752 · Chapter 8: Time Series Analysis and Forecasting IMSL STAT/LIBRARY
Description
Routine SPWF performs least-squares estimation of parameters for successive autoregressive models of a stationary stochastic process given a sample of n = NOBS observations {Wt} for t = 1, ¼, n.
Let
ˆ WMEAN
m =
be the estimate of the mean m of the stochastic process {Wt} where
1
known
ˆ 1
unknown
n t t
n W
m m
m m
=
ìï
= íïî
å
Consider the autoregressive model of order k defined by
( ) 0
k B Wt At k
f % = ³
where
t t
ˆ
W% =W -m
and
2
1 2
( ) 1 k 1
k B kB kB kkB k
f = -f -f - -L f ³
Successive AR(k) models are fit to the centered data using Durbin’s algorithm (1960) based on the sample autocovariances
1
ˆ ( ) 1
n k(
tˆ )(
t kˆ ) 0
t
k W W k
s n - m + m
=
=
å
- - ³Note that the variance
ˆ (0)
s*used in the fitting algorithm is adjusted by the amount
d=
WNADJaccording to ˆ (0) (1 ) (0) ˆ
s* = +d s
See Robinson (1967, page 96).
Iteration to the next higher order model terminates when either the expected mean square error
of the model is less than
EPSor when k =
MLFOP. The forecast operator
f= (
f1,
f2, …,
fk*)
Tfor
k*=
LFOPis contained in
FOP. See also Craddock (1969).
IMSL STAT/LIBRARY Chapter 8: Time Series Analysis and Forecasting · 753
Required Arguments
W
— Vector of length
NOBScontaining the time series. (Input)
PAR
— Vector of length
NPARcontaining the autoregressive parameters. (Input)
LAGAR— Vector of length
NPARcontaining the order of the autoregressive parameters.
(Input)
The elements of
LAGARmust be greater than zero.
PMA
— Vector of length
NPMAcontaining the moving average parameters. (Input)
LAGMA— Vector of length
NPMAcontaining the order of the moving average parameters.
(Input)
The elements of
LAGMAmust be greater than zero.
ICNST
— Option for including the overall constant in the model. (Input)
ICNST Action0 No overall constant is included.
1 The overall constant is included.
CNST
— Estimate of the overall constant. (Input)
AVAR— Estimate of the random shock variance. (Input)
AVAR
must be greater than 0.
ALPHA
— Value in the exclusive interval (0, 1) used to specify the 100(1
-ALPHA)%
probability limits of the forecasts. (Input) Typical choices for
ALPHAare 0.10, 0.05, and 0.01.
MXBKOR
— Maximum backward origin. (Input)
MXBKOR
must be greater than or equal to zero and less than or equal to
NOBS-max(
MAXAR,
MAXMA) where
MAXAR= max(
LAGAR(i)) and
MAXMA
= max(
LAGMA(j)). Forecasts at origins
NOBS-MXBKORthrough
NOBSare generated.
MXLEAD
— Maximum lead time for forecasts. (Input)
MXLEADmust be greater than zero.
FCST
—
MXLEADby (
MXBKOR+ 3) matrix defined below. (Output)
754 · Chapter 8: Time Series Analysis and Forecasting IMSL STAT/LIBRARY Column
Content
j
Forecasts for lead times l = 1, …,
MXLEADat origins
NOBS-MXBKOR-1 + j, j = 1, …,
MXBKOR+ 1.
MXBKOR
+ 2 Deviations from each forecast that give the 100(1
-ALPHA)%
probability limits.
MXBKOR
+ 3 Psi weights of the infinite order moving average form of the model.
Optional Arguments
NOBS
— Number of observations in the time series
W. (Input)
NOBS
must be greater than
ICONST+ max(
LAGAR(i)) + max(
LAGMA(j)).
Default:
NOBS= size (
W,1).
IPRINT
— Printing option. (Input) Default:
IPRINT= 0.
IPRINT Action
0 No printing is performed.
1 Prints the forecasts for lead times l = 1, …,
MXLEADat each origin
t = (NOBS-MXBKOR), …,
NOBS, the 100(1
-ALPHA)% probability limit deviations, and the psi weights.
NPAR
— Number of autoregressive parameters. (Input)
NPARmust be greater than or equal to zero.
Default:
NPAR= size (
PAR,1).
NPMA
— Number of moving average parameters. (Input)
NPMAmust be greater than or equal to zero.
Default:
NPMA= size (
PMA,1).
