0
$\begingroup$

I have a really basic question which I need clarifying on and would appreciate a quick response to this if possible.

I am forecasting conditional variance (volatility). I have in-sample from 1996-2007. Out of sample 2007-2016.

Do I report descriptive statistics for the in-sample or full sample?

I have my daily returns and will be using Automatic ARIMA forecasting on eviews to do my model selection for ARMA, this process uses information criteria to make the best selection (AIC, SIC). When I give the command for this, do I use the full sample or in-sample period?

After I get my AR and MA terms, I run an OLS, should I use the full sample or in-sample period for the OLS?......From this OLS I run residual diagnostics including an ARCH-LM test.

Next, I run my code for my GARCH. This code currently produces out of sample forecasts (2007-2016) for the conditional variance in a separate series. The code uses recursive rolling (an expanding window) to generate the forecasts. It is able to distinguish what is in-sample and out-0of-sample and recalculates and stores the coefficients to make the forecasts. Its perfect and does what I need. However the final GARCH equation which is saved shows outputs for the parameters for the full sample. I am trying to modify the code so I can run another estimation just for the in-sample period which will then allow me to report the in-sample parameters for the GARCH equation.

However, is this necessary? Do I have to report the parameters for the in-sample period, or can I report the parameters of the full sample?

As you can probably tell, the main issue comes in whether I am using the right sample, the techniques I have already and understand the whole process. This last step needs clarification.

$\endgroup$
2
$\begingroup$

Do I report descriptive statistics for the in-sample or full sample?

It depends on what you want to inform the reader about.

  • You might want to give him/her an impression of the whole dataset (not just the in-sample) to set a background for the problem you address.
  • You might want to motivate the choice of a model based on in-sample (not the whole sample) and test its performance out of sample.

<...> When I give the command for this, do I use the full sample or in-sample period?

If you are doing realistic out-of-sample testing where you hide the out-of-sample data at the model-building stage, you would only input the in-sample period.

After I get my AR and MA terms, I run an OLS, should I use the full sample or in-sample period for the OLS?......From this OLS I run residual diagnostics including an ARCH-LM test.

The same as above.

<...> However, is this necessary? Do I have to report the parameters for the in-sample period, or can I report the parameters of the full sample?

It depends on what do you want to address.

  • The model fit on the whole sample would be best suited for someone who wants to model or predict the actual future, as then all available data is welcome (assuming the data generating process has not changed over time and made the older data redundant).
  • The model fit on a particular training sample (a window) could be compared to the model fit on the whole sample to show how much variation there is between the full sample and a subsample.
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.