Compare Goodness of Fit for OLS and Kalman Filter Model

[workhard]

Hi all,

I wish to prove that the State space model that allows time-varying coefficient is better than the OLS technique that give us static coefficient.
However, the OLS estimation give me better results in terms of AIC, BIC and Loglikelihood for all models.
I have read some papers saying that Kalman filter approach produce instability parameters estimation at the initial stage. The exclusion of first few observations can avoid any unfair bias due to these start-up problems. However, I have tried to recalculate the AIC, BIC and Loglikelihood after I exclude the first two years of my data (my data sample from year 1988 to 2009, weekly data). It seems like the AIC, BIC and Loglikelihood do not improved for Kalman filter models.
Any idea whether we should always get better goodness of fit for Kalman Filter compared to OLS?
Is it a good way to use these statistical goodness of fit to compare these two techniques?