LDFCST
— Leading dimension of
FCSTexactly as specified in the dimension statement in the calling program. (Input)
LDFCST
must be greater than or equal to
MXLEAD. Default:
LDFCST= size (
FCST,1).
FORTRAN 90 Interface
Generic:
CALL NSBJF(W, PAR, LAGAR, PMA, LAGMA, ICNST, CNST, AVAR, ALPHA, MXBKOR, MXLEAD, FCST [,…])Specific: The specific interface names are
S_NSBJFand
D_NSBJF.
IMSL STAT/LIBRARY Chapter 8: Time Series Analysis and Forecasting · 755
FORTRAN 77 Interface
Single:
CALL NSBJF (NOBS, W, IPRINT, NPAR, PAR, LAGAR, NPMA, PMA, LAGMA, ICNST, CNST, AVAR, ALPHA, MXBKOR, MXLEAD, FCST, LDFCST)Double: The double precision name is
DNSBJF. Example
Consider the Wölfer Sunspot Data (Anderson 1971, page 660) consisting of the number of sunspots observed each year from 1749 through 1924. The data set for this example consists of the number of sunspots observed from 1770 through 1869. Routine
NSBJFis used to computed forecasts and 95% probability limits for the forecasts for an
ARMA(2, 1) model fit using routine
NSPE (page 727). With MXBKOR= 3, columns one through four of
FCSTgive forecasts given the data through 1866, 1867, 1868, and 1869, respectively. Column 5 gives the deviations from the forecast for computing probability limits, and column six gives the psi weights, which can be used to update forecasts once more data is available. For example, the forecast for the 102-nd observation (year 1871) given the data through the 100-th observation (year 1869) is 77.21, and 95% probability limits are given by 77.21
m56.30. After observation 101 (W
101for year 1870) is available, the forecast can be updated by using equation 7 with the psi weight (
y1= 1.37) and the one-step-ahead forecast error for observation 101 (W
101-83.72) to give
77.21 + 1.37 (W
101-83.72)
Since this updated forecast is one step ahead, the 95% probability limits are now given by the forecast
m33.22.
USE GDATA_INT USE NSPE_INT USE NSBJF_INT USE WRRRL_INT
INTEGER LDFCST, MXBKOR, MXLEAD, NOBS, NPAR, NPMA
PARAMETER (MXBKOR=3, MXLEAD=12, NOBS=100, NPAR=2, NPMA=1, &
LDFCST=MXLEAD)
!
INTEGER ICNST, LAGAR(NPAR), LAGMA(NPMA), NCOL, NROW
REAL ALPHA, AVAR, CNST, FCST(LDFCST,MXBKOR+3), PAR(NPAR), &
PMA(NPMA), RDATA(176,2), W(NOBS), WMEAN CHARACTER CLABEL(MXBKOR+4)*40, RLABEL(1)*6
!
EQUIVALENCE (W(1), RDATA(22,2))
!
DATA LAGAR(1), LAGAR(2)/1, 2/
DATA LAGMA(1)/1/
DATA RLABEL/’NUMBER’/, CLABEL/’%/Lead%/Time’, &
’%/Forecast%/From 1866’, ’%/Forecast%/From 1867’, &
’%/Forecast%/From 1868’, ’%/Forecast%/From 1869’, &
’ Deviation %/ for 95% %/Prob. Limits’, ’%/%/Psi’/
! Wolfer Sunspot Data for
! years 1770 through 1869 CALL GDATA (2, RDATA, NROW, NCOL)
756 · Chapter 8: Time Series Analysis and Forecasting IMSL STAT/LIBRARY
! Compute preliminary parameter
! estimates for ARMA(2,1) model CALL NSPE (W, CNST, PAR, PMA, AVAR)
!
! Include constant in forecast model ICONST = 1
! Specify 95 percent probability
! limits for forecasts ALPHA = 0.05
! Compute forecasts CALL NSBJF (W, PAR, LAGAR, PMA, LAGMA, &
ICNST, CNST, AVAR, ALPHA, MXBKOR, MXLEAD, FCST)
! Print results
CALL WRRRL (’FCST’, FCST, RLABEL, CLABEL, FMT=’(5F9.2, F6.3)’)
!
