Spectral forecasting uses frequency domain methods to fit a “generic” Box-Jenkins model to the data. The preparations (transformation and differencing) are the same as for a Box-Jenkins model.

 

It is a non-parametric method, so parsimony is not a well-defined concept. Technically, the estimates of cycles are smoothed to eliminate possibly spurious effects due to sampling error.

 

The procedure used is technically very complex. For those of you who are familiar with z-transforms, we describe the algorithm elsewhere.

 

Advantages

The advantages of spectral forecasting tend also to be disadvantages. It is an automatic procedure once you have chosen the preliminary transformation and trend model. Given adequate data, it can produce good forecasts quickly and painlessly. It can do better than Box-Jenkins, particularly when no model with just a few parameters seems to work well.

 

Disadvantages

As an automatic procedure, there are no diagnostics along the way to steer us away from problems. It is less tolerant of mistakes in preliminary transformation and requires more data (100+ observations) than parametric techniques.

 

@SPECFORE procedure

You can do spectral forecasting using the procedure @SPECFORE. The syntax for the procedure is:

 

@specfore( options )   series   fstart  fend   forecasts

 

series is the series to forecast and forecasts is the target series for the forecasts. fstart and fend gives the range to forecast.

 

The options select the transformation and trend model.

 

trans=[none]/log/root

Select the appropriate preliminary transformation.

 

diffs=Number of regular differences [0]

sdiffs=Number of seasonal differences [0]

[constant]/noconstant

Select NOCONSTANT to exclude the intercept from the differenced model.

 

Example

SPECFORE.RPF provides an example of spectral forecasting applied to the same interest rate spread series used in ARIMA.RPF.

 

Copyright © 2025 Thomas A. Doan