Sunday, August 7, 2011

Waking up to Neural Network committees

Time to wake up from my blogging slumber and write a few more posts.

A huge amount has happened in my trading life in the well over 12 months since my last post.  I finally completed the early stopping Neural Network library described in my last post of June 2010, wrote an EA or two and played around with it for ages without success.  I also tried the Nearest Neighbour algorithm, (see a subsequent post that I have not yet written but intend to), and became diverted onto many other ideas.

Then … back to the Neural Network idea: how about a committee of Neural Networks?  I had been reluctant to try this purely because of the processing time: as it is, the EA that I have written can take several hours to run a couple of years’ single backtest.

Well, it seems that this may have some promise.  You train a group of NNs with different topologies and parameters, then the simplest method of obtaining a result is to enter a trade when they all agree.  It is a well-established approach in the NN literature.  I’ve tried up to three NNs, although watching grass grow has now become a new fascinating pastime: some of the arrangements I have tried have taken a couple of days to backtest.  Just as with all NN techniques, this is far more art than science: what is a sufficiently different-looking topology to allow the NN to produce a result that is sufficiently independent of its peers?

Attached is a backtest from 1/1/2008 to present for a NN committee of two, trained with a factset of 6/12 EMA cross and both SL and TP set to 100