Forecast Model Selection Graph Details
Introduction
This chapter provides details on several aspects of the forecast model selection graph used in the SAS Forecast Server Procedures software.
The forecast model selection graph is a directed acyclic graph (DAG) defined through the use of the HPFSELECT procedure and the various model specification procedures. Nodes in the forecast model selection graph are the XML specifications, and the arcs are defined by the references in list specifications that are created via PROC HPFSELECT.
The forecast model selection graph supports a wide range of model selection and forecasting options and is an evolution of previous mechanisms. Earlier versions of SAS Forecast Server Procedures software permitted a one-level structure known as a model selection list. It held a set of alternative time series models evaluated to pick the best model for each time series realization where it was applied. With the addition of combined forecasts, the extension of the model selection list concept to the forecast model selection graph is a natural one. You can easily see the motivation with these two scenarios:
selecting between the best performing individual model forecast and a combined forecast
combining the best performing forecast from a model selection list with other forecasts
Given these scenarios, the need for the forecast model selection graph generalization becomes apparent. Indeed, supporting these two scenarios basically provides the framework for the generality of the forecast model selection graph. Subsequent sections of this chapter demonstrate by example use of these concepts and provide insight into any rules on usage imposed by the software.