Diagnostic Step

Create a Subset of Candidate Models

The purpose of the diagnostic step is to create a subset of the list of potential candidate models based on characteristics of the time series data. The list of potential candidate models is first limited to the model families that you specify. The following table shows the model families that might be considered during the diagnostic process.

Model Families for the Diagnostic Step

Family	Description
ARIMAX	ARIMA model that includes predictors that use ARIMA-REG identification order
ESM	Seasonal and nonseasonal exponential smoothing models
IDM	Intermittent demand (IDM) model
REGARIMA	ARIMA model that includes predictors that use REG-ARIMA identification order
UCM	UCM model with predictors

To create a subset of models, the time series data for each series is tested for certain characteristics that determine whether a type of model or model feature is appropriate for the data. For example, time series with seasonal trends (deterministic or stochastic) should be forecasta numerical prediction of a future value for a specified time period for each unique combination of BY variable values with models that have a seasonal component. Thus, a seasonality test is performed, and based on the results, the appropriate models are included in the list of candidate models. Likewise, inappropriate models are excluded from the list of candidate models.

The importance of the diagnostic step should not be underestimated. Applying a seasonal model to a nonseasonal time series, particularly one with a short history, can lead to over-parameterization or false seasonality. Applying a linear model to a nonlinear time series can lead to underestimation of the growth (or decline). Applying a non-intermittent model to an intermittent series will result in predictions biased toward zero.

If it is known that a time series has a particular characteristic, then the diagnostics should be overridden and the appropriate model should be used. For example, if the time series is known to be seasonal, the diagnostics should be overridden to always choose a seasonal model.

The following table shows the tests that are performed during the diagnostic process to identify model candidates.

Tests to Identify Candidate Models for the Diagnostic Step

Test	Description
INTERMITTENCY	Test for an Intermittent demand (IDM) model
MEAN	Test for a mean (constant) component to be included in the model
SEASONALITY	Test for a seasonal component to be included in the model
TRANSFORM	Test for a functional transformation to apply to the independent variable
TREND	Test for a trend component to be included in the model

Select Causal Factors for Candidate Models

Several causal factors might influence the dependent time series. The diagnostics for the multivariate time series determine which of the causal factors significantly influence the dependent time series. These diagnostics include cross-correlation analysis and transfer function analysis.

Once again, if it is known that a particular causal factor can influence the dependent time series, then the diagnostics should be overridden and the appropriate model should be used.

The following table shows the types of causal factors that might be used in the candidate models.

Causal FactorsThat Can Be Included in Candidate Models

Causal Factor Type	Description
INPUTS	Input variables that you supply
EVENTS	Event dummy variables created using an event definition
OUTLIERS	Outliers that you specify or identified using outlier detection
SEASONAL DUMMIES	Seasonal component specified using dummy variables