The RATS Software Forum

dear Tom:

I read codes of "an introduction to state space..." JFCC & SJK. In chart 3., UK data, I found the DLM estimate twice both in deterministic and stochastic models. The first estimate for @stampdiages with default option type=filter, the second estimate for level and irregular component with option type=smooth.
Why does not diagnostic test residue from rather smoothed state equation than filtered state equation?

hardmann

Each model is estimated once. The second model has one DLM used to do the estimation of the component variances and another to do the smoothed states.

Diagnostics are done based upon Kalman filtering because, if the model is correct, the Kalman filter vhats are serially independent, and thus it's relatively easy to create diagnostics. The Kalman smoothing statistics aren't independent, and, in fact, have a very complicated covariance pattern.