Fractals and Stochastic Calculus

Just another weblog

R/s Analysis to estimate the Hurst exponent

R/s analysis (or Rescaled Range analysis) was originally devised by Harold Edwin Hurst in its studies of the Nile discharge in 1951. It is a rather simple method, easily implemented in a program and that provides a direct estimation of the Hurst Exponent which is a precious indicator of the state of randomness of a time-series. It is especially interesting in revealing the existence of long-term dependence, which prevents, when it exists, the time-series to be reasonably modelised by a random walk.

I wish to expose here the minimal information for understanding the method, I will also provide a few references for those who wish to deepen their understanding of the matter.

Given a time-series with n elements X_1, X_2,...,X_n , the R/s statistic is defined as:

R/s(n)= \frac{1}{s}[Max_k{\sum_{i=1}^{k}(x_i-\bar{x})}-Min_k{\sum_{i=1}^{k}(x_i-\bar{x})}]

Where 1\leq k \leq n,
\bar{x} is the arithmetic mean
And s=\sqrt{\frac{1}{n}\sum(x_i-\bar{x})^2} is the standard deviation from the mean.

With this R/s value, Hurst found a generalization of a result found by Einstein in 1905 (Investigations on the Theory of the Brownian Movement ) as equation (11) (in the cited paper) in the following formula:

E[R/s(n)]=Cn^H \text{ as }n\rightarrow\infty \,\,\,\,(1)

Where H is the Hurst exponent.
From there, it is clear that we can obtain an estimation of the Hurst exponent pretty easily from an R/s analysis.
Several sites and articles propose a detailed methodology to implement R/s Analysis, I primarily use the approach exposed in a paper by O. Rose from February 1996:
Estimation of the Hurst Parameter of Long-Range Dependent Time Series
With a slight difference, however, I shall only plot one value of R/s for each value of d, in a manner similar to the following site: Estimating the Hurst Exponent
Anyway, I feel both articles are not very clear in their notations, and I therefore will detail the analysis I wish to implement.


Considering the time series above X_j (j=1,..,N)
We divide the time series into K^u (*) non-overlapping blocks of length d^u=\frac{N}{K^u}
And we fix: t_i=d^u(i-1)+1

Next we get a new time series W(i,k):


From there, we get the following rescaled range:



Taking the mean over i , we then get R/s(d^u):


Considering equation (1):            log(R/s(d^u))=log(C) +Hlog(d^u)

We can plot log(R/s(d^u)) vs log(d^u) for u varying, H is then the slope of the regression line which we simply get from the linear least squares method.
x_u=log(d^u) and y_u=log(R/s(d^u))

We get:

\boxed{H=\frac{U\sum_{u}x_uy_u -(\sum_{u}x_u)(\sum_{u}y_u)}{U(\sum_{u}x_{u}^2)-(\sum_{u}x_u)^2}} \text{ with }u \text{ varying from }1 \text{ to } U

(*): N and K^u are chosen adequately so that d^u is always an integer.


I implemented this method within an indicator for Metatrader 4 in order to compute a Fractalised Simple Moving Average for FOREX fluctuations. The implementation is not very interesting, partly because the computation time on this platform are not very good, I was therefore constraint to use the R/s method on a very limited number of data.
Anyway, the implementation can be seen on my other blog: Rescaled Range Analysis


October 14, 2009 - Posted by | Uncategorized

No comments yet.

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: