Hi, I'm hoping to use the fSeries GARCH modeling to perform multi-period predictions. However, the predict.fGARCH() function seems to be pretty sparsely documented. It seems like it's only able to predict from the end of the training data onwards, not based on new data that I provide. For example, if I train it on daily data from, say, 1/1/1990 - 12/31/2005, calling predict() will give me predictions starting at 1/1/2006, with (if I'm reading the source right) all subsequent error terms assumed to be zero. I'd like to be able to pass it an array of new data to predict over, so that I can use the model to predict, say, February of 2006 or June of 2000. This seems to be possible if you use the tSeries library (via the newdata parameter), but not with the fSeries library (which I need to use, since garchFit() is more robust than garch()). Besides writing my own GARCH predictor, is there any way to accomplish this with the provided code? This seems like a very common thing to do, but I can't seem to find it anywhere. Any help would be greatly appreciated. Thanks, -Mike
fSeries GARCH Prediction Questions
2 messages · Mike Kocurek, Brian G. Peterson
Mike Kocurek wrote:
I'm hoping to use the fSeries GARCH modeling to perform multi-period predictions. However, the predict.fGARCH() function seems to be pretty sparsely documented. It seems like it's only able to predict from the end of the training data onwards, not based on new data that I provide. For example, if I train it on daily data from, say, 1/1/1990 - 12/31/2005, calling predict() will give me predictions starting at 1/1/2006, with (if I'm reading the source right) all subsequent error terms assumed to be zero. I'd like to be able to pass it an array of new data to predict over, so that I can use the model to predict, say, February of 2006 or June of 2000. This seems to be possible if you use the tSeries library (via the newdata parameter), but not with the fSeries library (which I need to use, since garchFit() is more robust than garch()). Besides writing my own GARCH predictor, is there any way to accomplish this with the provided code? This seems like a very common thing to do, but I can't seem to find it anywhere. Any help would be greatly appreciated.
It would be typical to train the model over a rolling window, and make a constant 'n' step ahead prediction, hopefully learning from the prior history. As in real life, you 'in sample' period continues to grow, and you can see how well the model performs 'out of sample' 'n' steps ahead. Try rollapply or one of its cousins, with whatever appropriate window. Regards, - Brian