whittlehurst 0.7

Hurst exponent estimation using Whittle's method

Overview

This module provides an implementation of Whittle's likelihood method to estimate the Hurst exponent of a time series. The method fits a theoretical spectral density model to the periodogram of the time series realization. This implementation supports multiple spectral density approximations for fractional Gaussian noise (increments of fractional Brownian motion) and ARFIMA processes.

The Hurst exponent ($H$) controls the roughness, self-similarity, and long-range dependence of fBm paths:

• $H\in(0,0.5):~$ anti-persistent (mean-reverting) behavior.
• $H\in(0.5,1):~$ persistent behavior.
• $H=0.5:~ \mathrm{fBm}(H)$ is the Brownian motion.
• $H\rightarrow 0:~ \mathrm{fBm}(H)\rightarrow$ White noise.
• $H\rightarrow 1:~ \mathrm{fBm}(H)\rightarrow$ Linear trend.

Features

• Spectral density options:
  • fGn
  • arfima
  • fGn_paxson
  • fGn_truncation
  • fGn_taylor
• Flexible interface with an option for a custom spectral density callback.
• Good performance both in terms of speed and accuracy.
• Included generators for fBm and ARFIMA.

Installation

pip install whittlehurst

Usage

fBm and fGn

import numpy as np
from whittlehurst import whittle, fbm

# Original Hurst value to test with
H=0.42

# Generate an fBm realization
fBm_seq = fbm(H=H, n=10000)

# Calculate the increments (the estimator works with the fGn spectrum)
fGn_seq = np.diff(fBm_seq)

# Estimate the Hurst exponent
H_est = whittle(fGn_seq)

print(f"Original H: {H:0.04f}, estimated H: {H_est:0.04f}")

ARFIMA

import numpy as np
from whittlehurst import whittle, arfima

# Original Hurst value to test with
H=0.42

# Generate an ARFIMA(0, H - 0.5, 0) realization
arfima_seq = arfima(H=H, n=10000)

# No need to take the increments here
# Estimate the "Hurst exponent"
H_est = whittle(arfima_seq, spectrum="arfima")

print(f"Original H: {H:0.04f}, estimated H: {H_est:0.04f}")

Performance

Compared to other methods

Our Whittle-based estimator offers a compelling alternative to traditional approaches for estimating the Hurst exponent. In particular, we compare it with:

• R/S Method: Implemented in the hurst package, this method has been widely used for estimating $H$.

• Higuchi's Method: Available through the antropy package, it performs quite well especially for smaller $H$ values, but its performance drops when $H\rightarrow 1$.

• Variogram: Our variogram implementation of order $p = 1$ (madogram) accessible as from whittlehurst import variogram.

Inference times indicate per input sequence times, and were calculated as: $t = w\cdot T/k$, where $k=100000$ is the number of sequences, $w=32$ is the number of workers (processing threads), and $T$ is the total elapsed time. Single-thread performance is likely to be better, the results are mainly comparative.

The following results were calculated on $100000$ fBm realizations of length $n=1600$.

fGn spectral density approximations

The fGn spectral density calculations recommended by Shi et al. are accessible within our package:

• fGn: The default recommended spectral model. It relies on the gamma function and the Hurwitz zeta function $\zeta(s,q)=\sum_{j=0}^{\infty}(j+q)^{-s}$ from scipy. $$ f(\lambda) = , \frac{\sigma^2}{\pi}, \Gamma(2H+1), \sin(\pi H), (1-\cos(\lambda)),(2\pi)^{-(2H+1)} \left[ \zeta!\left(2H+1,, 1-\frac{\lambda}{2\pi}\right) + \zeta!\left(2H+1,, \frac{\lambda}{2\pi}\right) \right] $$
• fGn_Paxson: Uses Paxson's approximation with a configurable parameter K=50.
• fGn_truncation: Approximates the infinite series by truncation with a configurable parameter K=200.
• fGn_Taylor: Uses a Taylor series expansion to approximate the spectral density at near-zero frequency.

The following results were calculated on $100000$ fBm realizations of length $n=1600$.

References

• The initial implementation of Whittle's method was based on:
  https://github.com/JFBazille/ICode/blob/master/ICode/estimators/whittle.py

• For details on spectral density models for fractional Gaussian noise, refer to:
  Shuping Shi, Jun Yu, and Chen Zhang. Fractional gaussian noise: Spectral density and estimation methods. Journal of Time Series Analysis, 2024. https://onlinelibrary.wiley.com/doi/full/10.1111/jtsa.12750

License

This project is licensed under the MIT License (c) 2025 Bálint Csanády, aielte-research. See the LICENSE file for details.