📊 statistics

Exact Differentiable Inference for Fractional Stochastic Volatility: ADavies–Harte-Reparameterized Hamiltonian Monte Carlo Posterior of the Hurst Index, with Application to the VN30 Frontier Index

This paper presents an exact, differentiable Bayesian inference method for the Hurst index in fractional stochastic volatility models by integrating the Davies–Harte circulant-embedding algorithm into Hamiltonian Monte Carlo, thereby achieving $O(N \log N)$ computational efficiency and unbiased posterior estimates that refute the rough volatility hypothesis in favor of long-memory dynamics for the VN30 index, unlike biased Whittle approximations.

Original authors: Thanh-Phong Lam, Viet-Tam Tran

Published 2026-07-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Thanh-Phong Lam, Viet-Tam Tran

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict how "bumpy" the road of a stock market will be tomorrow. In finance, this bumpiness is called volatility. For a long time, experts believed these bumps were either "rough" (chaotic and unpredictable, like a rocky mountain path) or "smooth" (predictable, like a highway).

This paper is about a new, super-precise way to measure that bumpiness, specifically for the VN30, which is the top 30 stocks in the Vietnamese market. Here is the story of what the authors did, explained simply:

1. The Problem: The "Heavy Lifting" Mistake

To measure how bumpy the road is, mathematicians use a special number called the Hurst Index (let's call it H).

If H is less than 0.5, the road is "rough" (chaotic).
If H is greater than 0.5, the road has "long memory" (it remembers its past bumps and tends to keep going in the same direction).

The problem is that calculating H exactly is like trying to solve a giant jigsaw puzzle where every single piece is connected to every other piece. Doing this the old-fashioned way is so slow and heavy (like trying to lift a truck with your bare hands) that most people take a shortcut. They use a "spectral approximation" (let's call it the Whittle shortcut).

The Analogy: Imagine you are trying to weigh a giant elephant. The exact way requires a massive, expensive scale that takes hours to calibrate. The shortcut is to guess the weight based on how the elephant looks from a distance. It's fast, but it often gets the weight wrong. The authors found that this shortcut was consistently underestimating how "sticky" or "memory-filled" the VN30 market actually is.

2. The Solution: A "Magic Elevator" for Math

The authors built a new machine to do the exact calculation without the heavy lifting. They combined two powerful tools:

Davies-Harte Embedding: Think of this as a special recipe that turns a messy, tangled knot of data into a neat, straight line using a "Fast Fourier Transform" (a mathematical magic trick that sorts things quickly).
Hamiltonian Monte Carlo (HMC): This is like a smart robot that explores the landscape of possibilities. Instead of stumbling around in the dark, the robot uses "gradients" (like feeling the slope of a hill) to slide directly to the most likely answer.

The Result: They managed to make the "heavy lifting" (exact calculation) run as fast as the "shortcut." They turned a task that used to take cubic time (imagine a cube growing huge) into a task that takes almost linear time (growing in a straight line). It's like replacing a slow, winding dirt road with a high-speed maglev train that still goes to the exact same destination.

3. The Discovery: The VN30 is Not "Rough"

When they applied this new, precise machine to the VN30 data (daily stock returns from 2012 to 2026), they found something surprising:

The Shortcut's Guess: The old method said the market was "rough" (H < 0.5), meaning it was chaotic and forgetful.
The Exact Truth: Their new method found the Hurst Index is 0.58.

What does 0.58 mean? It means the VN30 market has long memory. It's not a chaotic rock slide; it's more like a river that remembers where it's been. If the market goes up, it has a tendency to keep going up for a while. The "roughness" people thought they saw was actually an illusion caused by the inaccurate shortcut.

4. The Proof: A "Taste Test"

To prove their new machine wasn't broken, they ran a simulation (a "taste test"):

They created fake data where they knew the answer was "long memory."
They ran the old shortcut: It got confused and said the answer was "rough."
They ran their new exact method: It got the answer right almost every time.

This confirmed that the old shortcut was biased and that their new method was telling the truth.

5. The Real-World Payoff: Better Forecasts

Finally, they tested if this "long memory" knowledge helped predict the future.

They built a forecast that looked at the past (using the long-memory idea).
They compared it to a standard "Random Walk" guess (which assumes the market is totally random).
The Result: Their long-memory forecast was 18.7% more accurate than the random guess. Even a very popular, standard model (GARCH) only beat them by a tiny margin (2.5%), and that standard model was essentially trying to fake long memory by pretending to be "almost integrated."

Summary

The authors built a fast, exact math tool that avoids the errors of older shortcuts. When they used it on the Vietnamese stock market (VN30), they discovered that the market is not chaotic and rough, but rather has a strong memory. This finding is not just a number; it means that understanding the market's past behavior is actually very useful for predicting its future, and the old ways of measuring it were misleading us.

Technical Summary: Exact Differentiable Inference for Fractional Stochastic Volatility

Problem Statement
The paper addresses the computational and inferential challenges associated with estimating the Hurst index ( $H$ ) in fractional stochastic volatility models. In these models, log-volatility is driven by a fractional Brownian motion (fBm), where $H < 1/2$ implies "rough" volatility and $H > 1/2$ implies long-memory volatility. While full-posterior Bayesian inference is desirable, it is traditionally bottlenecked by the non-Markovian nature of fBm, which requires inverting a dense $N \times N$ covariance matrix. This operation incurs an $O(N^3)$ computational cost, forcing practitioners to rely on biased spectral approximations (such as the Whittle pseudo-likelihood) or frequentist point estimators. These approximations can introduce significant bias, particularly in distinguishing between rough and long-memory regimes.

