Please see my response (towards the end of a long correspondence) for a question on UCM diagnostics ("Unobserved Components Model Model Diagnostic"). It explains how one-step-residuals are produced by the UCM procedure, including the fact that UCM residuals are different from the usual OLS residuals that are produced by the standard regression procedures such as PROC REG. Consequently, many common things that are true for standard regression model do not hold for UCM models, e.g., adding a regerssor in the model need not decrease Rsquare, Rsquare can be negative, etc.
In general, for time series models one-step-ahead residuals based Rsquare is not as useful a fit statistic as usual Rsquare is in standard regression modeling. Diagnostics for time series models is a little more involved and one must rely on a variety fit measures besides one-step-ahead residuals based fit measures. PROC UCM provides a wide variety of fit/diagnostic measures (both numerical and graphical) including:
1. one-step-ahead residual based fit measures and plots
2. information theoretic fit measures based on likelihood
3. delete-one-measurement based influence statistics
4. prediction errors based on with-holding last few measurements of the time series (see the BACK= option in the ESTIMATE and FORECAST statements)
Registration is open! SAS is returning to Vegas for an AI and analytics experience like no other! Whether you're an executive, manager, end user or SAS partner, SAS Innovate is designed for everyone on your team. Register for just $495 by 12/31/2023.
If you are interested in speaking, there is still time to submit a session idea. More details are posted on the website.