Abstract
We address the problem of parameter estimation for diffusion-driven stochastic volatility models through Markov chain Monte Carlo (MCMC). To avoid degeneracy issues, we introduce an innovative reparametrisation defined through transformations that operate on the time scale of the diffusion. A novel MCMC scheme which overcomes the inherent difficulties of time change transformations is also presented. The algorithm is fast to implement and applies to models with stochastic volatility. The methodology is tested through simulation-based experiments and illustrated on data consisting of US treasury bill rates.
Introduction
Stochastic Volatility (SV) models are fundamental tools in financial engineering, essential for capturing the complex dynamics of asset prices and interest rates. However, estimating the parameters of these models is computationally challenging. Because these processes are typically observed only at discrete intervals, researchers often employ Bayesian data augmentation. This approach involves using Markov chain Monte Carlo (MCMC) to "fill in" the missing paths between observations. While theoretically sound, this method faces a significant mathematical hurdle: as the discretization of the latent path becomes finer to improve accuracy, the MCMC algorithm often fails to converge.
The Challenge of Degeneracy
The primary obstacle in standard MCMC approaches for SV models is the property of quadratic variation. In a diffusion process, if the path is known perfectly over any interval, the volatility parameter is also known perfectly. In a Bayesian MCMC context, this creates a near-perfect correlation between the latent path and the model parameters. As the number of imputed points in the data augmentation increases, the algorithm becomes increasingly "frozen," unable to update the parameters without simultaneously updating the entire path in a way that is consistent with the new parameters. This leads to a total lack of convergence, a phenomenon known as reducibility.
Methodology: Decoupling via Time Change
To overcome this degeneracy, we propose a novel reparametrization based on time change transformations. Unlike traditional state-space transformations that rescale the values of the diffusion, our approach transforms the underlying time axis. This method effectively decouples the latent paths from the model parameters through a two-step process:
- Time Stretching: The original process $X$ is mapped to a new process $U$ on a time scale determined by the volatility parameter (e.g., $\sigma^2 t$).
- Standardization: The process $U$ is further transformed into a process $Z$, which behaves as a Brownian bridge.
The critical advantage of this transformation is that the law of $Z$ follows a standard Wiener measure, which is independent of the model parameters. By performing MCMC updates on the standardized process $Z$ rather than the original path $X$, the parameters and the path are decoupled, allowing the algorithm to move freely through the parameter space.
To implement this in practice, we utilize retrospective sampling. Because changing a parameter like $\sigma$ effectively alters the time grid of the observations, we developed a scheme to dynamically "fill in" new points using Brownian bridge properties. This ensures that the likelihood remains consistent across iterations despite the shifting time scales.
Empirical Evaluation
We evaluated the effectiveness of the time change transformation using both simulated and real-world datasets.
Simulation Studies
We tested the methodology on a 2D stochastic volatility model incorporating "leverage effects," where price shocks and volatility shocks are correlated. We simulated 501 observations to determine if the MCMC could accurately retrieve the true parameters. The results demonstrated high efficiency; the autocorrelation of the parameter draws remained low even as the number of imputed points increased from 30 to 50. This confirms that the algorithm is nearly invariant to the level of augmentation, successfully bypassing the curse of discretization. For instance, the model recovered a true $\sigma$ of 0.4 with a posterior mean of 0.406.
US Treasury Bill Rates
The methodology was also applied to a dataset of 1,809 weekly observations of the 3-month US Treasury bill rate, spanning from January 1962 to August 1996. The analysis confirmed the presence of significant stochastic volatility ($\sigma \approx 2.76$). Interestingly, the model found little evidence for mean reversion in interest rates during this specific thirty-year period, providing insights into the underlying dynamics of the US economy during that era.
Conclusion and Future Work
The time change reparametrization offers a robust solution to the convergence issues that have long plagued Bayesian inference for stochastic volatility models. By standardizing the "speed" of the process through time-stretching, we achieve a scheme that is flexible enough to handle leverage effects, state-dependent volatility, and multivariate models.
While the current MCMC scheme is highly effective, it is non-trivial to implement, requiring the management of multiple time scales and retrospective sampling. Future research will focus on extending this framework to jump-diffusion models, where prices can exhibit instantaneous snaps, and conducting comparative studies between time-stretching and state-rescaling strategies for complex multidimensional models.
Summary updated