Intervention and Monitoring
In Section 2.3.2 of Chapter 2 we introduced simple intervention ideas and considered in detail intervention into a first-order polynomial model. That intervention fed-forward information anticipating a major change in the level of a time series, modelled by altering the prior distribution for the level parameter to accommodate the change. Had the model used in that example been closed to intervention, then the subsequent huge change in the level of the series would have been neither forecast nor adequately estimated afterwards, the model/ data match breaking down entirely. In practice, all models are only components of forecasting systems which include the forecasters as integral components. Interactions between forecasters and models is necessary to adequately cater for events and changes that go beyond the existing model form. This is evident also in the illustrations of standard, closed models in Chapter 10 where deterioration in forecasting performance, though small, is apparent. In this Chapter, we move closer to illustrating forecasting systems rather than simply models, considering ways in which routine interventions can be incorporated into existing DLMs, and examples of why and when such interventions may be necessary to sustain predictive performance. The mode of intervention used in the example of Section 2.3.2 was simply to represent departures from an existing model in terms of major changes in the parameters of the model. This is the most widely used and appropriate method, although others, such as extending the model to include new parameters, are also important. An example dataset highlights the need for intervention.
KeywordsDiscount Factor Forecast Performance Observational Variance Seasonal Component Monitor Signal
Unable to display preview. Download preview PDF.