The RATS Software Forum

1. I would like to calculate one step ahead out-of-sample forecasts using EGARCH. Is there any code/RPF that can help me to do this? Appreciate any help you can provide. Thank you.
2. I have used the egarchbootstrap.rfp code to generate n-step forward forecasts. Is there a way to calculate the root mean squared error (RMSE) out of these forecasted values?

Regards,
Monty

imansur wrote:1. I would like to calculate one step ahead out-of-sample forecasts using EGARCH. Is there any code/RPF that can help me to do this? Appreciate any help you can provide. Thank you.

Isn't the "H" series the one-step forecast of the variance?

imansur wrote: 2. I have used the egarchbootstrap.rfp code to generate n-step forward forecasts. Is there a way to calculate the root mean squared error (RMSE) out of these forecasted values?

RMSE of what?

1. My understanding is that H series is for in-sample forecasts, not out-of-sample.
2. I ran egarchbootstrap.rfp with normal and “t” distributions and obtained n-step ahead forecasts. I would like to compare forecast accuracy of these two distributions. How do I do that?

imansur wrote: 1. My understanding is that H series is for in-sample forecasts, not out-of-sample.

They are one-step ahead forecasts of the variance (calculation of variance at t given data through t-1).

imansur wrote: 2. I ran egarchbootstrap.rfp with normal and “t” distributions and obtained n-step ahead forecasts. I would like to compare forecast accuracy of these two distributions. How do I do that?

Forecast accuracy of what? Variance? If so, what is the actual against which you want to compare it?

Just to clarify – I can use the conditional variance H(t) as a one-step ahead out-of-sample forecast for H(t-1) even when I use EGARCH? Thanks

h(t) is the predicted variance for t given t-1 which I assume is what you need. The one-step forecast for EGARCH is closed form since it's a function of "observables" (lagged residuals, lagged variance). It's the multi-step forecasts that need bootstrapping or simulation because the lagged variance term in step 2 (for instance) isn't observable and can't be computed using the law of iterated expectations as you can with a standard GARCH model.

Make sense now. Thank you for clarifying this for me.