Out of Sample Forecasting for ARIMA(0,1,1)-Garch(1,3)

[vsroy08111991]

Hi all,

I need to know how to get out of sample forecasts for ARIMA(0,1,1)-Garch(1,3) process. I have been using Eviews to analyse bond yield univariate series. The above specification proved to be the best. I also forecasted within the sample I had. Now I want to forecast for future. To be clear, my data set is monthly yield for the period Jan 2000 - April 2015. Now I would like to forecast for (say) May 2015 to June 2015. I realize that since the regressors (here MA(1), ARch and Garch variables) are not present up to June 2015, I am not getting any forecast for that period. But is there a way to estimate them and then forecast the bond yields?
Please be as elaborate as possible and if possible avoid codes as I am not comfortable with them. (still, if it can solve the problem, please go ahead) 

I am also sharing the set of estimation equations so that you can check them  as well to understand and help me in better way.

Thank you.

[EViews Gareth]

You need to extend the range of your workfile. Currently your workfile only has observations up until June 2015. If you wish to forecast beyond that, you'll need to have observations in your workfile for the periods you want to forecast.

To do that, from the workfile click on Proc->Structure/Resize Current Page, and then change the end date to the date you want.

[vsroy08111991]

You need to extend the range of your workfile. Currently your workfile only has observations up until June 2015. If you wish to forecast beyond that, you'll need to have observations in your workfile for the periods you want to forecast.

To do that, from the workfile click on Proc->Structure/Resize Current Page, and then change the end date to the date you want.

I want to forecast for the future when I don't have actual data to compare with my forecasts. It is like suppose someone trades in bonds and would like to predict what he bond yields would be like in the market for the next two periods. He does not have data to cross check. Is this not possible in any way?

Is Eviews just built to do ex post forecasting? I just want future expected values which follow my arima(0,1,1)-garch(1,3) specification? If this is not possible then there is no business prospect of Eviews as businesses always want to know (with certain degree of certainty) what lies in the future.

[EViews Gareth]

Did you actually try what I wrote?

[vsroy08111991]

Did you actually try what I wrote?

I did try that out... Check the update workfile.
But when you say that I need to have observations for the period I want to forecast, i.e., for the period May 2015 - (say) November 2015, they are in the future and I don't have observations for them. The series I am working with has "NA" in the cells for values of these periods as I have no idea what they are going to be and hence, my need to forecast.
If I insert any value in these cells, then that will just be a shot in the dark.

[EViews Gareth]

Having observations just means having rows - it doesn't mean you actually need data in those rows.

Once you have restructured the workfile to have those observations, you use the forecast button on your equation.

[vsroy08111991]

Having observations just means having rows - it doesn't mean you actually need data in those rows.

Once you have restructured the workfile to have those observations, you use the forecast button on your equation.

I tried what you mentioned. I increased the length of the series by 6 periods and used the forecast button on my equation but it did not give me any forecast. The forecast (saved in default as variablenamef) does not have any values in the last six rows which are out of sample. It has "NA" in those rows. The forecast result also shows the number of included observations as zero.
I have tried all the options that come in forecast dialog box. Where am I going wrong?

To elaborate once more, these are the steps that I did:
I have data from Jan 2000 to April 2015.
I estimated the ARIMA(2,1,1_-EGARCH(1,3) and saved the equation.
I increased the series length upto to October 2015 (2015m10).
Then I used the forecast button and set range as 2015m05 (May2015) - 2015m10 (Oct2015) under dynamic forecasting and MA backcast of estimation period.

Got an error message saying forecast can't be done because of missing values

So, I included a few periods of the actual data in the forecast period as well.
I got the forecast of only upto the sample end date i.e. April 2015.
I then tried other options in the forecast dialog box as well, but still it hasn't worked.

What should I do?

I am using Eviews 7.

I know I am making you go over and over the same thing but I must be missing something.

[EViews Gareth]

Ah. I just noticed the exact specification of your equation.

Essentially you have said that the dependent variable (Y for brevity's sake) depends upon D(Y). i.e. today's Y depends upon the difference between today's Y and yesterday's Y. Tomorrow's Y depends upon the difference between tomorrow's Y and today's Y. The only way we can calculate the value for tomorrow's Y is by knowing what the value for the difference is. But we have no way of knowing what the value of the difference is without knowing what tomorrow's value is. It is a circular problem.

What you're trying to do is mathematically impossible.

Most similarly specified equations would have a lag on the difference. i.e. Y depends upon D(Y(-1)).

[vsroy08111991]

Ah. I just noticed the exact specification of your equation.

Essentially you have said that the dependent variable (Y for brevity's sake) depends upon D(Y). i.e. today's Y depends upon the difference between today's Y and yesterday's Y. Tomorrow's Y depends upon the difference between tomorrow's Y and today's Y. The only way we can calculate the value for tomorrow's Y is by knowing what the value for the difference is. But we have no way of knowing what the value of the difference is without knowing what tomorrow's value is. It is a circular problem.

What you're trying to do is mathematically impossible.

Most similarly specified equations would have a lag on the difference. i.e. Y depends upon D(Y(-1)).

Thank you this solves the problem. I am finally getting predictions for my model. Thanks a ton!!