A powerful but relatively easy way to use neural networks is to start with a profitable trading system and use neural network technology to enhance it. Consider the simple dual moving average crossover trading system.
First, we optimize the system on a basket of markets with data from Jan. 1, 1980, to Jan. 6, 2006, with $75 deducted for slippage and commissions. The basket includes corn, crude oil, cotton, the dollar index, the Japanese yen, coffee, natural gas and 10-year T-notes.
The moving averages are optimized from five to 30 in incremental steps of 5 for the shorter average, and from 20 to 50 in steps of 10 for the longer average. The optimization shows that the 10, 30 combination produced the best results with $476,941 of profit during this period.
Now, here is the key question: If we could accurately predict the sign (+/-) of the 10-period minus 30-period moving average two bars into the future, how much more would the system make?
Believe it or not, the answer (on hypothetical testing) is the system would make $1.4 million on the same markets throughout the same period, which is an increase of $900,000. This is a jump of 300% caused by knowing this difference two days in advance. We cannot predict perfectly, but if the network is predictive at all, the original system is improved.
We have a profitable system without the network. If the network does nothing, we are still profitable. But what if the network hurts us? We can test for this by simulating the network failing with a fail-lag analysis, which tests the system with 1 and 2 additional days of lag. This gives an idea of the risk/reward of adding the neural network to this system.
Adding one day of lag dropped profits to about $425,000, and two days of lag dropped it to $358,000. So, in this one test, we can see that even if the network fails completely, then the system is still profitable. If the network is predictive, the results would be stellar.
Steve Zwick is editor-at-large for Futures magazine.