<Len, I know you'll read this. So, am I losing it? > I think we're both losing it for being up this late... just kidding(?). I think that conceptually you're theory is correct if you you do additional training once you update your net, but my best guess is that if when you update the model, you simply hit "predict", the program will give a prediction based on the parameters of your previosly trained net. Although the "window" will slide, I don't believe that the net has assimilated the new data at that point. There have been times when I just hit predict for months at a time with certain nets and when I looked at the scattergraphs they are no longer the crisp diagonals they were when I started out. Additional training at this point for an hour or so restores the integrity of the graphs and the nets and I think at that point you have the new net you're talking about. I retrain my nets on the weekends but use the batch files to make predictions during the week to avoid the situation you're talking about.
As you know I wrote to Andrew about Neuro's tendency to change its mind. This question was asked when the general market was in a downward trend a few months ago. My question and his responses follows:
A quick question. I'm trading a stock that Neuro gave a buy on 4 days ago, 75% confidence. I have a 4 month verify period set, and it appears to have a flawless record. Anyway, the stock been dropping (no surprise given the current market outlook) Anyway, Neuro appears to be "cheating". The first two days of the buy signal, have changed to hold. Why is this happening, and how can one have confidence in the signals, if Neuro can change its mind?
And Andrews response:
Did you train the network after the buy signal? NeuroStock does not record its past signals, it recalculates the signals on the fly from the current data and neural training. So, if you trained the network after the buy signal it will use the newly trained network to display all past signals. The newly trained network did not issue buy a signal where the older network did.
We ran extensive testing to insure that NS does not see the data in the verification period, so whatever is happening it is not NS peeking at the verification data and changing its mind.
Now, why did NS change from buy to hold signal? there are several
possibilities: since the buy signal (4 days ago) 4 data samples have moved from the verification period to the training period, and 4 days moved out of the training period; this new data might have altered NS'network in a way that affected the pattern interpretation used for the original buy signal. The more aggressive training strategies use very sensitive pattern match filters to issue as an early signal as possible; training the network affects the filters parameters (that's how they get the parameters to begin with) and because they are very sensitive they are more likely to change with additional training.
A neural network such as yours (60 neurons) contains hundreds of pattern matching filters, during training each filter seeks and learns a particular pattern from the data files. Furthermore, there are
additional neurons learning which patterns are significant and which ones are not, even learning the relationships between patterns (i.e., pattern of patterns.)
One fact that a lot of the neural books don't mention is that among the hundreds of filters there are always a few that have not found any good patterns to latch-on; but, once in a while during training these floating filters find a pattern that was not previously recognized and latch on to it, the new knowledge is then fused with that of the other filters before a final prediction is made. The end result is that the network will be a bit smarter and it may issue different signals for the same set of inputs.
I hope that this helped you, I tried to keep as much of the
techno-babble out of the documentation and user interface in order to keep the program simple and user friendly.
Andrew Cilia,
NeuroStock,
Jay, it seems from his response that by not training the net after updating the files will solve your delemna. Just hit the predict button.
I hope this helped, now go to bed!!!! |
