Convergence of Markov Switching Models'estimations

[BHV]

Hi everybody!

I'm estimating MSM (in mean and in mean and variance) using Eviews 8 on quarterly GDP/GDPperCap data.
However as soon as I move from two to three regimes my coefficients do not have any Standard Error, nor Z Stat and Prob (all indicate NA).
This is puzzling as in many articles in the literature, specifications using 3 regimes are to be found.
Moreover when comparing the different specifications, the ones with the best BIC, AIC and HQC are the specifications with three regimes.

I tried changing the seed but without success.
If someone could give me a hint as to why I stumble over this.
(And I hope, this is not a subject that has already been discussed, I couldn't find any related discussion).

Best regards

[BHV]

I found another post on the same subject saying this might be coming from the data. (sorry for double posting).
However the previous thread line did not really answer the question .. =/

Moreover I tried estimating the model again today and suddenly it started "working" for some countries but not all of them. I'm thinking it might be due to the starting point in the estimation or something like that.
Do you have any idea how to specify the estimation in order to ensure a possible convergence ?

Best regards
B

[EViews Glenn]

What you describe most probably is a starting value issue.

As we explain in the manual, Markov Switching models are particularly difficult to estimate as identification is sometimes tenuous. This tendency toward estimation problems is exacerbated was you increase the number of regimes.

You can never ensure covergence, but you can improve the estimation behavior by taking advantage of our options for pre-iteration, randomized coefficient, and other tools for sensitivity. Increasing the number of pre-estimation random trials, and possibly the number of pre-estimation iterations, should help. Setting your own random seed value will ensure that the results are replicable.

[BHV]

Thank you for your answer !

I will try increasing the figures you mentioned. (Is there any upper limit ?)
Moreover, as I know the approximate values my probabilities should take, I wanted to know how (in the code) to specify them. (I am new to eviews so I am sorry if this questions appears as trivial).

Thanks again !!