Methodology
The authors propose a framework that renders the exact likelihood of the Hurst index both differentiable and computationally efficient ( $O(N \log N)$ ). The methodology consists of three primary components:

Model Specification: The study utilizes a stochastic volatility SDE where the log-volatility follows a fractional Ornstein-Uhlenbeck process driven by fBm. To handle frontier market data (specifically the VN30 index), the observation model incorporates a liquidity-dependent mixture distribution. This accounts for "stale-price" artifacts (zero returns due to non-trading) by modeling them as a point mass at zero, distinct from genuine diffusive Gaussian returns.
Exact Marginal Posterior (Scalar $H$ ): For the scalar Hurst index, the authors derive an exact marginal likelihood by analytically integrating out the volatility-of-volatility parameter ( $\sigma$ ) under a Jeffreys prior. This results in a sampling-free posterior for $H$ that is computed via deterministic quadrature, eliminating Monte Carlo error.
Differentiable Davies–Harte Generator (Full Posterior): To enable full-posterior inference over the parameter vector $(\mu, \kappa, \sigma, H)$ $(μ, κ, σ, H)$ and the latent path, the authors embed the Davies–Harte circulant embedding method within an automatic-differentiation Hamiltonian Monte Carlo (HMC) sampler (specifically the No-U-Turn Sampler).
- The fBm path is reparameterized as a deterministic, FFT-differentiable function of whitened coordinates and $H$ .
- This allows gradients of the log-posterior to flow through the Fast Fourier Transform (FFT), enabling geometry-aware sampling without ever forming or inverting the dense covariance matrix.
- The computational cost per step is reduced from $O(N^3)$ to $O(N \log N)$ .

Key Contributions

Exact Differentiable Inference: The paper introduces the first scheme that achieves full-posterior Bayesian inference for fractional stochastic volatility at near-linear cost ( $O(N \log N)$ ) while maintaining exactness, avoiding the biases inherent in spectral approximations.
Analytic Marginalization: It provides an exact, sampling-free posterior for the scalar Hurst index by analytically marginalizing the scale parameter, allowing for precise measurement of the inferential gap between exact and approximate methods.
Dual-Verified Implementation: The methodology is implemented in Python and independently verified in R. The two tracks agree to within $4.9 \times 10^{-11}$ , ensuring the reproducibility of all reported numerical results.
Theoretical Validation: The paper includes a well-posedness theorem proving the existence and uniqueness of strong solutions for the model under specific conditions for $H$ (interpreting the integral in the Young sense for $H > 1/2$ and rough-path sense for $H \in (1/4, 1/2)$ ).

Results
The methodology was applied to the VN30 frontier index using $N=3585$ daily returns from 2012 to 2026.

Hurst Index Estimation: The exact posterior places the Hurst index at a mean of 0.5828 with a 95% credible interval of [0.5614, 0.6052]. The entire interval lies strictly above $0.5$, indicating long-memory volatility rather than roughness.
Comparison with Whittle Approximation: The standard Whittle spectral estimator yielded a MAP estimate of $H=0.56$ , shifting the inference toward the rough regime. Simulation calibrations confirmed that the Whittle estimator is biased toward roughness and suffers from severe under-coverage (12.5% coverage at true $H=0.70$ ), whereas the exact posterior remains near-unbiased with nominal coverage.
Model Comparison: The exact posterior odds favor the long-memory hypothesis ( $H > 1/2$ ) over the rough hypothesis ( $H < 1/2$ ) by a Bayes factor of $1.5 \times 10^{20}$ , effectively refuting the rough regime for this dataset.
Forecasting Performance: Out-of-sample forecasting on a 252-day holdout showed that a horizon-aware forecaster (Heterogeneous Autoregressive model) reduced one-day random-walk error by 18.7%. A tuned GARCH(1,1) model achieved a further 2.5% reduction, with its near-integrated persistence ( $\alpha + \beta = 0.96$ ) serving as a short-memory approximation of the long-memory dynamics identified by the Hurst posterior.
Computational Efficiency: Timing benchmarks confirmed the theoretical complexity claims. At $N=3585$ , the Davies–Harte FFT generator was approximately 1,100 times faster than the dense Cholesky factorization, with an empirical scaling exponent of 0.89 (near-linear) versus 2.64 (near-cubic) for the dense method.

Significance and Claims
The paper claims that the observed "roughness" in equity volatility may be an estimation artifact or a regime-specific feature, as the exact inference for the VN30 frontier index decisively supports a long-memory regime. The significance of the work lies in:

Removing Bias: Providing a bias-free alternative to the widely used Whittle approximation, which the authors demonstrate can qualitatively change the regime classification (from long-memory to rough).
Scalability: Making exact Bayesian inference for non-Markovian volatility models computationally feasible for large datasets, a task previously restricted to approximate methods.
Empirical Refutation: Offering a rigorous, data-driven refutation of the rough volatility hypothesis for the VN30 index, supported by a Bayes factor that effectively rules out the rough regime rather than merely disfavoring it.
Reproducibility: Establishing a high standard of reproducibility through dual-track implementation and precise numerical verification.

The authors conclude that while the rough regime is a valid feature for some developed markets, the exact differentiable inference scheme is necessary to distinguish genuine long-memory dynamics from estimation artifacts, particularly in frontier markets where liquidity constraints and data characteristics differ.