0
$\begingroup$

In DLM, what kind of preprocessing should be done before fitting the model?

$$Y_t = \beta_t'X_t+\epsilon_t$$ $$\beta_t = \beta_{t-1}+\eta_t$$

  1. Should I transform both $Y$ and $X$ into stationary time series? e.g. log, difference of log, etc.
  2. Should I centered and standardize $X$ like in Lasso/Ridge regression? If yes, should I do it variable-wise or time-wise?
  3. I think the any transformation will have an impact on the signal/noise ratio, does it matter? Because I think a high signal/noise ratio makes the estimation easier in state space model.

Any suggestions would be greatly appreciated. Thanks!

Improve this question
$\endgroup$
5
  • $\begingroup$ Hi: it's a state space model so you don't even stationarity. I would just leave as is and estimate. key thing is how stable $\beta$ is. If it's flying all over the place, that might be a sign that the model could be improved. note that you need the variances of $\epsilon_t$ and $\eta_t$ before you can run KF which I assume you estimate in some manner ? $\endgroup$
    – mlofton
    Mar 24, 2020 at 2:54
  • $\begingroup$ The variances of $\eta_t$ and $\epsilon_t$ can be estimated by combining KF and EM algorithm. But I think the estimated variance is actually impacted by the scale of $X_t$, like if $X_t$ is in km or miles, the variance can be different. And in SSM, I think a high $\frac{\sigma_\eta}{\sigma_\epsilon}$ leads to a more stable estimation. $\endgroup$
    – H.Yuanchen
    Mar 24, 2020 at 2:59
  • $\begingroup$ Hi: I can't say for sure but I don't think the units you use should matter ? Also, note that you can estimate the variances before the KF update by using the prediction error decomposition in Andrew Harvey blue book. I think the title is "kalman filter and structural models". You can also estimate them by KF-EM but my limited understanding of the latter leads me to believe that prediction error decomposition is easier. Hopefully others can chime in because I don't feel too confident giving advice on either A) standardizing or B) usefuleness of using EM instead of prediction error decomp. $\endgroup$
    – mlofton
    Mar 24, 2020 at 5:18
  • $\begingroup$ I'll check out prediction error decomposition. Can you please give the name of the book? I found he wrote 4 books about SSM. Thank you, mlofton! $\endgroup$
    – H.Yuanchen
    Mar 24, 2020 at 18:09
  • $\begingroup$ It's terse and compact ( each page is like 5 pages if someone else wrote them ) which is his general style. I hated it when I read it 20 years ago. Now I glance at it once in a while and realize how good it is. amazon.com/Forecasting-Structural-Models-Kalman-Filter-ebook/dp/… Note that his other books are very good also ( same description applies ) but they are not really focused on KF-SSM. $\endgroup$
    – mlofton
    Mar 25, 2020 at 17:39

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.