You can request that the X-13ARIMA-SEATS method select a model in a manner similar to the method used in X-11-ARIMA (Dagum 1988, 1983). Information about this model selection (PICKMDL) is based on the description in the X-13ARIMA-SEATS Reference Manual, Version 1.1 (US Bureau of the Census 2013c). You can request the PICKMDL method in one of the following ways:
by specifying the PICKMDL statement
by specifying more than one value for the _MODEL_ variable in the MDLINFOIN= data set (subset by BY group and series)
The default settings for the PICKMDL automatic model selection method classify a model as acceptable if all of the following conditions are true:
The absolute average percentage error of the extrapolated values within the last three years of data is less than 15%.
The p-value is greater than 5% for the fitted model’s Ljung-Box Q statistic test of the lack of correlation in the model’s residuals.
There are no signs of overdifferencing. Overdifferencing is indicated if the sum of the nonseasonal MA parameter estimates (for models with at least one nonseasonal difference) is greater than 0.9.
If a data set is specified in the MDLINFOIN= option and the data set contains more than one model for a series to be forecast, then the models described in the data set are candidates for the PICKMDL method of model selection. If the MDLINFOIN= option is not specified, then the candidate models are shown in Table 14, along with the order in which the models are considered. The order in which the model is considered is important when METHOD=FIRST is specified in the PICKMDL statement.
Table 14: PICKMDL Method Default ARIMA Models
| Order of Candidate Model | ARIMA Model Orders |
|---|---|
| 1 | (0 1 1)(0 1 1) |
| 2 | (0 1 2)(0 1 1) |
| 3 | (2 1 0)(0 1 1) |
| 4 | (0 2 2)(0 1 1) |
| 5 | (2 1 2)(0 1 1) |
No model is selected when none of the models in the MDLINFOIN= data set are acceptable. For more information about the output when no model is selected, see the section Final Automatic Model Selection Table.
The regARIMA model consists of a transformation, a regression component, and an ARIMA model component. For each series, the following conditions hold:
If no regression is specified in the MDLINFOIN= data set model but regressors are specified using the INPUT, EVENT, or REGRESSION statements, then the ARIMA models from the MDLINFOIN= data set are tested in conjunction with the regression variables specified in the INPUT, EVENT, and REGRESSION statements.
If no ARIMA model is specified in the MDLINFOIN= data set but an ARIMA model is specified using an ARIMA statement or TRANSFORM statement, then the regression information from each model specified in the MDLINFOIN= data set is used in conjunction with the ARIMA model specified by the TRANSFORM and ARIMA statements.
If no model information is specified in the MDLINFOIN= data set, then any model information specified by the TRANSFORM, INPUT, REGRESSION, EVENT, and ARIMA statements is used, and the PICKMDL statement is not in effect for that series.