Dear all,
I would like to make forecasting estimates (in-sample and out-of-sample) from markov-switching estimates (fixed and time varying probabilities).
I would like to understand what eviews can do. It seems we can only forecast the regime probabilities (in sample) without s.e. Moreover, the forecasted probabilities are not bounded between 0-1.
Thank you very much.
Forecast Markov Switching
Moderators: EViews Gareth, EViews Moderator
-
- EViews Developer
- Posts: 2672
- Joined: Wed Oct 15, 2008 9:17 am
Re: Forecast Markov Switching
EViews computes in-sample one-step ahead, filtered, and smoothed probabilities values.
EViews doesn't compute the one-step ahead out of-sample forecasts of the probabilities since the one-step ahead predicted is really only defined for the first period post-forecast as there is no out-of-sample filtering step. Now one can compute a dynamic n-step ahead predicted probability, in which the filtering step is by-passed out-of-sample. EViews doesn't provide these probabilities to the user right now, but I am not sure if this is what you want. Note also that these for time-varying probability models, the probability regressors would have to be available for the forecast-sample.
For many cases, these dynamic n-step ahead predicted probabilities can be generated from existing results using simple commands.
Note that all of the computed in-sample probabilities should lie between 0 and 1. If they don't in a particular case, I'd be interested in seeing the example. Standard errors are currently not computed. They are quite difficult for these models and would almost certainly be computed using re-sampling techniques, but this is not something we have implemented at this time.
EViews doesn't compute the one-step ahead out of-sample forecasts of the probabilities since the one-step ahead predicted is really only defined for the first period post-forecast as there is no out-of-sample filtering step. Now one can compute a dynamic n-step ahead predicted probability, in which the filtering step is by-passed out-of-sample. EViews doesn't provide these probabilities to the user right now, but I am not sure if this is what you want. Note also that these for time-varying probability models, the probability regressors would have to be available for the forecast-sample.
For many cases, these dynamic n-step ahead predicted probabilities can be generated from existing results using simple commands.
Note that all of the computed in-sample probabilities should lie between 0 and 1. If they don't in a particular case, I'd be interested in seeing the example. Standard errors are currently not computed. They are quite difficult for these models and would almost certainly be computed using re-sampling techniques, but this is not something we have implemented at this time.
Re: Forecast Markov Switching
Thank you very much for your response.
The probabilities don't lie between 0 and 1.
Is there a way to rescale the proba ? I will need them to compute the forecasted value of the dependent variable.
I have sent you my workfile.
Thank you very much
The probabilities don't lie between 0 and 1.
Is there a way to rescale the proba ? I will need them to compute the forecasted value of the dependent variable.
I have sent you my workfile.
Thank you very much
Last edited by Aktar on Fri May 27, 2022 3:46 am, edited 2 times in total.
Re: Forecast Markov Switching
Finally, I guess I can use the standard command from the equation object as "view" and then "regime results" ; "regime probabilities" to get the one step head predicted probabilities because there are the same as with the forecast command but with the good scale this time.
However, there is a mistake when computing the errors from MS objects. Seems that the software makes the difference between the actual y value and the predicted probabilities (the one with the wrong scale).
Could you let me know ?
Thanks
EDIT: the probabilities are finally used as weights and the predicted value of y is easily found. This matches well with the residuals given by eviews
However, there is a mistake when computing the errors from MS objects. Seems that the software makes the difference between the actual y value and the predicted probabilities (the one with the wrong scale).
Could you let me know ?
Thanks
EDIT: the probabilities are finally used as weights and the predicted value of y is easily found. This matches well with the residuals given by eviews
-
- EViews Developer
- Posts: 2672
- Joined: Wed Oct 15, 2008 9:17 am
Re: Forecast Markov Switching
Where is the workfile?
Who is online
Users browsing this forum: No registered users and 41 guests