The details on TSMODEL can be found in the Algorithms manual, which is accessible via Help > Documentation in PDF format. As far as I know, since the input data are time series and, therefore, cannot be randomly sampled, you would need to select a training sample via the usual variable selection methods and then apply it to your holdout sample. But, inevitably this means testing on a different time period from the estimation period.
For neural networks, when you run either procedure, save the model as an xml file (on the Export tab in the dialog box). Then, using Utilities > Scoring Wizard,select your xml file and specify any mappings needed for the computations.
--
Original Message:
Sent: 6/1/2023 10:31:00 AM
From: Ryan Burge
Subject: ARIMA Data Partitioning / Splitting & Neural Network Forecasting Options Q&A?
Hi everyone. A couple questions I could use your help with...
- For the SPSS Expert Modeler (ARIMA, in particular), do you know what the default model train/test split (e.g., 50/50, 70/30, 80/20, etc.?) is for the model generation against the base dataset? Can this be changed, or is it just a default in SPSS? I don't see details in either the SPSS EM documentation, in the model run outputs, or in the model set-up options/dialogs...
- Also, for the Neural Network analysis feature in SPSS, do you know how to use the model to actually forecast in SPSS using the trained/tested model?
Many thanks!
------------------------------
Ryan Burge
------------------------------