TL;DR
This paper provides a rigorous theoretical foundation for Approximate Maximum Likelihood Estimation (AMLE) in diffusion processes. It establishes the "missing manual" for researchers, proving exactly how many approximation terms and what sampling frequencies are needed to ensure that approximate estimates are just as statistically efficient as the (often uncomputable) true maximum likelihood.
What Problem Does This Paper Address?
Continuous-time diffusion processes—mathematical models like the Black-Scholes formula used to describe asset prices or interest rates—are a cornerstone of modern finance. Ideally, we would use Maximum Likelihood Estimation (MLE) to find the parameters of these models.
However, there is a major hurdle: we usually observe data at discrete intervals (e.g., daily or monthly), but the "transition density" (the probability of moving from one price to another over time) rarely has a closed-form formula. Without this formula, the true likelihood function is unknown, making exact MLE impossible for most models. While Yacine Aït-Sahalia previously proposed using asymptotic expansions to approximate these densities, the rigorous statistical properties—like exactly how fast these approximations converge and how many terms you actually need—remained partially explored.
Methods
The authors build upon Aït-Sahalia’s expansion method, which uses Hermite polynomials and Taylor series to approximate the transition density. The paper’s core contribution is a formal analysis of the Approximate Maximum Likelihood Estimator (AMLE) under two distinct asymptotic regimes:
- Fixed sampling interval ($\delta$), increasing terms ($J \to \infty$): Improving the approximation by adding more complexity to the formula.
- Fixed terms ($J$), shrinking sampling interval ($\delta \to 0$): Improving the approximation by observing the process more frequently (high-frequency data).
They establish the consistency and convergence rates for both scenarios and derive a first-order approximation for the Fisher information matrix, which is essential for calculating confidence intervals.
Datasets
As a theoretical statistics paper, it does not use real-world empirical datasets. Instead, the authors use extensive Monte Carlo simulations to validate their proofs. They test their methods on two classic financial models:
- The Vasicek Model: A mean-reverting model often used for interest rates.
- The Cox-Ingersoll-Ross (CIR) Model: A model that ensures interest rates stay positive (square-root process).
Key Insights and Ideas
- The $J \ge 2$ Requirement: For high-frequency data ($\delta \to 0$), the authors discovered that you need at least a two-term expansion ($J \ge 2$) to ensure the AMLE has the same asymptotic normality (efficiency) as the true MLE. A simple one-term expansion is often insufficient.
- Practical Guidance on Expansion Terms: The paper provides a "lookup table" (Table 1) for practitioners. For example, if you have monthly data ($\delta = 1/12$), a one-term expansion is perfectly adequate. However, if you only have data every nine months ($\delta = 3/4$), you might need 10 to 14 terms to maintain accuracy.
- The "Extra Bias" Term: The authors quantify the specific bias introduced by the approximation. They show that the AMLE error consists of the standard statistical error $O((n\delta)^{-1/2})$ plus an approximation error term $O(\delta^J)$ or $O(\delta^{J+1})$.
- Fisher Information Shortcut: They demonstrate that a first-order approximation of the Fisher information matrix can be obtained using only the one-term density expansion, and it remains accurate up to $O(\delta^2)$.
Key Results
- Convergence Rates: The paper proves that the AMLE is consistent. For fixed $\delta$, the convergence rate is $O_p\{\delta^{J+1} + (n\delta)^{-1/2}\}$. For shrinking $\delta$, it is $O_p\{\delta^J + (n\delta)^{-1/2}\}$.
- Asymptotic Equivalence: They define the specific condition ($n\delta^{2J+2} \to 0$) under which the AMLE is statistically indistinguishable from the true MLE.
- Simulation Success: In tests with the Vasicek and CIR models, the simulated bias and standard deviation closely matched the theoretical predictions, confirming that the AMLE performs reliably even with a small number of terms if the sampling frequency is high.
Limitations and Future Work
- Univariate Focus: The detailed convergence analysis in this paper is primarily focused on univariate (one-dimensional) diffusion processes. While the authors mention multivariate expansions exist, the rigorous error bounds for those cases are more complex.
- Parametric Assumptions: The method assumes the functional forms of the drift and diffusion are known. It does not address non-parametric or semi-parametric settings.
- Future Directions: The authors suggest that further work could involve higher-order approximations of the Fisher information matrix or extending these rigorous proofs to more complex models like stochastic volatility processes.
Summary updated