MarketGANs: Multivariate financial time-series data augmentation using generative adversarial networks

TL;DR

MarketGAN addresses the challenge of learning high-dimensional asset return distributions under limited data by embedding a factor-model structure as an economic inductive bias and using a temporal convolutional network backbone to generate time-varying factor loadings and idiosyncratic risks. It generates returns as a single joint vector via a conditional GAN, enabling accurate replication of cross-sectional dependence and tail co-movement beyond traditional factor-model bootstraps. Empirically on 98 S&P 100 stocks, MarketGAN closely matches marginal distributions and inter-temporal stylized facts, while substantially improving cross-sectional dependence; covariance estimates derived from MarketGAN samples translate into superior portfolio performance when factor information is informative, illustrating practical value. The work highlights a general framework for decision-making under data scarcity in high dimensions, combining economic structure with flexible generative modeling to produce coherent joint distributions for downstream optimization and risk assessment.

Abstract

This paper introduces MarketGAN, a factor-based generative framework for high-dimensional asset return generation under severe data scarcity. We embed an explicit asset-pricing factor structure as an economic inductive bias and generate returns as a single joint vector, thereby preserving cross-sectional dependence and tail co-movement alongside inter-temporal dynamics. MarketGAN employs generative adversarial learning with a temporal convolutional network (TCN) backbone, which models stochastic, time-varying factor loadings and volatilities and captures long-range temporal dependence. Using daily returns of large U.S. equities, we find that MarketGAN more closely matches empirical stylized facts of asset returns, including heavy-tailed marginal distributions, volatility clustering, leverage effects, and, most notably, high-dimensional cross-sectional correlation structures and tail co-movement across assets, than conventional factor-model-based bootstrap approaches. In portfolio applications, covariance estimates derived from MarketGAN-generated samples outperform those derived from other methods when factor information is at least weakly informative, demonstrating tangible economic value.

Figures (16)

• Figure 1: Architecture of MarketGAN.
• Figure 2: Marginal distributions of daily excess returns. FF-1, FF-3, and FF-5 denote the respective factor-model-based bootstrap methods.
• Figure 3: Boxplots of the first to fourth moments (mean, variance, skewness, kurtosis) of asset returns generated by MarketGAN and factor-model-based bootstrap methods. Each boxplot summarizes the 10th, 25th, 50th, 75th, and 90th percentiles across assets.
• Figure 4: ACF, VC, and Lev of real and synthetic equal-weighted excess return samples generated by the MarketGAN5 and FF-5 bootstrap.
• Figure 5: Cross-correlation of real and synthetic excess return samples generated by MarketGANs and factor-model-based bootstrap methods.