That was my point: for model development data, there are no hard and fast rules. I have no idea what type of data or software tools you have available. Asking for specifics like this is like asking "How many bricks will I need to build a house?" I don't know, how many rooms do you want? Do you want a chimney? How about a garage? Testing data is another matter. Though this will vary depending on how precisely you need to assess model performance, in most cases only a few thousand cases will be needed for testing each model. Common rules such at "train on 70% and test on 30%" miss the mark because once the test data reaches a certain size, the precision of the performance metric increases very slowly and is probably pretty close to the true value. Some authors (Weiss, Indurkhya, Kulikowski, Sklansky, Wassel) make a strong theoretical case for something around 5,000 test cases. -Will Dwinnell Data Mining in MATLAB
John Ehlers has written a bunch of really good papers on signal processing as applied to trading systems. Check out his "Technical Papers" section. The papers that I found the most useful to me were: "Inferring Trading Strategies from Measured Probability Density Functions" "The Joy of Detrending" "Using the Fisher Transform" Ehlers develops all sorts of ideas that are applicable to the OP's question.
I have already responded to this post, but remembered afterward that I had written a quick guide to utilizing neural networks here: Family Recipe For Neural Networks -Will Dwinnell Data Mining in MATLAB
Thanks. Data Mining in MATLAB is mine, and I co-author Data Mining and Predictive Analytics with Dean Abbott. -Will
There are a number of publications regarding NN here: http://yann.lecun.com/exdb/publis/index.html specifically look at http://yann.lecun.com/exdb/publis/index.html#lecun-98b Unfortunately I still do not have a strict algorithm to train NN on my market data. I may train NN several times with different algorithms and get different results. I also have small experience with SVM, RVM but found them to be very slow. I did not manage to train over 8000 data with RVM algorithm. And SVM did not give me an improvement in comparison with NN. Do you have specific tricks regarding learning NN? I would like to exchange ideas. Thanks.
RedRat, what data are you training with exactly? Generally you'd be feeding it some kind of cooked data like RSI samples rather than direct price data. I've also messed with open/close patterns where the NN is fed a series of 1's or 0's representing if the close is higher than the open. There's a good bit of poo-pooing of neural networks but if a statistical edge exists in a data set it'll find it.
I am training on 5min time bars with know-how preprocessing. First of all you should map all your data into the range say [-1 .. +1] and better if you map it with linear probability. For example if you map price = 1000 to be +1 and price 500 to be 0.5 it will not work. RSI is the better representation of market data but I've never tried to feed it for NN. As for pooling, I believe if somebody has the profitable algorithm he would not share it . So I suggest to exchange specific tricks which you can try to improve generalization and predictability.