I try to understand how to forecast with the rugarch package, especially the command [ugarchforecast][1].
I have problems with understanding the n.roll and the out.sample command. Can someone explain this exactly?
The n.ahead command is clear to me. It says how much steps I want to forecast ahead. If it is more than 1 period, then the previous data is not sufficient anymore and the method uses the forecast value of the one step ahead forecast, to calculate the two step ahead forecast. Ok I understand that.
The ability to roll the forecast 1 step at a time is implemented with
the n.roll
argument which controls how many times to roll the n.ahead forecast. The default argument
of n.roll = 0 denotes no rolling and returns the standard n.ahead forecast. Critically,
since n.roll depends on data being available from which to base the rolling forecast, the
ugarchfit function needs to be called with the argument out.sample being at least as
large as the n.roll argument, or in the case of a specification being used instead of a
fit object, the out.sample argument directly in the forecast function.
I do not want to use a rolling approach, so I do NOT want to reestimate the model every time step, so I do not understand what n.roll and out.sample gives me? What does it mean to "roll the forecast 1 step"? I mean, if I do n.ahead=1 forecasts, then I have constant parameters, since I do not reestimate the model every time step.. and then I do simple one step ahead forecasts. So I use one new observation to do the forecast.
[1]: http://www.inside-r.org/packages/cran/rugarch/docs/ugarchforecast
Best Answer
Ok, I found it out:
n.roll gives you the number of the n-ahead forecasts. So you roll the window by one observation and you do the next n-ahead forecast. Out.sample is used to define the observations which should be excluded in the estimation procedure (if used in ugarchfit) to do out of sample forecasting afterwards.
It means, you go to the next period and the the next n-ahead forecast, as I wrot above.
You may want to use fitted() and sigma(). These are the fitted values and due to the lagged character of GARCH this is already so to say a "forecast".