Lagged dependant variable in state-space

[ALFQZ]

Hi everyone,

I was hoping you could give me a few pointers in making the most out of my work so far. Basically I have...

@signal spainbonds = sv1 + sv2*spainbonds(-1) + sv3*spaincds(-1) + [var = exp(c(1))]

I'd like to see if the coefficient SV3 exceeds SV2 at any point in time, that is that log bond prices are led more by log CDS than lagged bond values. My graphs appear to show that it does twice. As you can see I am using the random walk model with recursive c. Apart from that everything else is just default settings.

I'd like to know how to improve my model. I am having the singular covariance problem. Firstly, is there anything in particular I should do since I have a lagged value of the dependant variable in my state equation?

Also, I have just assumed the starting values as the default. I've done a bit of searching on here and it seems I should try estimating by OLS and trying those figures as starting values which may cure my singular covariance message? So do I just go to Quick > Estimate equation (by OLS) and type in spainbonds spainbonds(-1) spaincds , then use the coefficient values produced from there?

My estimation output is as follows:

Sspace: SPAINTEST
Method: Maximum likelihood (Marquardt)
Date: 29/08/10 Time: 17:18
Sample: 10/08/2007 10/08/2010
Included observations: 783
Convergence achieved after 4 iterations
WARNING: Singular covariance - coefficients are not unique


Coefficient Std. Error z-Statistic Prob.


C(1) -46.50791 NA NA NA
C(2) -48.95228 NA NA NA
C(3) -7.347877 NA NA NA


Final State Root MSE z-Statistic Prob.


SV1 0.105769 0.009584 11.03656 0.0000
SV2 0.358188 0.034798 10.29329 0.0000
SV3 0.463744 0.034523 13.43289 0.0000


Log likelihood 644.0335 Akaike info criterion -1.637378
Parameters 3 Schwarz criterion -1.619512
Diffuse priors 3 Hannan-Quinn criter. -1.630508

My results here don't look too wild (or do they?), although i've already had to change the start date as the way i've constructed the model did NOT deal well with volatility around the start period.

Any input is appreciated, apologies if i've made any basic mistakes.

Thanks

[trubador]

There is a trade off between the flexibility and computational feasibility in State Space models. Changing the starting values or proper initialization of state variables would help to a certain degree. However, correct specification of the model matters the most. If your model is ill-defined, you will not be able to obtain a feasible solution. I suggest you to try alternative model specifications along with the usual remedies that have been mentioned elsewhere in the forum.