Interventions

An intervention is a special indicator variable, computed automatically by the system, that identifies time periods affected by unusual events that influence or intervene in the normal path of the time series you are forecasting. When you add an intervention predictor, the indicator variable of the intervention is used as a regressor, and the impact of the intervention event is estimated by regression analysis.

To add an intervention to the Predictors list, you must use the Intervention Specification window to specify the time or times that the intervening event took place and to specify the type of intervention. You can add interventions either through the Interventions item of the Add action or by selecting Tools from the menu bar and then selecting Define Interventions.

Intervention specifications are associated with the series. You can specify any number of interventions for each series, and once you define interventions you can select them for inclusion in forecasting models. If you select the Include Interventions option in the Options menu, any interventions that you have previously specified for a series are automatically added as predictors to forecasting models for the series.

From the Develop Models window, invoke the series viewer by selecting the View Series Graphically icon or Series under the View menu. This displays the Time Series Viewer, as was shown in Figure 43.2.

Note that the trend in the PETROL series shows several clear changes in direction. The upward trend in the first part of the series reverses in 1981. There is a sharp drop in the series towards the end of 1985, after which the trend is again upwardly sloped. Finally, in 1991 the series takes a sharp upward excursion but quickly returns to the trend line.

You might have no idea what events caused these changes in the trend of the series, but you can use these patterns to illustrate the use of intervention predictors. To do this, you fit a linear trend model to the series, but modify that trend line by adding intervention effects to model the changes in trend you observe in the series plot.