switchreg, mean of forecasted variable

[mse]

Hello,

I'm using Eviews 8 to estimate a MS-AR(2) model with a switching constant and variance on centered data:

yt=c_st+phi*yt(-2) +epsilon_st

The estimation works fine, but when I check the mean of the adjusted explained variable, I get something different form 0 (0.21).
It seems strange that the model is not able to capture the mean of the dependent variable.
Is there something wrong with the specification, or with the algorithm?

Thanks a lot in advance.
Please find attached my workfile.

[trubador]

The model output states that "Mean dependent var is 0.247063". As you include lagged terms in the model, sample period changes. The mean of centered dependent variable is no longer 0. You'll see that the estimation period is "1993m10 2007m06" and @mean(y,"1993m10 2007m06") should give you the correct average for comparison purposes.

As an aside, I'd like to tell you that the estimated model is not a MS-AR(2) model, it is a type of Markov Switching Dynamic Regression model, MS-DR(2).

[mse]

Indeed, thanks a lot.

[mse]

I would like to ask another question about the mean as I'm still a bit confused...

When I calculate the unconditional exectation of series y following such a MS-DR(2) process, I get

E(y)=E(c_st)/(1-phi), where

E(c_st)=c0*P(S=0)+c1*P(S=1) = (c1-c0)*P(S=1) + c0 = (c1-c0)*(1-p_00)/(2-p_11-p_00) +c0
p_ij is the transition probability

For the results in the workfile above I get E(y)=1 approximately, whereas the mean of the adjusted y is 0.22, which seems to be a significant difference.
Where can in come from?

Thank you.

[trubador]

In your calculation you use fixed transition probabilities. However, fitted values are computed via smoothed regime probabilities, which changes over time. Since each regime has an assigned probability, the result would be more-or-less equal to mean of dependent variable. In short, two calculations are essentially different and therefore would yield different results.