Exact and Approximate Maximum Likelihood Estimation for Multivariate Diffusions under High-Frequency Sampling
This paper establishes the asymptotic theory for exact and approximate maximum likelihood estimators of multivariate ergodic diffusions under high-frequency, long-span sampling by proving that a sufficiently high-order closed-form density expansion uniformly approximates the exact likelihood, thereby solving a longstanding open problem and characterizing the consistency, efficiency, and explicit bias properties of these estimators.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written by the author. For technical accuracy, refer to the original paper. Read full disclaimer
In the world of financial markets, prices do not move in straight lines; they jitter, drift, and surge in complex, continuous patterns. To understand these movements, economists and mathematicians use a tool called a diffusion model, which treats price changes as a fluid flow influenced by two main forces: a steady drift that pulls the price in a general direction, and a random, jittery volatility that pushes it around unpredictably. These models are essential for pricing complex financial contracts, managing risk, and understanding how markets behave over time. However, there is a significant hurdle in using them: while the rules that govern the movement are known, the exact mathematical description of how the price moves from one moment to the next is often impossible to write down in a simple formula. Because the true formula is missing, researchers have long relied on clever approximations to estimate the model's hidden parameters, such as the strength of the drift or the size of the volatility. The problem with these approximations is that no one could be certain how close they were to the truth, especially when the data comes in at very high speeds, with thousands of observations recorded in a single day.
A new study by Duy Nguyen addresses this uncertainty by finally establishing the mathematical behavior of the "exact" method, the one that would be used if the missing formula were known. The research focuses on a specific regime where data is collected frequently over a long period, a scenario common in modern high-frequency trading. The author proves that even without the exact formula, the best possible estimator—the one that maximizes the likelihood of the observed data—behaves in a predictable and optimal way. It consistently finds the true values of the drift and volatility, and as the amount of data grows, the estimates settle into a reliable pattern that allows for precise statistical testing. This result solves a long-standing open question in the field, confirming that the ideal method works as well as theory predicts, even when the underlying math is too complex to write down explicitly.
To reach this conclusion, the researcher did not try to solve the impossible equation directly. Instead, they compared the unknown exact method against a high-precision approximation known as a delta expansion. This approximation builds a detailed picture of the price movement by adding layers of correction to a simple baseline, much like refining a rough sketch into a detailed portrait. The study demonstrates that if enough layers of correction are added, the approximation becomes so accurate that it is indistinguishable from the exact method for all practical purposes. The difference between the two is so small that it vanishes as the data set grows, meaning the approximate method inherits all the desirable properties of the exact one. This finding provides a rigorous justification for using these approximations in real-world applications, assuring practitioners that their results are statistically sound.
The paper also investigates what happens when the approximation is not perfect, specifically when the number of correction layers is kept low to save computational effort. The analysis reveals a specific, predictable bias that creeps into the estimates. This bias does not affect all parameters equally; it follows a strict pattern based on the number of layers used. If the number of layers is even, the error primarily distorts the estimate of the drift. If the number of layers is odd, the error shifts to the volatility estimate. This alternating pattern was not visible in older methods and provides a clear guide for researchers: to get the most accurate results, one must choose the number of layers carefully to avoid biasing the specific parameter of interest. The study confirms these theoretical predictions through computer simulations, showing that the size and direction of the errors match the mathematical formulas exactly.
Furthermore, the research extends these findings to a broader class of models where the volatility is not constant but grows as the price level rises, a feature common in many real-world financial assets. In these cases, the standard mathematical tools often fail because the volatility can become unbounded. The author shows that by rescaling the problem to look at the movement relative to the current price level, the same high-precision approximations work just as well. This means the theory holds up even for models with heavy-tailed distributions, where extreme price moves are more likely than in standard models. The study concludes that for a wide range of financial models, the exact maximum likelihood estimator is a robust and efficient tool, provided the data is sampled frequently enough. It also clarifies the limits of approximation, showing exactly how many correction terms are needed to ensure the results are not skewed by systematic error.
The work does not claim to solve every problem in the field. It explicitly leaves aside models where the volatility can drop to zero, such as those used for interest rates, because the mathematical behavior near zero is fundamentally different and requires new tools. It also notes that if the data is sampled too slowly relative to the speed of the market, the approximations break down, and the bias becomes too large to ignore. However, within the realm of high-frequency data and models with non-zero volatility, the study provides a definitive answer. It confirms that the exact method is consistent and efficient, and it gives a precise recipe for how to use approximations without losing accuracy. By bridging the gap between the theoretical ideal and the practical approximation, the research offers a solid foundation for the statistical analysis of complex financial systems.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.