfolio, with realized 13-week return improvements of 4.2 to 13.7 percentage points. Risk results are regime-dependent: GAN portfolios improve downside behavior in shock windows, while historical covariance stays more defensive in calmer or recovery windows.
Conditional WGAN-GP covariance should therefore be treated as a forward-looking allocation signal, not a standalone calibrated risk model: raw scenarios require variance calibration, and GAN-based portfolios require concentration safeguards such as weight caps and turnover control. The contribution of this thesis is a portfolio-level evaluation linking scenario generation, covariance structure, allocation, concentration, and out-of-sample performance.