Optimizing Postprocessing of Predicted Signals

Neural network predictions train by adjusting their weights in order to reduce the difference between the neural network output and the desired output. If the neural network does not have adequate information in its inputs to make a perfect prediction, it will tend to adjust towards the mean (average) desired output value. When predicting signals, this mean value is typically around zero, which represents a hold signal.

Due to the noise and volatility associated with price data, predicted signals based on price data often remain around this mean of zero even when informative inputs are present. This does not mean that the predicted value does not have any meaning, just that the signal may be compressed towards zero.

To extract the predicted information and turn it into a tradable signal, genetic optimization is used to locate the best values to amplify the predicted output. This optimization occurs automatically for all predicted signals. By default, it is based on the trading style which was used to create the desired output.

The settings associated with signal postprocessing can be viewed or modified from the Signal Postprocessing Settings Dialog.

&  For help with general optimization principles, see Understanding Genetic Optimization.

&  For more help with predictions, see Predicting and Modeling Financial Data.