Jour. Ind. Soc. Ag.

Statistics
54(1), 2001 : 43-51
State Space Modelling Versus ARIMA Time-Series
Modelling
S. Ravichandran
l
and Prajneshu
Indian Agricultural Statistics Research Institute, New Delhi-llOOl2
(Received : June, 2000)
SUMMARY
Box-Jenkins Autoregressive Integmted Moving Avemge (ARIMA)
procedure is genemlly used for analyzing time-series data. In this article,
another approach, which is quite promising, viz. State space modelling
approach using Kalman filtering technique, is studied in detail. The
advantage of this technique is that it can take into account the time
dependency of the underlying pammeters. As an illustration, the two
approaches are compared for modelling and forecasting all-India marine
products export data.
Key words : ARlMA time-series procedure, Kalman filtering technique,
Marine products export data, State space modelling.
1. Introduction
In agriCUlture, data are generally collected over time and Box-Jenkins
ARIMA time-series modelling procedure is used to analyze these data (see e.g.
Box, et al. [2] ). In this approach, the underlying parameters are assumed to
be constants but in reality this assumption is rarely met. Evidently, the
parameters are time-dependent, and to handle such a situation another approach,
viz. State space modelling using Kalman filtering technique may be effectively
used. This procedure has been mostly used in control engineering (Meinhold
and Singpurwalla [5]) but, to the best of our knowledge, it has not so far been
used to model data from Indian agriCUlture. Accordingly, the purpose of this
paper is to highlight importance of this dynamical modelling technique so that
researchers may start applying it in their respective fields of studies. As an
illustration, the two procedures are compared to model data of all-India marine
products export.
Present address : Scientist, Computer Centre, Indian Council of Agricultural
Research, Krishi Bhawan, New Delhi-llOOOI
44 JOURNAL OF THE INDIAN SOCIETY OF AGRICULTURAL STATISTICS
2. State Space Model
As in Box-Jenkins approach, state space modelling is applicable only when
the time-series data is made stationary. This latter approach represents a
multivariate time-series through auxiliary variables (called state variables), some
of which mayor may not be directly observable. The state vector summarizes
all the infonnation from the present and past values of the time-series relevant
to the prediction of future values of the series. The observed time-series are
expressed as linear combinations of the state variables. The state space model
is also called a "Markovian representation," or "Canonical representation" of
a multivariate time-series process. It is defined with the help of two equations,
viz. "State transition equation" and "Measurement equation". Let X( be the r x I
vector of observed variables, after differencing and subtracting the sample mean
and let 1 be the state vector of dimension s, s 2: r, where the first r components
of ( consist of Xl' Let XI + kII denote the conditional expectation of XI +k based
on the infonnation available at time t. Then the last s r elements of 1 consist
of elements of XI + k II where k > 0 is specified. Then the model is defined
as
(1)
Eq. (1) is the state transition equation. F is the transition matrix of order
s x s. This matrix detennines the dynamic properties of the model. The
s x r coefficient matrix G is the input matrix which detennines the variance
structure of eq. (1). For model identification, the first r rows and r columns
of G are set to an identity matrix of order r x r. The input vector e is a sequence
l
of independent, nonnally distributed random vectors of dimension r with mean
vector 0 and covariance matrix Lee' The random error e( is sometimes called
the "Innovation vector". In addition to the state transition equation, state space
models usually include a measurement or observation equation that gives the
observed values XI as a function of the state vector 1' Since the observed
values XI are included in the state vector, the observation equation here merely
represents the extraction of first r components of the state vector. Hence the
measurement or observation equation is given by
(2)
where Ir is an identity matrix of dimension r x r.
--_..... _-------
45 STATE SPACE MODEWNG VERSUS ARIMA TIMESERIES MODEWNG
3. Estimation Procedure
State space procedure employs canonical correlation analysis for the
identification of state space model (see e.g. Aoki [1]). First fit a sequence of
unrestricted vector autoregressive (VAR) models and compute Akaike's
Information Criterion (AIC) for each model given by
1\
AIC
p
= n In (I I:p j) + 2pr
2
(3)
1\
where I:p is an estimate of innovation variance matrix obtained by fitting an
autoregressive model of order p. The smallest AIC value determines the number
of autocovariance matrices to be included in the canonical correlation analysis.
It may be mentioned that in this paper, the symbol ,"" indicates transpose of
a matrix. The vector autoregressive models are estimated using the sample
autocovariances :
r
i
= E(X
t
X'I j)
Thus the Yule-Walker equations are :
fro
r
1
r'l ro

r'
p 2
:::
...
r'_I]

