Estimating Multivariate TVTP Markov Switching model

[sohansust]

Dear All,
I am trying to estimate a multivariate markov switching model where the transition probabilities will be estimated with Filardo's (1998) procedure.
To more elaborate Filardo (1998) choose a information variables for transition probabilities. I want to use Composite leading Indication (CLI) as a information variable.
Note that I have Eviews 8 but here it estimates AR as a independent variable.
Thnaks

[EViews Glenn]

You'll have to be more specific about the Filardo paper to which you refer and the model in question.

[sohansust]

You'll have to be more specific about the Filardo paper to which you refer and the model in question.

Thanks for your reply.
I have used the explanatory variables on List of non-switching regressor and it worked.
thanks

[sohansust]

Dear Glenn,
Now I am facing the convergence problem.
I run the model once and found convergence, but when I run the same model the following day, it say' convergence not achieved after certain iteration'.
How would I solve the problem?
Please help.

[startz]

You probably have different starting values for the parameters the second time you run it.

[trubador]

Startz is right. By default, EViews uses random search for initial values. You can either choose "User supplied" option as the starting method or save/specify the "Seed" value of random number generator for further use. I'd like to remind you that the estimation of time varying transition probabilities requires a thorough examination of the model, since it is not easy to find a robust specification.

[sohansust]

Startz is right. By default, EViews uses random search for initial values. You can either choose "User supplied" option as the starting method or save/specify the "Seed" value of random number generator for further use. I'd like to remind you that the estimation of time varying transition probabilities requires a thorough examination of the model, since it is not easy to find a robust specification.

Thank Trubador and Startz,
But If I change the initial value, will be the estimated parameter be the same?
I tried and found a slight difference in parameter. Can you give me some reference, if possible, for the robust specification of TVTP model?
Thanks
Sohansust

[trubador]

This is an optimization model with too many local roots. So you do not have to worry about the differences in results as long as they are within the convergence tolerance. If you want to fix the solution, you can try the suggestions that I made in my previous post.

[mse]

Hello,

This is probably a stupid question, but I'd like to clarify how to set the User defined initial values of the parameters.
I try to estimate a markov switching model with two regimes, switching constant and sigma and a non-switching lag.
The output that I get is:

Code: Select all

Dependent Variable: PC               
Method: Switching Regression (Markov Switching)               
Date: 06/06/14   Time: 11:59               
Sample (adjusted): 1993M07 2013M10               
Included observations: 244 after adjustments               
Number of states: 2               
Initial probabilities obtained from ergodic solution               
Ordinary standard errors & covariance using numeric Hessian               
Convergence achieved after 20 iterations               
               
Variable    Coefficient    Std. Error    z-Statistic    Prob.
               
Regime 1               
               
C    1.225047    0.214160    5.720233    0.0000
LOG(SIGMA)    0.402751    0.076826    5.242350    0.0000
               
Regime 2               
               
C    -1.524567    0.322153    -4.732426    0.0000
LOG(SIGMA)    0.752415    0.079126    9.509030    0.0000
               
Common               
               
PC(-2)    0.868448    0.025987    33.41805    0.0000
               
Transition Matrix Parameters               
               
P11-C    2.810124    0.463121    6.067793    0.0000
P21-C    -2.375720    0.462287    -5.139060    0.0000
               
Mean dependent var    0.104078        S.D. dependent var        5.892272
S.E. of regression    1.948373        Sum squared resid        907.2817
Durbin-Watson stat    1.303704        Log likelihood        -507.5203
Akaike info criterion    4.217379        Schwarz criterion        4.317708
Hannan-Quinn criter.    4.257786         

Then, when I switch to Representations to know the names of the parameters I see that the constant and and the lag parameter have the same name.

Code: Select all

Estimation Command:
=========================
SWITCHREG(TYPE=MARKOV,HETERR,S) PC C @NV PC(-2)

Estimation Equation:
=========================
1: PC = C(1) + C(3)*PC(-2)

1: SIGMA = @EXP(C(2))

2: PC = C(3) + C(3)*PC(-2)

2: SIGMA = @EXP(C(4))

Substituted Coefficients:
=========================
1: PC = 1.22504665939 - 1.52456652777*PC(-2)

1: SIGMA = @EXP(0.402750911969)

2: PC = -1.52456652777 - 1.52456652777*PC(-2)

2: SIGMA = @EXP(0.752415464536)

I hope that this error is present in the representations window only.
Still, how should I call the parameters to initialize them?

Thanks

[EViews Glenn]

Looks like an error in the representations view. We'll take a look.

[edit] I'm not seeing the reported behavior in the representations view. Can you tell me the date on your version of EViews? And perhaps post the workfile with the equation in question.

Code: Select all

Estimation Equation:
=========================
1: G = C(1) + C(5)*G(-2)

1: SIGMA = @EXP(C(2))

2: G = C(3) + C(5)*G(-2)

2: SIGMA = @EXP(C(4))