[Paper Review] MarketGANs: Multivariate financial time-series data augmentation using generative adversarial networks

Jeonggyu Huh, Seungwon Jeong+3 more|arXiv (Cornell University)|Jan 25, 2026

Stock Market Forecasting Methods0 citations

TL;DR

MarketGANs propose a factor-based GAN framework with a temporal convolutional network backbone to generate high-dimensional multivariate asset returns, preserving cross-sectional dependence and time-varying risk factors for data augmentation and portfolio applications.

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.

Motivation & Objective

• Motivate data scarcity in high-dimensional asset return modeling and the need for joint distribution learning over point prediction.
• Introduce MarketGAN, a factor-based generative framework that embeds asset-pricing factors as an economic inductive bias.
• Model time-varying factor loadings and idiosyncratic volatilities via neural networks to capture cross-sectional and inter-temporal dependence.
• Demonstrate that MarketGAN-generated data improve distributional fidelity and portfolio performance under weakly informative factor information.

Proposed method

• Use a factor-model structure with r_t+1 = α_t + β_t F_t+1 + σ_t ⊙ ε_t+1 to generate returns.
• Generate α_t, β_t, σ_t as stochastic, time-varying objects via conditional GANs with a temporal convolutional network backbone.
• Anchor generated coefficients to ex-post regression estimates via multiplicative corrections: α_t = α̂_t ⊙ (1 + f_α(…)), β_t = β̂_t ⊙ (1 + f_β(…)), σ_t = σ̂_t ⊙ (1 + f_σ(…)).
• Employ a Wasserstein GAN with gradient penalty (WGAN-GP) for stable adversarial training.
• Use a TCN-based discriminator that operates on sequences to assess realism of return paths.
• Treat MarketGAN both as a data generator of joint return distributions and as a stochastic coefficient factor model.

Experimental results

Research questions

• RQ1 Can MarketGAN learn and reproduce the joint high-dimensional distribution of asset returns while preserving cross-sectional dependence and tail co-movement?
• RQ2 Do MarketGAN-generated samples improve portfolio-relevant metrics and covariance estimation relative to factor-model-based bootstrap methods, especially when factor information is weakly informative?
• RQ3 How well do MarketGAN-generated returns replicate stylized facts such as volatility clustering and leverage effects across many assets?

Key findings

• MarketGAN generates synthetic returns that closely match marginal distributions and higher-order moments.
• It reproduces intertemporal stylized facts like volatility clustering and leverage effects.
• MarketGAN substantially improves fidelity of cross-sectional dependence and tail co-movement over factor-model-based bootstrap benchmarks.
• Covariance estimates from MarketGAN samples yield superior portfolio performance when factor information is weakly informative.
• The gains arise from generating coherent high-dimensional return distributions rather than solely improving marginal predictions.