EViews.com

We are having problems in the input specification. Our dependent variable is employment rate and independent variable is business process outsourcing (bpo) revenues (contribution to gdp growth). Both variables are I(1), data is quarterly. Here are our questions:
1. Are we right in inputting them to the converted dlog(variable)*100. For dependent variable: dlogemployment c, for non-switching: ar(1) ar(2) ar(3) ar(4) and for probability regressors: dlogbpo. If not, what is your suggested specification? Our professor did not discuss this topic that's why we are self-studying it.
2.Moreover, how do we interpret the transition matrix parameters - the first output that eviews will show
3. How about the tvtp table outputted? Is it simply the probability of the variable being in that low/high regime?

We have already read a lot of tutorials and papers using the procedure, but the lesson remained vague to us. We would be very very happy and grateful then if you could help us here. Thank you very much!

As to 1. that's really up to you.

As to the other questions, it will help facilitate discussion if your workfile contains the results you'd wish to discuss.

For number 1, I just copied the Filardo example provided in Eviews website (the one with the title Time-Varying Transitions) but did not really understand why the conversion is needed. Before doing Markov chain we have done preliminaries like unit root test (all variables are i(1)), we also did VAR and IRF.
Here are the outputs (number 1 specification but probability regressors is: c dlogbpo(-1))
Dependent Variable: DLOGEMPLOYMENT
Method: Switching Regression (Markov Switching)
Date: 06/03/15 Time: 05:56
Sample (adjusted): 1999Q3 2014Q4
Included observations: 62 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=1513392891)
Convergence achieved after 24 iterations

Variable Coefficient Std. Error z-Statistic Prob.

Regime 1

C 1.741799 0.153182 11.37081 0.0000

Regime 2

C -1.653990 0.158657 -10.42494 0.0000

Common

AR(1) -1.073105 0.143104 -7.498795 0.0000
AR(2) -0.981948 0.150152 -6.539705 0.0000
AR(3) -0.935431 0.143409 -6.522816 0.0000
AR(4) -0.082335 0.130168 -0.632531 0.5270
LOG(SIGMA) -0.379627 0.099063 -3.832176 0.0001

Transition Matrix Parameters

P11-C -2.517707 1.032917 -2.437473 0.0148
P11-DLOGBPO(-1) -0.013859 0.080978 -0.171142 0.8641
P21-C 2.819470 1.089539 2.587765 0.0097
P21-DLOGBPO(-1) -0.010724 0.076108 -0.140908 0.8879

Mean dependent var 0.102722 S.D. dependent var 1.707459
S.E. of regression 1.189563 Sum squared resid 79.24335
Durbin-Watson stat 1.425896 Log likelihood -80.00416
Akaike info criterion 2.889021 Schwarz criterion 3.263219
Hannan-Quinn criter. 3.036195

Inverted AR Roots -.00+.94i -.00-.94i -.10 -.97

Which ones are needed to be interpret? And can you list the interpretations?

This is the 2nd output.
Equation: UNTITLED
Date: 06/03/15 Time: 06:04
Transition summary: Time-varying Markov transition
probabilities and expected durations
Sample (adjusted): 1999Q3 2014Q4
Included observations: 62 after adjustments

Time-varying transition probabilities:
P(i, k) = P(s(t) = k | s(t-1) = i)
(row = i / column = j)
1 2
Mean 1 0.067148 0.932852
2 0.938162 0.061838

1 2
Std. Dev. 1 0.008880 0.008880
2 0.006563 0.006563


Time-varying expected durations:

1 2
Mean 1.072078 1.065965
Std. Dev. 0.010221 0.007488

Are the values 0.0671480 and .061838 simply the probability of the variable being in that low/high regime? How do we explain the other values here?
Also, I remember last time I have a much higher value here, around 0.9 instead of this 0.6 but I have not save my work so I started all over again. Is this because of the seed.

I have also provided the graph of filtered regime probabilities (both 1 and 2 - so both high and low regime) in the workfile attached

The transition probability coefficients are logit coefficients for the specified probability (see the equation for the transition probabilities in the Background section of the manual). P11 represents the probability of staying in regime 1 given that you are in regime 1. P21 are the coefficients for the transition from regime 2 into regime 1.

