TLDR: Researchers propose a new diffusion-based generative model for financial time series that incorporates Geometric Brownian Motion (GBM) into its noise process. This approach, unlike standard models, injects noise proportionally to asset prices, accurately reflecting real-world financial characteristics like heteroskedasticity. Evaluated on historical stock data, the model successfully reproduces key stylized facts such as heavy-tailed returns, volatility clustering, and the leverage effect more realistically than existing generative models, offering a more robust framework for synthetic financial data generation.
Generating realistic synthetic financial time series is a crucial task in modern quantitative finance, with wide applications in risk modeling, developing trading strategies, and stress testing. As financial markets become increasingly data-driven, there’s a growing need for generative models that can produce plausible and high-fidelity data trajectories.
Traditional deep generative models, such as Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs), along with standard diffusion models, have shown promise. However, many of these approaches treat financial time series as generic numerical sequences. This often leads to a significant oversight: they fail to account for domain-specific structures inherent in financial data, such as “heteroskedasticity,” where volatility scales with asset prices. This can result in unrealistic outputs, especially during market shifts or extreme conditions.
A new research paper, titled “A diffusion-based generative model for financial time series via geometric Brownian motion,” proposes a novel solution to this challenge. Authored by Gihun Kim, Sun-Yong Choi, and Yeoneung Kim, the study introduces a groundbreaking diffusion-based generative framework that integrates Geometric Brownian Motion (GBM) directly into its forward noising process. GBM is a cornerstone of the Black–Scholes theory, a fundamental model in financial mathematics.
Unlike conventional score-based models that typically add noise uniformly, this new method injects noise proportionally to asset prices at each time step. This design inherently reflects the observed heteroskedasticity in financial time series, meaning that larger price movements naturally experience larger fluctuations. By carefully balancing the drift and diffusion terms, the model ensures that the resulting log-price process aligns with established score-based generative models, making it theoretically sound.
For the reverse-time generative process, which is responsible for creating the synthetic data, the model employs a Transformer-based architecture. This architecture is adapted from the Conditional Score-based Diffusion Imputation (CSDI) framework and has been specifically refined to better capture the complex temporal dependencies and localized volatility structures found in financial returns. The researchers fine-tuned the model’s capacity, increasing convolutional channels and embedding dimensions, which proved crucial for accurately modeling subtle features like the “leverage effect.”
The model was rigorously evaluated using historical stock data from the S&P 500 index, focusing on stocks with over 40 years of price history to ensure a diverse dataset encompassing various market regimes. The empirical evaluations demonstrated that this GBM-inspired model excels at reproducing key “stylized facts” of financial time series more realistically than conventional diffusion models and GANs.
These stylized facts include:
Heavy-tailed Return Distributions
Real-world asset returns often exhibit “heavy tails,” meaning extreme price movements occur more frequently than predicted by a normal distribution. The GBM model successfully reproduced these heavy tails, with tail exponents closely matching empirical observations, unlike other models that tended to produce lighter tails.
Volatility Clustering
This phenomenon describes how periods of high volatility tend to be followed by more high volatility, and similarly for low volatility. The GBM-based model effectively captured this persistent temporal dependence in volatility, showing a gradual and consistent decline in autocorrelation of absolute returns, mirroring real market behavior.
Also Read:
- Advancing Mammogram Generation with Precise Lesion Control
- AI-Powered Tabular Data Synthesis: Speed and Accuracy Combined
Leverage Effect
The leverage effect refers to the tendency for negative returns to be followed by increased volatility. The proposed model accurately replicated this asymmetry, exhibiting a strong negative correlation between past returns and future volatility, a critical feature for realistic risk modeling.
The study also investigated how different noise schedules (linear, exponential, cosine) interact with the GBM-based forward process, revealing that exponential or cosine scheduling combined with GBM best reproduces the characteristics of financial data. Overall, the findings indicate that by embedding fundamental financial theory into the noise structure, this model not only improves statistical fidelity but also enhances interpretability, offering a more robust foundation for generating realistic financial data.
This research represents a significant step forward in financial time series modeling, moving beyond simply simulating price dynamics to incorporating the underlying noise process itself. This novel perspective has profound implications for derivative pricing, risk measurement (such as Value-at-Risk), and the development of advanced market simulation tools.
