Parameter Estimation for the Vasicek Process with Microstructure Noise

Technical Summary: Parameter Estimation for the Vasicek Process with Microstructure Noise

Problem Statement
High-frequency financial data, while valuable for modeling mean-reverting behaviors in variables like interest rates and volatility, are inherently contaminated by market microstructure noise. This noise causes the observed process ($Y_t$) to deviate from the underlying true Vasicek diffusion process ($X_t$). Existing literature on parameter estimation for the Vasicek process often relies on stringent assumptions, such as the noise being independent and identically distributed (i.i.d.) or strictly white noise. However, empirical evidence suggests that microstructure noise frequently exhibits serial dependence, heteroskedasticity, and non-Gaussian characteristics. The paper addresses the theoretical gap in establishing strong consistency for parameter estimators when the noise follows a more realistic $\alpha$-mixing dependence structure, rather than the simplified independence assumptions common in prior works.

Methodology
The authors propose a two-stage estimation framework combining kernel smoothing with maximum likelihood estimation (MLE):

1. Kernel Smoothing (Denoising): To mitigate the impact of microstructure noise ($\epsilon_t$), the observed data $Y_{i\Delta_n}$ are first processed using a kernel smoother. The latent process $X_{i\Delta_n}$ is estimated as a weighted average of neighboring observations:
   $$\hat{X}_{i\Delta_n} = \frac{\sum K(\frac{i\Delta_n - j\Delta_n}{h_n}) Y_{j\Delta_n}}{\sum K(\frac{i\Delta_n - k\Delta_n}{h_n})}$$
   where $K$ is a kernel function and $h_n$ is the bandwidth. This step aims to recover a cleaner signal before parameter estimation.

2. Maximum Likelihood Estimation: Using the smoothed estimates $\hat{X}_{i\Delta_n}$, the authors construct estimators for the Vasicek parameters: the mean reversion speed ($\kappa$), the long-run mean ($\alpha$), and the volatility ($\sigma$). These estimators are derived from the log-likelihood function of the Vasicek process, adapted to use the smoothed data points.

3. Theoretical Proofs: The paper establishes the strong consistency of these estimators ($\hat{\kappa} \to \kappa$, $\hat{\alpha} \to \alpha$, $\hat{\sigma}^2 \to \sigma^2$ almost surely). The proof techniques rely on:

   • Truncation: Decomposing the error terms into bounded and unbounded components to handle the unbounded nature of the Vasicek process.
   • Moment Inequalities for Mixing Sequences: Utilizing exponential inequalities tailored for $\alpha$-mixing sequences (specifically referencing results by Yang et al., 2023, 2024) to bound tail probabilities, rather than relying on inequalities for i.i.d. sequences.
   • Asymptotic Conditions: The analysis assumes a long-span high-frequency setting where the sampling interval $\Delta_n \to 0$ and the total time horizon $n\Delta_n \to \infty$, with specific constraints on the bandwidth $h_n$ relative to $\Delta_n$.

Key Results

• Theoretical Consistency: The paper proves that under $\alpha$-mixing conditions for the noise and appropriate selection of the bandwidth and sampling interval, the proposed estimators are strongly consistent. This extends previous results that were limited to independent noise settings.
• Numerical Simulations: Simulations with sample sizes ranging from $n=100$ to $n=5000$ and varying noise variances ($\sigma_\epsilon = 0.005, 0.01, 0.02$) demonstrate that:
  • As sample size increases, parameter estimates converge to true values, and standard deviations decrease.
  • As noise variance decreases, estimation accuracy improves.
  • The mean-reversion speed parameter ($\kappa$) is the most sensitive to noise variance, requiring larger sample sizes to suppress bias, while the long-term mean ($\alpha$) is the least affected.
• Empirical Analysis: Using 5-minute price data of the Shanghai Stock Exchange Composite Index (SSECI), the model was applied to real-world data.
  • Kernel smoothing effectively reduced volatility in the log-return series.
  • The estimated parameters allowed for the construction of a prediction model.
  • The model achieved a confidence interval coverage rate of 92.32%, which the authors note is "almost consistent with the theoretical probability level," indicating that the prediction model fits the data features well and performs reliably despite the presence of microstructure noise.

Significance and Claims
The paper claims to bridge a specific theoretical gap regarding the robustness of parameter estimation for the Vasicek process under complex noise environments. Its primary contribution is the relaxation of the independence assumption on microstructure noise, replacing it with an $\alpha$-mixing framework that better reflects the serial dependence found in actual financial markets.

By employing truncation techniques and moment inequalities for mixing sequences, the authors provide a rigorous proof of strong consistency where classical i.i.d.-based methods would fail. The work demonstrates that reliable parameter estimation is attainable even when noise exhibits weak dependence, provided the sampling frequency and bandwidth are properly calibrated. The empirical application to SSECI data validates the method's practical utility, showing that the proposed model fits data features well and offers stable prediction capabilities despite the presence of microstructure noise.