December 2014 Newsletter

 December 2014

Ensemble Neural Nets in FOREX Trading ModelsEnsemble

by Marge Sherald, CEO

Mission:  find three trading strategies that update as new bars arrive.

  • Build a Trading Strategy based on signals from all three to increase probabilities of success.
  • Limit trading time to London hours, the largest FOREX market.
  • Match pattern matching neural nets to the patterns of major European banks, which trade predominantly in the London market.

According to a December 2014 Technical Analysis of STOCKS AND COMMODITIES article by Imran Mukati, following the banks “provides an opportunity for individual pattern traders to exploit the movements of London in particular.”

The Particulars

If you attended the first NeuroShell Trader webinar, you learned about three different neural network types that update the models when new data arrives in the chart.  Specifically, I used GRNN nets from the Adaptive Net Indicators Add-on, Adaptive TurboProp 2, and Recurrent Nets from the Neural Indicators Add-on.

The model for the webinar was based on a chart for the S&P 500 EMini, but I wanted to see if the same neural net types with the same inputs would work on a FOREX chart.

I used the following indicators as inputs to the model:

  • Momentum from the Change Category
  • RSI and Stochastic %K from the Price Momentum Category
  • Money Flow Index from the Volume Category

After some experimentation, it appears that 15 minute bars capture market movement for the AUDUSD pair the best when using these particular neural net models.

The models trade 100000 unit lots with commissions of $4 in and out of a trade, with a 0.3 pip spread on one side only.  (I used the Trader’s interface to FXCM for data and trading so the costs reflect the recently lowered FXCM fees at the time the model was created.)  The models are set to trade between the hours of 5:30 am to 10:30 am eastern time in the US to correspond to the London market.


All three models were successful in the out-of-sample period in their own right. When combined, they increased the probability of success.  Note the blue line in the lower subgraph in the chart above that displays the equity curve for the combined model.  The trading rules found by the optimizer for the model are displayed on the right.

Increasing the probability of success does not necessarily mean ensemble models will earn more money than an individual model, but it does indicate that the trading signals are more likely to be correct.  Check out Ensemble Estimators for the mathematical proof.

Click here to download this chart.  In order to correctly view the model, you need to own all of the Ward Systems Group Add-ons described in the article.  Like all of our models this one is intended for instructional use only as market conditions are constantly changing. 

Was this article helpful?

Related Articles