Professional Documents
Culture Documents
=
A = = (4)
( 1) / [ ( -k 1)]
d
d d
k
= I + I + (5)
, defined as
1
0
( ) , 0
t x
x e t dt x
I = >
}
, represents the gamma function.
Obviously, {
t
X } is a FARIMA (p, a, q) process in which -0.5,0.5) ( d only if
t
d
X is an
ARMA (p, q) process. The only difference between FARIMA(p,d,q) and ARIMA(p,d,q) lies in the
value of d, and the former only take an integer among 0,1 and 2, but the latter is a fraction among
-0.5,0.5) ( . They share the same process except for the different value of d, converted to ARMA (p, q)
model after taking d-order difference on the time series. The accurate formula to compute the d-order
fractional difference of {
t
X } can be obtained from the above formulas, that is
0 0
( )
d
d k
j t t t t j
k
k j
W X B X X
= =
= A = = [
(6)
and
j) - + 1 ( ) + 1 (
) + 1 ( (-1)
=
j
d j
d
j
.
t
W so obtained can be judged as a ARMA sequence and then
prediction can be made according to ARMA model. A set of mature theory and methods about
FARIMA model based on ARMA model have been established.
Theorem
[3]
If 0<d<0.5, then the process FARIMA (p, d, q) is an asymptotic second-order
self-similar process and H=d+1/2, where H is the value of its Hurst parameter.
It is obviously that FARIMA (0, d, 0) (0<d<0.5) and FARIMA (p, d, q) process can both describe the
long correlation of the self-similar traffic. Easily seen from the definition of FARIMA, FARIMA (p, d,
q) is normal ARMA (p, q) process and short-related when d=0, and long-related when ) 5 . 0 , 0 ( d . In
addition, FARIMA (0, d, 0), the simplest form of FARIMA (p, d, q) when p=q=0, is defined as d-order
fractional difference of Gaussian white noise shortly named fractional differential noise. And the
operation is called fractal filtering. FARIMA (0, d, 0) is equivalent with FGN that the differential
coefficient d indicates the intensity of long-range dependence and the Hurst parameter, similar to FGN
process, can be calculated by H=d+0.5. Compared with the ARMA process and the FGN process,
- 90 -
Parameter Estimation and Application of Time-varying FARIMA Model
Xun-yi Ren, Yang-Yu, Jun-feng Zhang, Li-Ma, Xiao-dong Ma
International Journal of Advancements in Computing Technology, Volume 3, Number 3, April 2011
FARIMA (p, d, q) process can build model based on long-range and short-range dependence at the
same time, which reflects its flexibility.
3. Parameter Estimation of Traditional FARIMA Model
Known from the analysis in section II, the mentality of FARIMA(p,d,q) model modeling is that
t
X
is a FARIMA(p,d,q) model if
t t
d
a B X B ) ( = ) ( and
t
W (
t
d
t
X W = ) is a ARMA(p,q) model.
After fractional difference,
t
X is equivalent with
t
W . So FARIMA model can be built through two
steps, firstly calculating the value of d to achieve fractional difference, and then to obtain the
corresponding ARMA (p, q) model.
3.1. Estimation of the Parameter d
It's known from d=H-1/2 that the estimation of d can be converted to the calculation of the value of
H. A number of methods to calculate the value of H have been proposed, such as R/S method, variance
aggregation method, periodogram method
[4]
. Now the variance of wavelet coefficients based method is
mostly used and its idea is as follows:
If {x(t)} is a statistical self-similar process and the self-similarity coefficient is H, then the wavelet
coefficient
k j
d
,
can be obtained after wavelet transformation to x(t).
-j/2
, ,
-j/2
/ 2
( 1/ 2)
0,
( ), ( ) 2 ( ) (2 )dt
2 (2 ) ( - )2 ( 2 )
2 ( ) ( - ) ( (2 ) 2 ( ))
2
j
j k j k
j j j
jH j j jH
j H
k
d x t t x t t k
x u u k du t u
x u u k du x u x u
d
+
+
= =
}
=
}
=
}
=
(7)
In the formula (1), ) (
,
t
k j
is got by the expansion and translation to ) (t .
) (
,
t
k j
( k)dt - 2 ( 2 = ) (
-j -j/2
,
t t
k j
) is a binary orthogonal wavelet, R being its regularity index.
Then ] [
,k j
d Var is the variance of
k j
d
,
and
( 1/ 2) (2 1) (2 1) j
, 0, 0,
[ ] [2 ] =2 [ ] C2 =C2
j H j H j H
j k k k
Var d Var d Var d
+ + +
= = (8)
Taking the logarithmic data from both sides of the formula (2) and then making a linear fitting under
the minimum mean square error, a straight line can be obtained, whose slope is and function is
log
2
Var[d
j,k
]. Then the value of H can be got by 2H+1.
3.2. Parameter Estimation of Corresponding ARMA Process
The commonly used estimation method of ARMA (p, q) parameter can be taken and achieved
through the following two steps:
.Decide the order of p and q. Choose a small value for (p, q) such as 0, 1 and 2(p and q can't be
both 0) according to the experience of fitting FARIMA (p, d, q) model. Then do modeling examination
based on commonly used model selection standards such as AIC, BIC and decide the best value of
(p,q). Repeat the above steps by increasing the value of p and q(generally not more than 5) if no
suitable one found.
- 91 -
Parameter Estimation and Application of Time-varying FARIMA Model
Xun-yi Ren, Yang-Yu, Jun-feng Zhang, Li-Ma, Xiao-dong Ma
International Journal of Advancements in Computing Technology, Volume 3, Number 3, April 2011
.Estimate the parameters
i
and
j
. After the first step, all the parameters
(
2
2 1 2 1
, ,..., , , ,..., ,
q p
) can be got through the parameter estimation, like the approximate
maximum likelihood method.
4. Parameter Estimation of Time-Varying FARIMA model
4.1. Estimation of d (t)
Similar to traditional FARIMA model, time-varying d-order fractional difference of {
t
X } can be
obtained from the time-varying model with d(t) in place of d in the formula (6). Therefore, the
estimation of the value of d(t) is the key point.
This article uses improved time-varying H(t)
[5]
algorithm to estimate d(t). As to the network traffic
sequence } ,..., 2 , 1 = ), ( { N k k X , define
1 -
0 =
j) + X(k
1
= ) (
m
j
m
m
k X as its local time-lapse mean
sequence, where ) (k X
m
is a m-size time window added to the sequence.
Define
2
0
1
[ ( )-X ( )]
k
m
k
i
X i i
k
o
=
= as the sample variance of {X (k)} relative to the time-lapse
average value.
( 1) k i
-
i k i o o
o o o
A + + +
= is the difference of the variance. Window width m of
i
o
A
got
by different ( ) m X i meets the Hurst-exponent exponential law
[6]
, namely,
H
i
cm o
A
and the
linear relationship in the logarithmic diagram. Based on the above definition, the m-size time window
slides in the entire sampling region at