Overview of SAS High-Performance Forecasting Software

SAS High-Performance Forecasting software provides a large-scale automatic forecasting system. The software provides for the automatic selection of time series models for use in forecasting time-stamped data.

Given a time-stamped data set, the software provides the following automatic forecasting process:

1. accumulates the time-stamped data to form a fixed-interval time series

2. diagnoses the time series with time series analysis techniques

3. creates a list of candidate model specifications based on the diagnostics

4. fits each candidate model specification to the time series

5. generates forecasts for each candidate fitted model

6. selects the most appropriate model specification based on either in-sample or holdout-sample evaluation by using a model selection criterion

7. refits the selected model specification to the entire range of the time series

8. creates a forecast score from the selected fitted model

9. generate forecasts from the forecast score

10. evaluates the forecast using in-sample analysis

The software also provides for out-of-sample forecast performance analysis.

For time series data without causal inputs (input variables or calendar events), the HPF procedure provides a single, relatively easy-to-use batch interface that supports the preceding automatic forecasting process. The HPF procedure uses exponential smoothing models (ESM) and intermittent demand models (IDM) in an automated way to extrapolate the time series. The HPF procedure is relatively simple to use and requires only one procedure call.

For time series data with or without causal inputs (input variables or calendar events or both), the software provides several procedures that provide a batch interface that supports the preceding automatic forecasting process with more complicated models. These procedures must be used in the proper sequence in order to get the desired results. Forecasting time series of this nature normally requires more than one procedure call.

Input variables are recorded in the time-stamped data set. These input variables might or might not be incorporated in time series models used to generate forecasts.

Calendar events are specified with the HPFEVENTS procedure. These event definitions are used to generate discrete-valued indicator variables or dummy variables. These event definitions are stored in a SAS data set. These indicator variables might or might not be incorporated in time series models used to generate forecasts.

Given the specified calendar events and input variables, the HPFDIAGNOSE procedure diagnoses the time series and decides which, if any, of the calendar events or input variables are determined to be useful in forecasting the time series. The HPFDIAGNOSE procedure automatically generates candidate model specifications and a model selection list by using time series analysis techniques. These model specifications and model selection lists can then be used to automatically generate forecasts.

The user can specify model specifications with one of the following model specification procedures:

• The HPFARIMASPEC procedure enables the user to specify one of the family of autoregressive integrated moving average with exogenous inputs (ARIMAX) models.

• The HPFESMSPEC procedure enables the user to specify one of the family of exponential smoothing models (ESM).

• The HPFEXMSPEC procedure allows the forecast to be generated by an external source.

• The HPFIDMSPEC procedure enables the user to specify one of the family of intermittent demand models (IDM).

• The HPFSELECT procedure enables the user to specify a model selection list. The model selection list refers to one or more candidate model specifications and specifies how to choose the appropriate model for a given time series.

• The HPFUCMSPEC procedure enables the user to specify one of the family of unobserved component models (UCM).

Regardless of whether the model specifications or model selection lists are specified or automatically generated, the HPFENGINE procedure uses these files to automatically select an appropriate forecasting model, estimate the model parameters, and forecast the time series.

Most of the computational effort associated with automatic forecasting is time series analysis, diagnostics, model selection, and parameter estimation. Forecast scoring files summarize the time series model’s parameter estimates and the final states (historical time series information). These files can be used to quickly generate the forecasts required for the iterative nature of scenario analysis, stochastic optimization, and goal seeking computations. The HPFSCSUB function can be used to score time series information.