ro
<Ill


=[[:]
(4)
and
fro

r'l
ro
rp
2
r'_l]
r'p_2
ro
'Pl

'P
P
p
=
(5)
Here <l>f are the coefficient matrices for the past observation form of the
vector autoregressive model, and 'Pi are the coefficient matrices for the future
observation form which are given respectively as
p
X
t
= L <I>f X
t
-
i
+e(
(6)
j = 1
where e is a vector white noise sequence with mean vector 0 and covariance
t
matrix I:
p
. This is the forward autoregressive form based on the past
46 JOURNAL OF THE INDIAN SOCIETY OF AGRICULTURAL STATISTICS
observations. The backward autoregressive fonn based on the future
observations is written as
p
X
t
::::; L 'Pf X
t
+i +11t
(7)
i = 1
where 111 is a vector of white noise sequence with mean vector 0 and covariance
matrix .Q
p
' Elements of state vector are determined with the help of a sequence
of canonical correlation analysis of the sample autocovariance matrices through
the selected order. This analysis computes the sample canonical correlations
of the past with an increasing number of steps into the future. Variables that
yield significant correlations are added to the state vector and the remaining
are excluded from further consideration. Once the state vector is detennined,
the state space model is fit to the data. The parameters in F, G and l:ee are
estimated using maximum likelihood procedure. Here
L::::; nJ2 In (Il:ee I) 112 trace ( ~ ! EE') (8)
where E is a r X n matrix of innovations given as
(9)
After the parameters are estimated, forecasts are obtained from the selected
state space model using Kalman filtering technique (Harvey [3]). The m-step
ahead forecast of Z t +m is Z t +mit ' where Z t +mit denotes the conditional
expectation of Z t +m given the infonnation available at time t, i.e.
(10)
where the matrix
H ::::; [lr 0]
The m-step ahead forecast error is
m-l
Zt+m- Zt+mlt::::; L 'Pjt+m-i
i =0
(11)
and its variance is
m-I
V
z
, m ::::; L 'Pi l:ee 'P'j
i=O
(12)
47 STATE SPACE MODEWNG VERSUS ARIMA TIME-SERIES MODEWNG
Letting Vl., 0 = 0, V l., m can be compute<,t recursively using Kalman
filtering technique as
(13)
Vz, m = Vz, m 1 + 'I'm - 1 I:ee '1" m 1
Thus the variance of m-step ahead forecast error of XI +m' obtained from
eq. (13), is
Vx,m H Vz,mH'
(14)
Goodness of fit of the selected model is assessed using Akaike's
Information Criterion (AIC), Schwartz- Bayesian Information Criterion (SBC),
Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) (Shumway
and Stoffer [6]). Lower the values of these statistics, better is the fitted modeL
It may be pointed out that none of the standard statistical packages, viz.
SPSS, GENSTAT, MSTAT are capable of handling state space analysis of
time-series data. To the best of our knowledge, only Statistical Analysis System
(SAS) software package contains programs for fitting state space models. SAS
modules, viz. PROC STATESPACE and PROC ARIMA, available in SAS [8]
are utilized for data analysis.
4. An Illustration
As an illustration, the univariate data pertaining to all-India marine
products export during the period 1975-'98, obtained from various reports of
MPEDA [4], is considered. The data is presented in Table 1 for ready reference.
Following Venugopalan and Prajneshu [7], ARIMA time-series modelling
technique is first applied to analyze the data. Autocorrelations (r
k
) and Partial
autocorrelations (<I>kk) up to lag 10 are worked out. Since the computed r
k
values
do not tail off towards zero, the original series is found to be nonstationary
and the stationarity is achieved after first differencing, i.e. d = 1. Patterns of
r
k
and <l>kk in respect of differenced data show that the partial autocorrelations
cut off after lag 1 and autocorrelations tail off towards zero after initial spikes.
However, to save space, r
k
and <l>kk values are not presented. Hence, for
describing the given data, ARIMA (1, I, 0) is identified as the best model
which is given by
(15)
where Il is a constant and <1>1 denotes the AR parameter. The above equation
is fitted to the given data and the estimates of Il and the <1>1 are obtained as
" "
Il 16.99, <1>1 = 0.87
----------------------...........- ~ ~ - ~
48 JOURNAL OF THE INDIAN SOCIETY OF AGRICULTURAL STATISTICS
Table 1. All India marine products export data
Year Quantity
(in thousand tonnes)
1975-'76 54.5
1976-'77 66.8
1977-'78 66.0
1978-'79 86.9
1979-'80 86.4
1980-'81 75.6
1981-'82 70.1
1982-'83 78.2
1983-'84 92.7
1984-'85 86.2
1985-'86 83.7
1986-'87 85.8
1987-'88 97.2
1988-'89 99.8
1989-'90 110.2
1990-'91 139.4
1991-'92 171.8
1992-'93 208.6
1993-'94 244.0
1994-'95 307.3
1995-'96 296.3
1996-'97 378.2
1997-'98 385.8
Subsequently, state space modelling procedure is applied to the
transformed data. Ale values are worked out to find out the number of
autoregressive models fit to the series. Table 2 shows that the smallest Ale
value is at lag 1. which determines the number of autocovariance matrices
analyzed in the canonical correlation phase. The selection of a first-order
autoregressive model by the Ale statistic looks reasonable for the given data
because partial autocorrelations for lags greater than I are not significant. Next
the Yule-Walker estimates for the selected autoregressive models are obtained.
For the given data, number of lags comes out as I. After the autoregressive
order selection process has determined the number of lags, canonical correlation
analysis phase selects the state vector. Once that is selected, the state space
model is estimated by maximum likelihood procedure. Information from the
canonical correlation analysis and from the preliminary autoregression analysis
is used to form preliminary estimate of the parameter of state space model.
49 STATE SPACE MODEWNG VERSUS ARIMA TIME-SERIES MODEWNG
Table 2. AIC values for state space model
Lag 0 2 3 4 5
AIC 111.79 106.97 108.60 110.55 112.51 113.37
Lag 6 7 8 9 10
AIC 114.35 115.90 117.78 119.76 121.36
This preliminary estimate is used as starting value for the recursive estimation
process.
For the given data, the final maximum likelihood estimate is obtained as
F(I, 1) = 0.549. Hence the fitted state space model is
Xt+l = Var[et+d = 250.76 (16)
The identified ARIMA (1, 1,0) and state space models given respectively
by eqs. (15) and (16) are fitted to the data ignoring a few observations towards
the end, say three, and forecast values on the basis of the identified models
are compared with the actual observations for assessing accuracy of the fitted
models. Forecast values for 1995-'96, 1996-'97, 1997- '98 using ARIMA and
state space models are respectively 364.50, 416.41, 463.72 and 348.06, 376.44,
398.03 thousand tonnes against the respective values of 296.3, 378.2,
385.8 thousand tonnes, thereby indicating that the latter approach yields closer
forecast values. Further, goodness of fit statistics, viz. AIC, SBC, RMSE and
MAE presented in Table 3 also show that the state space model provides lower
Table 3. Goodness of fit statistics
Statistic Model
ARIMA (1, 1,0) State Space Model
AIC 155.73 106.97
SBC 157.62 108.27
RMSE 13.87 9.45
MAE 12.59 6.87
values as compared to ARIMA model. Thus, for the data set under consideration,
state space modeHing technique performs better than the ARIMA approach.
Finally, forecasts of aU-India marine products for the next five years starting
from 1998-'99, using state space model, are respectively 415.87, 431.68, 446.36,
460.15 and 472.28 thousand tonnes. In order to get visual insight, the graph
of fitted state space model along with data points is exhibited in Fig.I.
50
JOURNAL OF mE INDIAN SOCIETY OFAGRICULTURAL STATISTICS
325
300
275
..
250

