I have a 6 state Kalman Filter (Unscented).

When I use a diagonal matrix only for Q (i.e only the diagonal has covariances), I get a "smooth" plot of estimate against actual.

If I use the entire Q matrix (i.e include all covariances) I get an estimate that oscillates around actual wildly.

Is it incorrect to just use diagonal?


I am using 6 correlated Equity Stocks Price timeseries. In constructing the covariance matrix I use the difference between consecutive prices. I do NOT log the data.

  • $\begingroup$ Are the state process noises correlated? If the "true" system has them correlated, then you should use the off-diagonal entries. If you don't know, just use the diagonal entries. $\endgroup$ – Peter K. Nov 27 '15 at 17:05
  • $\begingroup$ Yes - they are correlated - they are all Equity Price time series. Using the full matrix makes the Kalman unusable - what might I be doing wrong? $\endgroup$ – ManInMoon Nov 27 '15 at 17:12
  • $\begingroup$ Can you copy paste the matrices? $\endgroup$ – Nir Regev Nov 27 '15 at 17:44
  • $\begingroup$ This question and the book that the (broken) links go to may be of some interest. $\endgroup$ – Peter K. Nov 27 '15 at 17:47
  • $\begingroup$ Yes - I had already seen those. But it's nice to talk to someone with personal experience. I am now wondering if I should weight all "off diagonal" values by some constant - perhaps that would make the estimate less jumpy. Or maybe I should have used log difference etc. $\endgroup$ – ManInMoon Nov 27 '15 at 17:59

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.