NA values in Markov Switching regressions

[hbee9211]

Hello,

I am using Eviews student version and trying to run Markov regime switching regressions on daily data for different companies from great depression era. So far I have limited the regressions to 2 regimes for each company. However, for any given 'AR()' terms (I start with 5 (representing a 'week') and reduce it by one term to see the effect of prices of previous days), the results sometimes yield 'N/A' for values of 'coefficient', 'std error', 'z statistic' & 'probability' for each regime. I have to re-run the switching regression for a different seed each time till I get values for these variables. I have tried changing the options in 'Initial Regime Probabilities', 'Start Method' , 'Generator', & 'Randomized Estimates' tabs, but it is of no help.

[EViews Glenn]

Without looking at the data it's difficult to say what will help. Even more extensive use of the randomization tools is, I think, probably the best thing for you to do. I will point out that what you are experiencing is not uncommon for Markov models which can be very difficult to estimate.

[hbee9211]

Let me know how can I send you the data. It is time series of daily data. Following is an example of one such result:

Dependent Variable: REY
Method: Switching Regression (Markov Switching)
Date: 07/12/14 Time: 00:58
Sample (adjusted): 2 2367
Included observations: 2366 after adjustments
Number of states: 2
Initial probabilities obtained from ergodic solution
Ordinary standard errors & covariance using numeric Hessian
Random search: 25 starting values with 10 iterations using 1 standard
deviation (rng=kn, seed=1832293665)
Failure to improve objective (non-zero gradients) after 9 iterations
WARNING: Singular covariance - coefficients are not unique

Variable Coefficient Std. Error z-Statistic Prob.

Regime 1

C 132.6174 NA NA NA

Regime 2

C 31.96534 NA NA NA

Common

AR(1) 0.997970 NA NA NA
LOG(SIGMA) 0.135700 NA NA NA

Transition Matrix Parameters

P11-C 8.042026 NA NA NA
P21-C -21.11086 NA NA NA

Mean dependent var 80.79420 S.D. dependent var 45.23749
S.E. of regression 2.371069 Sum squared resid 13279.09
Durbin-Watson stat 1.999348 Log likelihood -3699.495
Akaike info criterion 3.132286 Schwarz criterion 3.146916
Hannan-Quinn criter. 3.137612

Inverted AR Roots 1.00

[EViews Glenn]

Post the workfile here or send it to support@eviews.com along with your serial number, build-date and reference this thread.

[RFJ]

Hello,

I'm sorry to divert the discussion to ask a rather silly question, but I'm a newbie here. I'm trying to estimate a two-state Markov model with time varying transition probabilities. How do you estimate a model with the state variable depending on its previous lags? The estimation window allows for using constant probability matrix under 'c' or any other variable included in the workfile, but I can't figure out how to estimate the model with the lagged state variable as a probability regressor.

Again, sorry for the silly question and thanks!

[EViews Glenn]

Do you want to use the unobserved state or the value of the dependent variable?

[RFJ]

I want to use a lag of the unobserved state variable as a probability regressor as in Jeanne and Masson, 2000

[EViews Glenn]

I'm not sure how one uses an unobserved variable as a probability regressor (or as any kind of regressor). It is, after all, unobserved.

I was not familiar with the Jeanne and Masson (2000, JIE) paper. My quick (under 5 minute) look indicates that they estimate a Markov switching model intercept model for a devaluation probability in terms of other economic variables.There is nothing non-standard about this specification. Focusing on p. 341 ff, the devaluation probability is an observed variable obtained prior to estimation and I see no evidence that the 2 x 2 matrix of transition probabilities (Theta) is specified as time-varying.

I must admit that I skimmed the first part of the paper so I didn't work out the theoretical model, but unless I'm misunderstanding things, there is no unobserved variable as a probability regressor in this paper.