c
II
225

S
c

200
~
175
"

150
w
125
100
75

Year
50
1985
1995
1975 1980
1990
Fig. 1. Fitting of all-India marine products export data using state space model
5. Concluding Remarks
The importance of state space modelling using Kalman filtering technique
for the dynamic linear models is highlighted. It is suggested that researchers
start using this technique for modelling and forecasting time-series data. It will
also be worthwhile to study Generalized Kalman Filter (GKF) and Extended
Kalman Filter (EKF) involving nonlinear state space models. However, this is
a very difficult task as the software for fitting such models is not readily
available. Work in the direction is in progress and shall be reported in due
course of time.
ACKNOWLEDGEMENT
The authors are grateful to the referee for his valuable suggestions.
51
STATE SPACE MODEWNG VERSUS AR1MA T1MESERIES MODEWNG
REFERENCES
[1] Aoki, M., (1987). State Space Modelling of Time Series. Springer Verlag, New
York.
[2] Box, G.E.P., Jenkins, "G.M. and Reinsel, G.C., (1994). Time Series Analysis:
Forecasting and Control. 3rd edition. Prentice Hall, New Jersey.
[3] Harvey, A.C., (1984). A unified view of statistical forecasting procedures. J.
Forecasting. 3, 245-275.
[4] Marine P(oducts Export Development Authority, (1996). Statistics of Marine
Products Exports. India.
[5] Meinhold, RJ. and Singpurwalla, N.D., (1983). Understanding Kalman filter.
Amer. Statistician, 37, 123-127.
[6] Shumway, R.H. and Stoffer. D.S., (2000). Time Series Analysis and its
Applications. Springer Verlag, New York.
[7] Venugopalan, R. and Prajneshu, (1996). Trend analysis in all-India marine
products export using statistical modelling techniques. Ind. J. Fish.. 43,
107-113.
[8] SAS, (1990). SAS User's Guide, Version 6.12. SAS Institute Incorporation,
U.S.A.