An Experiment on the Hurst Exponent based on FARIMA

Pu Chen, Ni Li, Jie Xue, Ting Zhao, Chen Liu2017 · DOI

Abstract

Long-Range Dependence (LRD) is a fundamental characteristic of modern network traffic, often modeled using the Fractional Autoregressive Integrated Moving Average (FARIMA) framework. However, a persistent discrepancy exists between the theoretical Hurst exponent ($H$) used to generate synthetic series and the value measured via Rescaled Range (R/S) analysis. This study investigates how noise distributions and "negative timeline" initialization methods influence this estimation gap. Experimental results indicate that lognormal noise and Autoregressive (AR) process initializations significantly improve the accuracy of the measured Hurst exponent, bringing it closer to the theoretical target.

Introduction

In network engineering, traffic patterns exhibit self-similarity and long-range dependence, where current events are statistically correlated with those in the distant past. While traditional ARMA models capture short-term fluctuations, they fail to represent this "memory." The FARIMA(p, d, q) model addresses this by incorporating a fractional difference order ($d$), where the Hurst exponent is defined as $H = d + 0.5$. Despite its theoretical robustness, generating synthetic traffic that accurately reflects a target $H$ remains challenging due to the "estimation gap" observed in finite time series.

Methodology

The researchers generated synthetic FARIMA(0, 0.25, 0) sequences, targeting a theoretical Hurst exponent of $H = 0.75$. To optimize computational efficiency, a coefficient recursion formula was used for noise sequence calculation, avoiding intensive Gamma function operations.

The experiment varied three primary parameters:

  1. Noise Distributions: Standard Normal, t-distribution (variances of 1 and 25), and Lognormal.
  2. Initialization ("Negative Timeline"): Methods to simulate history before $t=0$, including all-zero starts, normal distribution noise, and AR processes (AR(2) and AR(8)).
  3. Data Volume: Sequence lengths ranging from 5,000 to 20,000 points.

The resulting data were analyzed using Rescaled Range (R/S) analysis to determine the measured Hurst exponent.

Results and Discussion

The choice of noise distribution and initialization method proved critical to the accuracy of the Hurst exponent estimation. Across most tests, the measured $H$ was lower than the theoretical 0.75, suggesting that finite data volume and history naturally suppress measured LRD.

Hurst Exponent measurements across different noise distributions and initialization methods.

As shown in the figure, the Lognormal distribution consistently outperformed the Standard Normal distribution, which showed the highest sensitivity to data length and the largest discrepancy. Furthermore, initializing the "negative timeline" with an AR process yielded significantly better results than all-zero or random noise starts. Interestingly, increasing the complexity from AR(2) to AR(8) provided diminishing returns, suggesting that the type of initialization is more important than its mathematical depth.

The most accurate result ($H = 0.7490$) was achieved using a t-distribution with AR initialization on a 20,000-point sample. Lognormal noise remained the most stable across varying data lengths, maintaining values above 0.74.

Conclusion

Closing the gap between theoretical and measured Hurst exponents requires careful consideration of model initialization and noise characteristics. This study demonstrates that lognormal noise and AR-based history simulation are superior for generating accurate LRD time series. These findings provide a more reliable foundation for simulating network traffic and optimizing performance management in real-world environments.

Summary updated

Abstract

Based on the theory of time series analysis for the definition of FARIMA model, we generate time seriesin accordance with the FARIMA model, and measure its Hurst exponent and fractal order by using R/S analysis method. The experimental results show that the Hurst exponent used to generate the series has difference from the Hurst exponent of the real sequence measured by using R/S analysis. The further analysis indicate that, the noise types and the estimation of negative time-series data series length both have effect on the deviation.
BibTeX
@misc{chen2017experiment,
  author = {Pu Chen and Ni Li and Jie Xue and Ting Zhao and Chen Liu},
  title = {{An Experiment on the Hurst Exponent based on FARIMA}},
  year = {2017},
  doi = {10.2991/icmmct-17.2017.237},
}
Figure