"KUL" <m9828353@urc1.cc.kuleuven.ac.be wrote:
QUESTION:
What is Multiple Time Series Analysis ( MTSA) ?
ANSWER:
I'm glad you asked.
Univariate time series simply uses the past values of
one series to develop an autoprojective relationship which
goes by many names... ARIMA , RATIONAL EXPECTATIONS to name
two. It explicitly uses the previous values BUT since these
previous values captured the impact of omitted causal series
(X for eXogenous) thus IMPLICITELY captures the effect of
these omitted X's. Please see
http://www.autobox.com/t1c9.html for an enlightening
discussion if the dual role that an ARIMA plays. See
http://www.autobox.com/t1a13a.html for modern procedures for
ARIMA MODEL identification. Note that Pulses, Seasonal
Pulses, Level Shifts and Local Time Trends have to be taken
into consideration along with potentially changing variance ,
parameters and even model form. More on this can be found at
http://www.autobox.com/teach.html.
Transfer Functions are a form or MTSA where the
assumption is made that the X's cause the Y and not
vice-versa . A single equation model with one or more inputs
an be identified and estimated and checked diagnostically.
Please read "Lies My Mother Never Told Me" which summarizes
the roles of regression vis-a-vis Transfer Functions at
http://www.autobox.com/t1c6.html.
A regression is a particular case of a transfer function
and assumes bunches of things. A Transfer Function can
reduce to a simple multiple regression if the data so
indicates and in-model step-down testing reduces unneeded
structure. For a discussion of "REGRESSION AS A SUBSET OF
TRANSFER FUNCTIONS" please see
http://www.autobox.com/t1c8.html.
A more general MTSA is when there are multiple dependent
series and possibly multiple input series. This happens
quite naturally when you are interested in estimating or
predicting simultaneously products that are either
substitutes or complements or have in general a cross
dependence. This is called VECTOR ARIMA and a number of
vendors unable to deploy VECTOR ARIMA promote a subset called
VAR (NO IMA). For downloadables and explanatory material on
VARIMA please see http://www.autobox.com/mts.html. Using
VARIMA, variables can be either endogenous or exogenous.
Endogenous variable prediction can then be based on a
combination of lags of the series of interest and appropriate
lags of other variables (endogenous or exogenous) in the
model.
I hope this helps. If you wish to pursue more please
either see http://www.autobox.com/referenc.html or sign up
for the seminar on AUTOBOX coming up in June in Washington
D.C. at the International Institute of Forecasters annual
meeting. http://ifsm2.ifsm.umbc.edu/ISF/
P.S. The same Gaussian assumptions as in ARIMA must be
implemented in either Transfer Functions or Vector ARIMA.