Let us say I have a time series I want to forecast. I have decided I am going to use ppr, nnet and svm. I calibrate my model with these 3 algorithms (let us assume I use all the data for the calibration; I do not distinguish here between training and testing: all data are used for training, just for simplification of the exposition). Now, I get a new data point for the time series. I use the predict method with the 3 calibrations. This is fast. I get a new data point. I use the predict method. And so on. But at some point in time, I want to recalibrate my 3 algorithms. This takes time. What I would like is the following:
Let us assume I have 1000 data points. I calibrate my algorithms to this data set. Let us assume that I get a new data point every minute, and let us assume it takes half an hour for the calibration of one of the algorithms. What I want is that for the first 30 data points, I am using the "old" predict. Then, after this point, I start a calibration with 1030 data points. But this will take a long time (until I have acquired 1060 points). What I want is to use the old predict (the calibration when there were 1000 data points). When the calibration with the 1060 points is over and I get the data point No. 1061, I would like that automatically, when I call the predict method, the new predict is called, not the old one.
Is there an easy way to do this? Of course, the ideal would be to have the calibration in one computer and the predict in another, but I guess that this adds too much complexity to the mix ...
This archive was generated by hypermail 2.1.8 : Fri 02 Dec 2005 - 02:26:17 EST