Abstract
Simulating financial markets requires capturing the dual characteristics of individual stock mean-reversion and complex cross-sector correlations. This paper details a hierarchical framework developed by J.P. Morgan AI Research to generate realistic synthetic stock price time series. The approach combines multivariate Ornstein-Uhlenbeck (OU) processes for sector-level dynamics with Arbitrage Pricing Theory (APT) to model individual stock behavior. Evaluation against historical S&P 500 data demonstrates that the framework produces synthetic series with high visual and statistical fidelity, effectively replicating market return distributions and inter-sector dependencies.
Introduction
Financial time series generation is a critical task for training machine learning models and testing trading strategies. However, stock prices are not simple random walks; they exhibit mean reversion and are influenced by systematic factors. Traditional univariate models often fail to capture the "complex market of interacting stocks," treating assets in isolation. This research addresses the gap by proposing a model-based framework that generates a coherent market environment grounded in financial theory.
Methodology
The framework utilizes a two-tier hierarchical architecture to manage the complexity of the S&P 500 universe:
- Multivariate Ornstein-Uhlenbeck (OU) Process: To model the broad market (SPY) and 11 sector ETFs, the authors employ a multivariate OU process. Unlike standard Brownian motion, the OU process is mean-reverting, pulling prices toward a long-term trend. The multivariate formulation uses a covariance matrix to ensure realistic co-movements between sectors (e.g., Technology and Financials).
- Arbitrage Pricing Theory (APT): Individual stocks are generated by treating their returns as a linear combination of the 11 systematic sector factors plus idiosyncratic risk. Parameters are estimated from historical data starting April 1, 2020, using linear regression.
Results and Evaluation
The framework was evaluated using the S&P 500 universe. Qualitatively, the synthetic price traces for major sectors like Information Technology (XLK) and Financials (XLF) closely mirror the trends and volatility seen in historical data.
Quantitatively, the model's performance was validated using Kernel Density Estimation (KDE) to compare return distributions. Statistical similarity was measured via the Kolmogorov-Smirnov (KS) test. For 8 of the 12 primary entities (SPY plus 11 sectors), the synthetic distributions were statistically indistinguishable from real data (p-values > 0.05).
While the model excels in most areas, certain sectors like Consumer Staples (XLP) and Communication Services (XLC) showed lower p-values, indicating areas for future refinement in capturing specific volatility profiles. Contour plots further confirmed that the model successfully replicated the correlation structures between the overall market and specific sectors.
Conclusion
The proposed framework successfully bridges the gap between simple stochastic models and the complex reality of financial markets. By leveraging the hierarchical structure of sectors and individual stocks, it provides a robust tool for generating high-fidelity synthetic data. Future work will focus on refining parameter estimation for outlier sectors and testing the utility of this data in training deep reinforcement learning agents.
Summary updated