Switch Regression Model Problem

[startz]

The out of sample state probabilities aren't observed. What statistical test would you like to apply?

[nicotrader]

What do they need the probability if they can not be tested out of sample?

[startz]

What do they need the probability if they can not be tested out of sample?

I have no idea what you need the probability for.

[nicotrader]

Never heard trading rules?
What is the Markov Switch if not to develop the PROBABILITIES REGIME?!

[startz]

Never heard trading rules?
What is the Markov Switch if not to develop the PROBABILITIES REGIME?!

Actually, I have heard of trading rules. That doesn't mean I can read your mind.

If you have a specific statistical test in mind, say what it is and someone here may be able to help you.

[nicotrader]

What i want is a test OPERATING, derived from a test STATISTICAL.
So, i need to analyze the probability of regime change, in sample and out of sample.
If i can have the probability of state, only in sample, how can i verify the model's ability to capture the behavior of latent behavior?
Also how can i use the parameters of a model that I trusted on a different sample of data if I can not make an out of sample?

[EViews Gareth]

I think you may be better off hiring a statistician to provide some answers for you. This forum is for asking about how to do thing in EViews. It generally requires/assumes the user knows what they want to do statistically, and need to know how to do it in EViews. I think you're a few steps before that currently.

[nicotrader]

Sorry but i do not.
I'm here to ask why it is not possible to estimate the probability regime change, out of sample.
I am also here to ask how i can save the parameters of a model estimated to use it on a different sample.
That is, the base of the modeling ....

[trubador]

Because MS models do not work in that way. It really does not matter what you expect from a model. You can always study and learn the statistical properties of a model from a textbook. The forum is here to fill in the blanks or point you in the right direction to do so.

Your model has a fixed (time-invariant) transition probability matrix, so it will remain the same out-of-sample as well. If you like, you can retrieve the values as follows:

Code: Select all

youreq.maketransprobs tprob

I understand that it will not make much sense from a trading point of view. If you are interested in "capturing the latent behavior", then you need to find a set of independent variables that have an explanatory power over the probabilities of regime changes. This is called "time varying transition probability" model. Of course, you'll need future values of those independent variables to make out-of-sample predictions for transition probabilities.

[nicotrader]

for time invariant mean that the output probability is not different if use 1000 or 2000 observation of same total sample?

[trubador]

I am not sure what you mean by "output probability". In general, MS models have two types of probabilities (i.e. transition and regime). You seem to be interested in the transition probabilities, which define the probability of switching from regime i in period t-1 to regime j in period t. In general, these are time-invariate, meaning fixed over the "estimation" sample. If you estimate the model regarding the period 1 1000, then the transition matrix will be the same over the forecasting period 1001 2000 as well. Of course, you can always re-estimate the model by extending your data with adding the new(est) information flow.

[nicotrader]

No i am interested to Probability of regime.
if i estimate again the model with new and old data the probability of state change totally.

This rappresentation is useful in order to extrapolate concrete indications for operations in the financial trading.
For example, in the analysis of realized variance or Kernel volatility, it is supposed, that the model is able and statistically significant in describing those who are phases of high and low volatility. Not being able to test the probability of regime, with out of sample test, it is impossible to tell if the model is efficient.

[trubador]

Unfortunately, EViews does not produce out-of-sample forecast values of regime probabilities at the moment.

[startz]

Unfortunately, EViews does not produce out-of-sample forecast values of regime probabilities at the moment.

Even if EViews did, there would be no way to test them as the out-of-sample regimes are not observed. This fundamental issue has nothing to do with EViews.

Trubador gave the OP sound advice (as usual) above

Because MS models do not work in that way. It really does not matter what you expect from a model. You can always study and learn the statistical properties of a model from a textbook. The forum is here to fill in the blanks or point you in the right direction to do so.