QUESTION:

What are the shortcomings of using the SAS proc AUTOREG with a time trend and seasonal dummies?

ANSWER:

Specifically, the problems that seem most pressing are:

For series that are generally stable (not trending or trending slowly), this is not an appropriate model.

Examination of the estimation results often shows that the time trend is statistically insignificant (and

could justifiably be eliminated). Retaining the trend adds a spurious element to the model and to the

forecasts (misspecification bias).

Further, the pattern of outliers sometimes leads to strong upward or downward time trends when no such

trend would be detected when the outliers are handled properly. For example, a large positive outlier in

the early history forces the regression on time to be negatively sloped as it tries to minimize the SSE,

similar biasing results if there are large outlier spikes in the recent history, but not the early history.

SAS' problems can be resolved in several ways:

Outliers can be modeled with dummy variables using either autoreg or proc arima (Transfer Functions -

more flexible than autoreg), but this can require extensive analyst time in identifying the outliers and

building and maintaining the database code and the model management code.

Code could be developed to automatically detect the outliers, generate the dummy variables, and

dynamically modify the forecasting model accordingly. The outliers detected here are only pulses with no

ability to identify seasonal pulses, level shifts or time trends.

Sometimes a time trend is estimated but visual inspection reveals a simple shift in the level of the factor.

Forecasting with the time trend gives spurious forecasts, while modeling the level shift gives much more

believable results. Another possible problem is that there may be a number of time trends.

Again, considerable analyst time needs to be expended in identifying these shifts, developing and

maintaining the data management code and the model management code. Once again, code could be

developed to do these things automatically - automatically detect level and trend shifts and dynamically

managing the requisite dummy variables and model specifications.

AUTOBOX - does all these things (and more) automatically. It would save many hours of analyst time, give

you a much improved starting point and put you years ahead of the way you do things now.