Can the Hurst exponent be greater than one? Does it mean that the time series follows a random walk or that it's not stationary?
A hurst exponent, H, between 0 to 0.5 is said to correspond to a mean reverting process (anti-persistent), H=0.5 corresponds to Geometric Brownian Motion (Random Walk), while H >= 0.5 corresponds to a process which is trending (persistent).
The hurst exponent is limited to a value between 0 to 1, as it corresponds to a fractal dimension between 1 and 2 (D=2-H, where 0 < H < 1). I often think of it more along the lines of how much space the 'wandering' fills up (between 1 to 2 dimensions) and also how jagged or noisy the process may be (more noisy-> lower hurst, more smooth -> higher hurst).
You could calculate an H>1, but it would not have any meaning using the accepted definition and fractal dimension boundaries (fractions between integer dimensions must always be less than one).
Also, from "Estimating the Hurst Exponent," R. Racine.
"Applied to financial data such as stock prices, the Hurst Exponent can be interpreted as a measure for the trendiness: H < 0:5 high volatility, stock price is anti trended, H = 0:5, stock price behaves like a brownian process, no trend, H > 0:5 stock price has a trend."
*For more insight, any of Benoit Mandelbrot's books are easily accessible and instructive on the topic.
The Hurst value can be coded to plot values >1.
An example of how to tame Hurst values >1
Following Weron, once (2) is calculated, the Hurst exponent H will be 0.5 plus the slope of (R/S)n −E(R/S)n. However, if we calculate this modified R/S analysis in this way, results show a Hurst exponent, for some random series, with values higher than 1, which makes no sense. For this reason, we have followed a different procedure than in Ref. .
This procedure lies in adding a final step to the classical R/S analysis which consist in calculating
log Hn = log (R/S)n − log E (R/S)n + log(n)/2
where E (R/S)n is given by (2).
Then find H by linear regression on log Hn = log c + H log n.
(3) The distribution for the Hurst exponent calculated as stated previously (which we will note by R/S-AL), resembles in this case a normal one with a mean of 0.49 and a standard deviation of 0.04 (with n = 16).
Note that the distribution of the Hurst exponent calculated using standard R/S analysis cannot be approximated by a normal distribution.
We would like to remark that since formula (2) was derived for the series with underlying normal distribution, modified R/S analysis should be studied deeply to check its correctness for other series (for example, the series with Hurst exponent different from 0.5).
In every case only Whittle gives an estimation for Hurst which is always .99. Also the Periodogram estimates Hurst to be greater than 1.
Long-Range Dependence in a Changing Internet Traffic Mix
The concepts and definitions for self-similarity and long-range dependence given in section 1 assume that the time series of arrivals is second-order stationary (a.k.a. weakly stationary). Loosely speaking, this means that the variance of the time series (and more generally, its covariance structure) does not change over time, and that the mean is constant (so the time series can always be transformed into a zero-mean stochastic process by simply subtracting the mean). The obvious question this raises is whether Internet traffic is stationary. This is certainly not the case at the scales in which the time-of-day effects are important (traffic sharply drops at night), so Internet traffic is usually characterized as self-similar and long-range dependent only for those scales between a few hundred milliseconds and a few thousand seconds.
For example, the UNC link showed an increase in traffic intensity during morning hours as more and more people become active Internet users. However, it is still useful to study time series using the self-similarity and long-range dependence framework, and this is possible using methods that are robust to non-stationarities in the data (i.e., methods that first remove trends and other effects from the data to obtain a second-order stationary time series). In some cases, non-stationarities are so complex, that conventional models fail to accommodate them, which can result in estimates of the Hurst parameter greater than or equal to 1. We examine some of these cases in section 5.
The wavelet-based tools for analysis of time series are important because they have been shown to provide a better estimator (and confidence intervals) than other approaches for the Hurst parameter . These methods also are robust in the presence of certain non-stationary behavior (notably linear trends) in the time series.
SiZer provides a useful method for finding statistically significant local trends in time series and is especially useful for finding important underlying structure in time series with complex structure. SiZer is based on local linear smooths of the data, shown as curves corresponding to different window widths in the top panel of Figure 4 (a random sample of time series values is also shown using dots). These curves are very good at revealing potential local trends in the time series, and provide a visualization of structure in the time series at different scales, i.e. at different window widths used for smoothing. See  for an introduction to local linear smoothing. Two important issues are: which of these curves (i.e., scales) is the right one, and which of the many visible trends (at a variety of different scales) are statistically significant (thus representing important underlying structure) as opposed to reflecting natural variation?
Summary: This paper provide an insight into the Sizer tool which can used to detect tends in the traffic. however, I need to look further into this paper and see how actually this tool be used. there is also an indication that trends can be the part of the analysis. but i am a bit confused here.
Wavelet Analysis of Long-Range-Dependent Trafﬁc Patrice Abry and Darryl Veitch
It can be implemented very efﬁciently allowing the direct analysis of very large data sets, and is highly robust against the presence of deterministic trends, as well as allowing their detection and identification. There are always come conflicting studies regarding the nature of independence in the dataset.
