Beyond Visual Realism: Toward Reliable Financial Time Series Generation

TL;DR

The paper addresses the gap between superficial realism and practical usability in financial time series generation by focusing on stylized facts such as fat tails, volatility clustering, and leverage effects. It introduces Stylized Facts Alignment GAN (SFAG), which integrates differentiable structure-preserving losses $L_{ ext{GPD}}$, $L_{ ext{ACF}}$, $L_{ ext{Lev}}$, and $L_{ ext{CFVC}}$ into the GAN objective, forming $L_{ ext{SFAG}} = L_{ ext{adv}} + \lambda_1 L_{ ext{GPD}} + \lambda_2 L_{ ext{ACF}} + \lambda_3 L_{ ext{Lev}} + \lambda_4 L_{ ext{CFVC}}$ and trained under WGAN-GP with gradient penalty $\ ext{gp}=10$. Through experiments on the Shanghai Composite Index (2004–2024), SFAG achieves closer fidelity to key stylized facts and yields robust momentum backtests, outperforming Standard GAN and WGAN-GP in both structural accuracy and trading viability. The work demonstrates that enforcing structure-preserving objectives during training bridges the gap between visual realism and financial usability, and it suggests backbone-agnostic applicability to diffusion or transformer-based generators. This approach offers a principled path toward reliable, risk-aware synthetic financial data for stress testing, scenario analysis, and privacy-preserving learning.

Abstract

Generative models for financial time series often create data that look realistic and even reproduce stylized facts such as fat tails or volatility clustering. However, these apparent successes break down under trading backtests: models like GANs or WGAN-GP frequently collapse, yielding extreme and unrealistic results that make the synthetic data unusable in practice. We identify the root cause in the neglect of financial asymmetry and rare tail events, which strongly affect market risk but are often overlooked by objectives focusing on distribution matching. To address this, we introduce the Stylized Facts Alignment GAN (SFAG), which converts key stylized facts into differentiable structural constraints and jointly optimizes them with adversarial loss. This multi-constraint design ensures that generated series remain aligned with market dynamics not only in plots but also in backtesting. Experiments on the Shanghai Composite Index (2004--2024) show that while baseline GANs produce unstable and implausible trading outcomes, SFAG generates synthetic data that preserve stylized facts and support robust momentum strategy performance. Our results highlight that structure-preserving objectives are essential to bridge the gap between superficial realism and practical usability in financial generative modeling.

Figures (1)

• Figure 1: Visual comparison of six stylized facts between real and generated data. SFAG successfully preserves key properties such as unpredictability, volatility clustering, fat tails, leverage effects, coarse-fine correlations, and gain/loss asymmetry.