The table below represents the descriptive statistics for those fitted transition probabilities for the estimation sample. In addition to the .067148 for Regime 1 into 1, there is the corresponding .93282 probability for transitioning from Regime 1 into Regime 2. The fact that you are seeing different results when you estimate again is described in detail in the manual. Briefly, Markov switching results are often sensitive to starting values and can change dramatically depending on where you begin your estimation.

hi
could you tell me that why we must use c in dependent variale ?
if we use tvtp , is it trusted ????
please answer to my question
t is very important for me

thanks so much

I'm sorry but I don't understand the questions.

Dear Eviews Glenn,

I am working on a TVP- markov Regime switching for crude oil futures basis and I am interested to know if inventories can explain the regime switches from backwardation to contango and vice-versa. The estimated parameter results regarding the transition matrix parameters,( I mean P11-C, P11-CHLOGINV, P21-C, P21-CHLOGINV) are only specified for the slope coefficient related to the switch from regime 1 to regime 2 (P21-CHLOGINV). However, I want to know the slope coefficient for switch from regime 2 to 1 and its significance as well. I have searched the forum and I know about the logit relationship between the probabilities and the estimated parameters, but, what is my problem is the magnitude for P12-CHLOGINV and it's significance based on related statistics.
I am really wondering if there is a method in eviews to calculate the mentioned coefficients or not?

Kind Regards,

If you know the probability of staying in regime 1, then 1 minus that probability is the probability of switching from 1 to 2.

The coefficients only have meaning with respect to the logit relationship, but in some sense, you could have normalized the 1 transition to 1 relationship and reversed the sign on those coefficients to get the coefficients for transitioning from 1 to 2.

Dear Glenn,

An interesting point for me, Thanks a lot for your help.

Warmest Regards,

For details on how to think about the coefficients, I'd recommend looking at any book that describes logit analysis.

Hi,
I need your help to know how to test the hypothesis of change of regime for this model:

Rit = alpha t + beta1 (RM-Rf)t + beta2 DEFt + beta 3 TERMt + et
Rit = return of bond (i) at time (t)
RM-Rf, TERM and DEF are the risk factors.

In my research, I let the data tell me whether there is a set of times when betas and alpha are different than in other times.
So I estimate a Markov Switching regime model allowing alpha, beta1, beta2, beta3 vary between 2 regimes.
For this, I tested the hypothesis of change of regime by carrying the Wald test H0: c(1)=c(5), c(2)=c(6), c(3)=c(7), c(4)=c(8).
My question : Is this the right test for the hypothesis of change of regime.
I know the other tests for heteroskdasticity and number of regime
Waiting impatiently for your answer
Thank you
Ella

Unfortunately, testing for the presence of multiple regimes is quite difficult and the standard test you suggest does not work. See "Markov Regime-Switching Tests: Asymptotic Critical Values," by Carter and Steigerwald, Journal of Econometric Methods, 2013.

Thank you Startz for your reply but there is no multiple regime, there is only 2 regime: one stress regime where the 3 betas become significative and high. I don't think that this too difficult.
The tests that I made are :
I began by assuming the existence of 2 regimes
1/ Wald test for hypothesis of switch regime H0: c(1)=c(5), c(2)=c(6), ...
2/ test of Arch effect (Q statistics for square residuals,Q statistics for residuals,)

Then, I tried those tests with 3 regimes.

After thoses tests, I found good results I found that the model that suits the first portfolio is the markov switching (2 regimes) with heteroscedasticity of residuals (betas more significative in stress regime, pii are signifcatives, both p11 and p22 >90%)

But I started to doubt the 1st test (Wald) with the second portfolio where the probability of p11 (p-value > 5%) is much less than p22 (11%<<<97%) and smoothed prob (regime 1: stress) is often = 0,00002 That means that there is no regime switch for the second portfolio?

2 regimes is "multiple." The Wald tests are not valid.

In this case, what is the right test to verify the hypothesis of the existence of 2 regimes?