Singular covariance in SSpace estimation.

[rpn4]

Hello,

I'm estimating a SSpace model using E-views and sometimes I get the Warning note of singular covariance - coefficients are not unique. However, my interest in the model is solely to generate the state series, and not to get the coefficients; and the state series seems okay to me. I estimated the model over and over again, and keep getting the same state series. Does it mean that although the coefficients are not unique the state series is robust and I can use it? If not, what do you suggest I try? I've already tried to increase the number of observations, take log of the variables... What else can I do? Stating initial values may help solving the problem?

Cheers.

[rpn4]

In case somebody is also facing a similar problem, it turns out that stating initial values does help solving the problem.

[BobJ]

Thanks rpn4 for clarifying this. I had a similar question regarding the warning message of singularity in SSpace estimation. I'm also only interested in obtaining state series. Can you please calirify how did you set (or change) the initial condition values ?

[rpn4]

All you have to do is set the initial parameter (use @param, as in the manual). I used the OLS estimates as initial parameters.

[tin_terres]

Hi rpn4 i'm having the same problem. My state equation is a VAR(1) so i used proxy of those variables an estimated my VAR(1). But using this numbers didn't solve my problem. Do you know what else i can do?

[SnakeS2k]

After an Initial calibration with starting value that give me a convergence even if with a failure (read below).

Sspace: DIEBOLD_KALMAN_FILTER
Method: Maximum likelihood (Marquardt)
Date: 10/02/14 Time: 09:16
Sample: 7/19/1999 1/17/2014
Included observations: 3711
Failure to improve Likelihood after 5 iterations

Coefficient Std. Error z-Statistic Prob.

C(1) 0.092766 0.000488 190.2573 0.0000
C(2) 0.884205 167.8895 0.005267 0.9958
C(3) 0.882498 487.2424 0.001811 0.9986
C(4) 0.881565 640.8659 0.001376 0.9989
C(5) 0.881183 291.3343 0.003025 0.9976
C(6) 0.882804 113.5213 0.007777 0.9938
C(7) 0.885380 285.5461 0.003101 0.9975
C(8) 0.887086 436.4433 0.002033 0.9984
C(9) 0.888180 394.5858 0.002251 0.9982
C(10) 0.888930 277.8070 0.003200 0.9974
C(11) 0.889475 356.6718 0.002494 0.9980
C(12) 0.889888 360.8989 0.002466 0.9980
C(13) 0.890212 167.3506 0.005319 0.9958
C(14) 0.890471 29.54602 0.030138 0.9760
C(15) -0.511781 21.84499 -0.023428 0.9813
C(16) 0.345823 0.010791 32.04801 0.0000
C(17) 0.571374 0.001811 315.5247 0.0000
C(18) -0.670912 0.002203 -304.5966 0.0000
C(19) 0.852860 0.001551 549.7949 0.0000
C(20) -0.242892 47.89067 -0.005072 0.9960
C(21) -0.034781 0.003810 -9.127893 0.0000
C(22) 0.237985 0.001962 121.2782 0.0000
C(23) 0.079919 8.81E-05 907.5590 0.0000
C(24) 0.789928 0.000326 2425.828 0.0000
C(25) -4.907890 157.9000 -0.031082 0.9752
C(26) -0.616572 0.002950 -209.0331 0.0000
C(27) -0.877873 5.30E-05 -16576.05 0.0000
C(28) -0.024501 0.004694 -5.219184 0.0000
C(29) -1.055007 0.014213 -74.23042 0.0000
C(30) 1.113292 5.55E-06 200420.6 0.0000
C(31) 0.740281 0.010353 71.50653 0.0000
C(32) -0.040536 0.006699 -6.051130 0.0000

Final State Root MSE z-Statistic Prob.

BETA1 0.582461 1.319767 0.441336 0.6590
BETA2 -0.487141 1.406621 -0.346320 0.7291
BETA3 -5.681428 0.000000 NA 0.0000

Log likelihood -72383.46 Akaike info criterion 39.02746
Parameters 32 Schwarz criterion 39.08109
Diffuse priors 0 Hannan-Quinn criter. 39.04655

I wasn't able to obtain same results into a second run and the warning error appear again.

Sspace: DIEBOLD_KALMAN_FILTER
Method: Maximum likelihood (Marquardt)
Date: 10/02/14 Time: 11:26
Sample: 7/19/1999 1/17/2014
Included observations: 3711
Failure to improve Likelihood after 6 iterations
WARNING: Singular covariance - coefficients are not unique

Coefficient Std. Error z-Statistic Prob.

C(1) 0.077873 NA NA NA
C(2) 0.883231 NA NA NA
C(3) 0.892654 NA NA NA
C(4) 0.894846 NA NA NA
C(5) 0.894302 NA NA NA
C(6) 0.888040 NA NA NA
C(7) 0.888380 NA NA NA
C(8) 0.893207 NA NA NA
C(9) 0.897714 NA NA NA
C(10) 0.900508 NA NA NA
C(11) 0.901803 NA NA NA
C(12) 0.902065 NA NA NA
C(13) 0.901677 NA NA NA
C(14) 0.900900 NA NA NA
C(15) 1.866118 NA NA NA
C(16) 0.628026 NA NA NA
C(17) 0.727110 NA NA NA
C(18) -0.861727 NA NA NA
C(19) 0.894171 NA NA NA
C(20) 1.019792 NA NA NA
C(21) -0.283466 NA NA NA
C(22) -0.114529 NA NA NA
C(23) -0.506432 NA NA NA
C(24) 0.911805 NA NA NA
C(25) -3.474517 NA NA NA
C(26) -0.457498 NA NA NA
C(27) -0.603912 NA NA NA
C(28) -0.491701 NA NA NA
C(29) 0.873802 NA NA NA
C(30) 1.360664 NA NA NA
C(31) 1.476205 NA NA NA
C(32) 1.076109 NA NA NA

Final State Root MSE z-Statistic Prob.

BETA1 0.655795 0.362550 1.808842 0.0705
BETA2 1.249283 1.253480 0.996652 0.3189
BETA3 -3.252186 0.000000 NA 0.0000

Log likelihood -70457.83 Akaike info criterion 37.98967
Parameters 32 Schwarz criterion 38.04330
Diffuse priors 0 Hannan-Quinn criter. 38.00875

Any advice?
Thanks

[SnakeS2k]

I defined better my parameters. But there is something in eviews that is not clear to me.
I try to explain it:

Step 1: set initial value for parameters
=> "convergence"
Step 2: try again estimate
=> "near singular Matrix"

If I put to zero c vector of parameters, where Eviews stores parameters calibrated in step 1 => convergence again.
It seems to me that after first calibration, Eviews starts, not from @param set in the code, but from parameters of the first calibration stored in "c".

Help please

[SnakeS2k]

Anyone?

[EViews Glenn]

I just double checked, and EViews 8 does use the parameters set in the @param statement.