END
Output
FCST
Deviation Lead Forecast Forecast Forecast Forecast for 95%
Time From 1866 From 1867 From 1868 From 1869 Prob. Limits Psi 1 18.28 16.62 55.19 83.72 33.22 1.368 2 28.92 32.02 62.76 77.21 56.30 1.127 3 41.01 45.83 61.89 63.46 67.62 0.616 4 49.94 54.15 56.46 50.10 70.64 0.118 5 54.09 56.56 50.19 41.38 70.75 -0.208 6 54.13 54.78 45.53 38.22 71.09 -0.326 7 51.78 51.17 43.32 39.30 71.91 -0.286 8 48.84 47.71 43.26 42.46 72.53 -0.169 9 46.53 45.47 44.46 45.77 72.75 -0.045 10 45.35 44.69 45.98 48.08 72.77 0.041 11 45.21 44.99 47.18 49.04 72.78 0.077 12 45.71 45.82 47.81 48.91 72.82 0.072
Comments
1. Workspace may be explicitly provided, if desired, by use of
N2BJF/DN2BJF. The reference is:
CALL N2BJF (NOBS, W, IPRINT, NPAR, PAR, LAGAR, NPMA, PMA, LAGMA, ICNST, CNST, AVAR, ALPHA, MXBKOR, MXLEAD, FCST, LDFCST, PARH, PMAH, PSIH, PSI, LAGPSI)
The additional arguments are as follows:
PARH
— Work vector of length equal to
IARDEG+ 1.
PMAH
— Work vector of length equal to
IMADEG+ 1.
PSIH
— Work vector of length equal to
MXLEAD+ 1.
PSI
— Work vector of length equal to
MXLEAD+ 1.
IMSL STAT/LIBRARY Chapter 8: Time Series Analysis and Forecasting · 757 LAGPSI
— Work vector of length equal to
MXLEAD+ 1.
2. If the
Wseries has been transformed using a Box-Cox transformation with parameters
POWERand
SHIFT, the forecasts and probability limits for the original series may be obtained by application of routine
BCTR (page 689) with the same parameters andargument
IDIRset equal to one.
Description
Routine
NSBJFcomputes Box-Jenkins forecasts and their associated probability limits for a nonseasonal ARMA model given a sample of n =
NOBSobservations {W
t} for t = 1, 2, …, n.
Suppose the time series {W
t} is generated by a nonseasonal ARMA model of the form ( )
B Wt 0( )
B At t{0, 1, 2, }
f = +q q Î ± ± K
where B is the backward shift operator,
q0=
CONST,
(1) (2) ( )
1 2
(1) (2) ( )
1 2
( ) 1 ( ) 1
l l l p
p
l l l q
q
B B B B
B B B B
f f f
q q q
f f f f
q q f q
= - - - -
= - - - -
L L
p = NPAR and q = NPMA. Without loss of generality, we assume
1 (1) (2) ( )
1 (1) (2) ( )
l l l p
l l l q
f f f
q q q
£ £ £ £
£ £ £ £
L L
so that the nonseasonal ARMA model is of order (p¢, q¢) where p¢ = lf(p) and q¢ = lq(q). Note that the usual hierarchal model assumes
( ) 1
( ) 1
l i i i p
l j j j q
f q
= £ £
= £ £
The Box-Jenkins forecast at origin t for lead time l of Wt+l is defined in terms of the difference equation
0 1 (1) ( )
1 (1) ( )
ˆ ( ) [ ] [ ]
[ ] [ ] [ ]
t t l l p t l l p
t l t l l q t l l q
W l W W
A A A
f f
q q
q f f
q q
+ - + -
+ + - + -
= + + +
+ - - -
L L where
1
0, 1, 2,
[ ]
ˆ ( ) 1, 2,
ˆ (1) 0, 1, 2,
[ ]
0 1, 2,
t k t k
t
t k t k
t k
W k
W W k k
W W k
A k
+ +
+ + -
+
= - -
= íìïïî =
ì - = - -
= íïïî =
K K
K K The 100(1 - a)% probability limits for Wt+l are given by
758 · Chapter 8: Time Series Analysis and Forecasting IMSL STAT/LIBRARY
1 1/ 2 2 / 2
1
ˆ ( )
t1
l j Aj
W l za - y s
=
ì ü
± í + ý
î
å
þwhere z
(1-a/2)is the 100(1
-a/2) percentile of the standard normal distribution,
2
AVAR
sA =
and {
yj} are the parameters of the random shock form of the difference equation. Note that the forecasts are computed for lead times l = 1, 2, …, L at origins t = (n
-b), (n
-b + 1), …, n where L =
MXLEADand b =
MXBKOR.
The Box-Jenkins forecasts minimize the mean square error ˆ
2[
t l t( )]
E W+ -W l
Also, the forecasts may be easily updated according to the equation
1 1
ˆ
t( ) ˆ
t( 1)
l tW+ l =W l+ +y A+