Statistical Methods for Data with Long-Range Dependence by Jan Beran
A Practical Guide to Measuring the Hurst Parameter
Richard G. Clegg
This paper describes, in detail, techniques for measuring the Hurst parameter. Measurements are given on artificial data both in a raw form and corrupted in various ways to check the robustness of the tools in question. Measurements are also given on real data, both new data sets and well-studied data sets. All data and tools used are freely available for download along with simple “recipes” which any researcher can follow to replicate these measurements. 2Measuring the Hurst Parameter
While the Hurst parameter is perfectly well-defined mathematically, measuring it is problematic. The data must be measured at high lags/low frequencies where fewer readings are available. Early estimators were biased and converged only slowly as the amount of data available increased. All estimators are vulnerable to trends in the data, periodicity in the data and other sources of corruption. Many estimators assume specific functional forms for the underlying model and perform poorly if this is misspecified. The techniques in this paper are chosen for a variety of reasons. The R/S parameter, aggregated variance and periodogram are well-known techniques which have been used for some time in measurements of the Hurst parameter. The local Whittle and wavelet techniques are newer techniques which generally fare well in comparative studies. All the techniques chosen have freely available code which can be used with free software to esti- mate the Hurst parameter.
The problems with real-life data are worse than those faced when measuring artificial data. Real life data is likely to have periodicity (due to, for example, daily usage patterns), trends and perhaps quantisation effects if readings are taken to a given precision. The naive researcher taking a data set and running it through an off-the-shelf method for estimating the Hurst parameter is likely to end up with a misleading answer or possibly several different misleading answers.
Various techniques are tried to filter real-life traces in addition to making measurements purely on the raw data. These methods have been selected from the literature as commonly used by researchers in the field. Often in such cases, a high pass filter would be used to remove periodicity and trends, however, since LRD measurements are most important at low-frequency that is an obvi- ously inappropriate technique. The techniques used to pre-process data before estimating H are listed below.
• Transform to log of original data (only appropriate if data is positive). • Removal of mean and linear trend (that is, subtract the best fit line Y = at + b for constant a and b). • Removal of high order best-fit polynomial of degree ten (the degree ten was chosen after higher degrees showed evidence of over-fitting).
Long-Range Dependence: Now you see it, now you don’t! By Thomas Karagiannis CSE Dept., UC Riverside [email protected] Michalis Faloutsos CSE Dept., UC Riverside [email protected] Rudolff H. Riedi ECE Dept., Rice University [email protected]
This paper is very good in illustrating the possible difficulties that one may face in realizing the real goals of achieving LRD and SS. Some of the main focus of this paper are: Not one single estimator is good in estimating LRD. for example, Whittle although provides robust result.
All of most of the estimator wrongly judge Hurst value in case of periodicity present in the data. The results are obscure by the presence of such signal. Using interpolation one can smooth the time series that is affected by missing data, removal outliers etc. It is very important to remove the periodic component. This can be done by processing and decomposing the signal.
Authors recommend to plot the signal at various scales to reveal different characteristics. Over the last few years, the network community has started to make heavy use of novel concepts such as self- similarity and Long-Range Dependence (LRD). Despite their wide use, there is still much confusion regarding the identifi- cation of such phenomena in real network traffic data.
In this paper, we show that estimating Long-Range Dependence is not straightforward: there is no systematic or definitive methodol- ogy. There exist several estimating methodologies, but they can give misleading and conflicting estimates. More specifically, we arrive at several conclusions that could provide guidelines for a systematic approach to LRD. First, long-range dependence may exist even, if the estimators have different estimates of the Hurst exponent in the interval 0.5-1. Second, long-range dependence is unlikely to exist, if there are several estimators that fail to es- timate the Hurst exponent. Third, we show that periodicity can obscure the analysis of a signal giving partial evidence of long- range dependence. Fourth, the Whittle estimator is the most accurate in finding the exact value when LRD exists, but it can be fooled easily by periodicity. As a case-study, we analyze real round-trip time data. We find and remove a periodic component from the signal, before we can identify long-range dependence in the remaining signal.
To extract the useful information from the raw RTT data, we applied typical time series methodologies like, interpola- tion to recover from loss (so that our signal would not have discontinuities), removal of outliers and smoothing. Ap- plying the estimators in the RTT signal, resulted in non- consistent estimations, in the sense that some of the estima- tors showed long-range dependence for some of our datasets.
The evaluation of each estimator is achieved through three different Fractional Gaussian Noise (FGN) generators. FGN generators are often used to synthesize long-range depen- dence series with a specific Hurst value. The first is devel- oped by Paxson , while the second is described in . The third is based in the Durbin-Levinson coefficients. Due to space limitation, we only present results from the genera- tor developed by Paxson. However, findings are similar for the other two generators. We show that the estimators are quite sensitive and can be deceived to report LRD. In particular we apply the esti- mators in synthesized signals such as cosine functions with noise or signals that show trend.
The definition of LRD assumes stationary sig- nals. In this case, we intend to identify the impact of non-stationarity on the estimators. Thus, we created various signals with slow and fast decaying or increas- ing trends. Such signals include combination of White Gaussian Noise and cosine functions with trend. In every case only Whittle gives an estimation for Hurst which is always .99. Also the Periodogram estimates Hurst to be greater than 1.
A reporting of the Hurst exponent is meaningful, only if it is accompanied by the method that was used, as well as the confidence intervals or correlation coefficient. •Researchers should not rely only on one estimator in deciding the existence of long-range dependence (e.g. ). As we saw, several of the estimators (Whittle, Pe- riodogram) can be overly optimistic in identifying long- range dependence. •For efficient characterization, it may be necessary to process and decompose the signal. •A visual inspection of the signal can be very useful, pro- viding a qualitative analysis and revealing many of its features, like periodicity. We recommend plotting the signal at several different scales, since each scale can reveal different characteristics.