A very pleasing result so far for my forward test of the Breakout Neuro EA V1_0. I first wrote about the Neural Network approach I'm using in August
A recap on how this EA works:
The base EA is a simple 4 bar breakout on a 15 minute chart with 100 pip stoploss and takeprofit. Not surprisingly this is miles from being profitable in its own right. A 19-10-1 neural network provides an entry filter which is very effective (at the moment) at entering swing trades in the long direction.
The NN was optimised using an early stopping approach: backpropogate against a factset from 6 to 3 months ago, and select the NN configuration which delivers the best profit/drawdown on data in the last 3 months.
So how come I haven't broken out the champagne? The EA is making money isn't it? The problem is that the approach of optimisation and run until failure, or regular optimisation, is not backtestable. So, I thought, why not incorporate regular optimisation into the EA itself?
So I started to laboriously code the NN backpropagation algorithm into MT4 ... and gave up. It's not impossible, but translating what was a neat bit of OOP into a C-style environment, with no debugging facility, was painful. I also realised how many loops and arrays are involved and became concerned that MT4 would run too slowly.
So, for now, the opposite approach. I'm in the middle of writing a C++ dll which will spoon-feed everything to the EA, including the entries. All the EA will need to do is send out the latest closed bar info every 15 minutes and ask if there is an entry. Stay tuned for the results - in a week or two.