QUESTION:
Please explain, in words
not equations, the role of pre-whitening and transfer function model
identification.
ANSWER:
The procedure for transfer
function model identification outlined by Box and Jenkins uses the cross
correlations between
two prewhitened series
to tentatively identify model form. The first step to this process is to
develop an ARIMA model for
each of the
user-specified input time series in the equation. Each series must then be made
stationary by applying the
appropriate differencing
and transformation parameters from its ARIMA model. The stationary time series
are, in turn,
"prewhitened". Prewhitening refers
to the process of applying a given set of autoregressive and moving average
factors
to a stationary series.
Each input series is prewhitened by its own ARIMA model AR (autoregressive) and
MA (moving
average) factors. The
output series is also prewhitened by the input series AR and MA factors. If
there is more than one
input series, then the
stationary output series is prewhitened once for each different input.
Prewhitening is necessary
because it removes the intrarelationship in the
individual series. This allows you to more accurately assess the
interrelationship
between the input and the output series. The cross correlations between the
prewhitened input and
output reveal the extent of this
interrelationship. The cross correlations
can be converted to estimates of the impulse
response weights or regression weights. The
pattern in the impulse response weights indicate can suggest a tentative
transfer function model. By applying these
impulse response weights to the input series to predict the output series,
one can generate a
preliminary estimate of the noise series. Following the rules for ARIMA model
identification, the
patterns in the autocorrelations and partial
autocorrelations of the tentative noise process give clues as to the initial
form of the noise model.
Given the identified transfer function and noise model, one can proceed to the
model
estimation/diagnostic
checking phase.
The process of
identifying a transfer function model, as outlined by Box and Jenkins, is in
itself a multi-step
procedure. The very first step is to develop
the optimal ARIMA model for each time series that is to be included in
the transfer function
equation. These ARIMA models play a critical role in the analysis of the
relationship between
each input series and
the output series. AFS suggests that
the ARIMA model used in prewhitening be kept as simple
as possible, but not too
simple. We suggest that interventions,
both mean and variance, should not be included as it is
possible to mask interrelationships by
fitting overly complicated intrarelationships.
AUTOBOX offers option to use
differencing factors in
the "Identification phase". AFS has found that over differencing can
lead to poor identification.
Box and Jenkins, faced
with "expensive computing costs" took the shortcut and used the ARIMA
model differences,
possibly different, in
the Transfer Function identification phase. We suggest that the user try it
both ways, that is with
and without the differencing factors.