Rachev-Ratio Portfolio Optimization with TimeGAN Scenarios: Replacement versus Supplementation

Abstract

This study evaluates the reliability and investment consequences of scenarios from a specified time-series generative adversarial network (TimeGAN) in Rachev-ratio portfolio allocation. Using 12 exchange-traded funds, we compare historical, replacement and hybrid policies across five training seeds with delayed execution and transaction costs. Synthetic probability mass varies independently of draw count under a fixed training protocol; optimal generator fidelity is not assumed. Generated scenarios understate volatility and tail losses and distort dependence. Limited longer-training diagnostics improve reconstruction loss without improving cross-asset correlation accuracy. Higher-mass hybrids earn larger average net returns in some configurations, but all conditional generated-versus-historical return intervals include zero at the reference block length. Full-period TimeGAN means exceed resampling controls, while direct paired return, Sharpe and realized-tail-ratio intervals include zero. Replacement raises turnover and is sensitive to the modified objective's denominator floor; changing that floor does not uniformly improve realized downside risk. Exploratory generation-error associations do not identify a causal investment-performance channel. The results illustrate financial model risk from the interaction of imperfect scenarios and a tail-sensitive allocation criterion, supporting joint evaluation of fidelity, numerical behavior and realized outcomes within the tested implementation.