The showcase pages over the last few months have been geared mainly towards the discretionary trader. Although we prefer the discretionary approach, systematic trading does have some advantages that should not be overlooked. Aside from the reduced stress associated with a 100% objective method, systems traders can backtest their systems to produce reports that would take many hours for a discretionary trader to generate. This showcase will present a starting point for those of you who are interested in creating your own trading systems using our tools and some AI techniques.
Many have heard of neural networks. These used to be a hot topic on Wall Street, but fell out of favor for a time when the public discovered they did not represent the holy grail of trading after all, and that a lot of hard work went into creating a good model. Basically, a neural net takes in a certain number of inputs and uses these inputs to predict a given output. For traders, the output we try to predict is future price/indicator movement. If this can be done accurately, we have the basis for a profitable trading system. We have created a number of profitable systems using neural nets, so pay no attention to those who say this can't be done.
We will now discuss a recent net we have created, and will demonstrate how it has been performing recently. We told the net to try and predict the change in a 9 bar MA_Momentum curve 6 places into the future. The inputs we fed the net were:
6, 9, and 13 bar MA_Momentum curves
Cycle Forecaster (both raw and difference from last value)
These inputs were minimally processed, and then fed into a neural network. The net was trained on data from 1/98-1/99, and produced encouraging results. Below is a chart of the out of sample period (the net did not know about this data when it was being trained):
The yellow line at the bottom is the neural forecast. We have built a simple system that buys and sells on zero line crossovers. Below is a performance report:
As you can see, the results have been very good over the entire year. This shows that this method does have the ability to make good trading decisions and should be explored further. There are a few bothersome statistics such as the $30,000 maximum drawdown, but on the whole the numbers are very promising.
This represents the core of a possible trading system. The neural net could be used as a signal generator, while various filters and exits would need to be developed to guard against losses. Stop losses, entry patterns, and time based exits would all be good things to look at adding to the basic crossover system. Neural nets are very good at finding relationships between different data streams. If you keep things simple and feed the nets useful data, you will be able to create profitable models as we have done here.
Back to Archives