On the approximate maximum likelihood estimation for diffusion processes

Jinyuan Chang, Song Xi Chen

Abstract

The transition density of a diffusion process generally lacks an explicit expression, which complicates full maximum likelihood estimation (MLE) for discretely observed sample paths. Building on the asymptotic expansion framework proposed by Aït-Sahalia, this paper establishes the consistency and convergence rates of the Approximate Maximum Likelihood Estimator (AMLE). Our analysis elucidates the relationship between the number of expansion terms, the sampling interval, and the resulting statistical efficiency. We identify specific conditions under which the AMLE achieves the same asymptotic distribution as the full MLE and propose a first-order approximation for the Fisher information matrix.

Introduction

Continuous-time diffusion processes are fundamental to modern financial modeling, serving as the backbone for describing asset prices, interest rates, and volatility. To estimate the parameters of these models, Maximum Likelihood Estimation (MLE) is the preferred statistical approach due to its efficiency. However, a significant practical hurdle exists: while models are defined in continuous time, data is observed at discrete intervals. For most diffusion processes, the transition density—the probability of moving from one state to another over a specific time step—does not have a closed-form solution. This absence of a likelihood function renders exact MLE impossible for the vast majority of models used in practice.

The AMLE Framework

To bridge this gap, researchers utilize the Approximate Maximum Likelihood Estimator (AMLE). This method relies on asymptotic expansions of the transition density, typically using Hermite polynomials and Taylor series to approximate the unknown likelihood. While previous work by Aït-Sahalia introduced these expansions, this research provides the rigorous theoretical "manual" for their application, focusing on two distinct asymptotic regimes:

  1. Fixed Sampling Interval ($\delta$): Where the approximation is improved by increasing the number of terms ($J \to \infty$) in the density expansion.
  2. Shrinking Sampling Interval ($\delta \to 0$): Where the approximation is improved by the increasing frequency of data (high-frequency asymptotics).

Theoretical Contributions and Convergence Rates

The core of this work is the formal proof of consistency and the derivation of convergence rates for the AMLE. We demonstrate that the error in the AMLE is composed of two parts: the standard statistical error inherent in any estimation and an approximation error introduced by the expansion.

For a fixed sampling interval $\delta$, the convergence rate is established as $O_p\{\delta^{J+1} + (n\delta)^{-1/2}\}$. In the high-frequency setting where $\delta$ shrinks, the rate is $O_p\{\delta^J + (n\delta)^{-1/2}\}$. A critical finding of this research is the condition for asymptotic equivalence: when $n\delta^{2J+2} \to 0$, the AMLE becomes statistically indistinguishable from the (uncomputable) true MLE. This provides a clear target for researchers to ensure their approximations do not sacrifice statistical power.

Key Insights for Practitioners

The $J \ge 2$ Requirement

One of the most significant discoveries for high-frequency data analysis is that a simple one-term expansion is often insufficient. To ensure the AMLE maintains the same asymptotic normality and efficiency as the true MLE when $\delta \to 0$, at least a two-term expansion ($J \ge 2$) is required.

Balancing Terms and Frequency

The paper provides practical guidance on the trade-off between sampling frequency and expansion complexity. For example, with high-frequency monthly data ($\delta = 1/12$), a single-term expansion is often adequate for high precision. Conversely, if observations are sparse (e.g., every nine months, $\delta = 3/4$), practitioners may need between 10 and 14 terms to maintain the integrity of the estimation.

Fisher Information Shortcut

Calculating the Fisher information matrix is essential for determining confidence intervals, but it can be computationally intensive. We demonstrate that a first-order approximation of this matrix can be derived using only a one-term density expansion. This approximation remains accurate up to $O(\delta^2)$, offering a significant computational shortcut without a meaningful loss in precision for many applications.

Simulation and Validation

To validate these theoretical bounds, the study employed Monte Carlo simulations using two benchmark financial models: the Vasicek Model (mean-reverting) and the Cox-Ingersoll-Ross (CIR) Model (square-root process). The simulations confirmed that the observed bias and standard deviation in the AMLE align closely with the predicted theoretical rates. Even with a limited number of expansion terms, the AMLE performs reliably provided the sampling frequency is sufficiently high relative to the complexity of the model.

Conclusion

By establishing the rigorous statistical properties of the AMLE, this research transforms a useful heuristic into a mathematically grounded tool. It provides the necessary bounds to ensure that approximate methods do not introduce hidden biases, allowing researchers to apply maximum likelihood principles to a much broader class of diffusion processes with confidence.

Summary updated

Abstract

The transition density of a diffusion process does not admit an explicit expression in general, which prevents the full maximum likelihood estimation (MLE) based on discretely observed sample paths. A\"ıt-Sahalia [J. Finance 54 (1999) 1361--1395; Econometrica 70 (2002) 223--262] proposed asymptotic expansions to the transition densities of diffusion processes, which lead to an approximate maximum likelihood estimation (AMLE) for parameters. Built on A\"ıt-Sahalia's [Econometrica 70 (2002) 223--262; Ann. Statist. 36 (2008) 906--937] proposal and analysis on the AMLE, we establish the consistency and convergence rate of the AMLE, which reveal the roles played by the number of terms used in the asymptotic density expansions and the sampling interval between successive observations. We find conditions under which the AMLE has the same asymptotic distribution as that of the full MLE. A first order approximation to the Fisher information matrix is proposed.
BibTeX
@article{chang2011approximate,
  author = {Jinyuan Chang and Song Xi Chen},
  title = {{On the approximate maximum likelihood estimation for diffusion processes}},
  journal = {The Annals of Statistics},
  year = {2011},
  doi = {10.1214/11-aos922},
}
Figure