Optimizing Prediction Inputs and Settings

Many different settings associated with neural network predictions can be optimized to produce more effective signals and values. By default, TradingSolutions automatically optimizes many new predictions when they are created.

Ä    Note: This section discusses the optimization of prediction inputs and settings. This is separate from the optimization which occurs for the postprocessing of predicted signals. For more information about that, see Optimizing Postprocessing of Predicted Signals.

Optimizing Prediction Fields

The following settings associated with each prediction can be optimized:

·      Input selection

·      "Optimizable function" input parameters

·      Memory depth

·      Topology selection

·      Hidden layer processing elements

·      Weight update sample size

·      Learning momentum rate

·      Learning step size

·      Advanced learning criteria

For predictions of signals, such as the optimal signal, the optimization defaults to maximizing the fitness of the signal. The calculation of this fitness is defined in the trading style being used for postprocessing. This trading style defaults to the trading style used to define the desired output. It can be viewed and modified from the Signal Postprocessing Settings Dialog.

For predictions of values other than signals, the error in the prediction is minimized. The error is defined as the difference between the output of the network and the desired output.

Optimizing Prediction Inputs

When creating a prediction, you can select any combination or existing fields and optimizable functions as inputs. During optimization, these inputs can be enabled or disabled until the set of inputs giving the best results can be determined. In addition, the constant parameters associated with optimizable functions can be optimized.

By default, all inputs are included in optimization. Individual existing field inputs can be excluded from optimization by selecting it and pressing Modify Input…, which will display the Modify Prediction Input Dialog. Individual optimizable function inputs are always included in optimization. The settings associated with optimizing individual parameters for optimizable functions can be modified by selecting it and pressing Modify Input…, which will display a Modify Field Dialog: Function Inputs page.

After optimization, any optimizable functions which are included in the best solution will automatically create new fields based on the selected parameters and these fields will be added to the input list. If you elect to re-optimize the prediction, TradingSolutions will ask to remove these created fields from the prediction so that the previous results of the optimizable function do not affect the current results.

After optimization, inputs which are being used by the prediction will have a check next to them on the Modify Field Dialog: Prediction Inputs page. References to any other inputs, including optimizable functions, can be removed without modifying the value of the prediction.

Ä    Note: Predictions optimized with TradingSolutions v2.1 and earlier require any disabled inputs to reproduce their value since the inputs to the neural network were allocated based on the number and order of the overall inputs. Emulation of this previous requirement is controlled form the Modify Training Settings Dialog.

Optimizing Other Neural Network Settings

By default, TradingSolutions optimizes the inputs and memory depth associated with each prediction. It also automatically maintains all of the other settings associated with the neural network to their default values based on the number of inputs and memory depth being used. These automatic settings are based on the same rules of thumb used when modifying the inputs and memory depth through the interface.

Optimization can also be set to modify the settings in the neural network directly. This is controlled from the Training Optimization Settings Dialog. In order to keep the samples-to-weights ratio at a reasonable level, the optimization can be set to penalize low samples-to-weights ratios. This prevents the optimization from selecting a neural network which is likely to specialize only the training data due to having too many free parameters available.

Optimizing New Prediction Fields

Prediction fields are created using the Predict a Value Wizard. The optimization of new predictions is controlled from the Select Options page. The optimization settings for individual inputs can be accessed from the Select Inputs page.

Two types of optimization defaults are available for optimizing predictions – brief and full optimization. The primary difference between these two settings is the number of generations and maximum time allowed for optimization. Other optimization settings can be used by pressing the Optimization Settings… button. The defaults for brief and full optimization can also be adjusted from the Modify Options Dialog: Field Creation page.

Optimizing Existing Prediction Fields

Prediction fields which have already been created can be optimized (or re-optimized) from the Modify Field Dialog: Training Settings page. Selecting Re-optimize neural network inputs and settings on save will cause the prediction to be optimized from its current settings when the OK button is pressed. Fields which are defined for an entire group can be re-optimized by unchecking and re-checking this control.

Ä    Note: As mentioned above, TradingSolutions will request to remove any inputs associated with optimizable functions when a field is re-optimized.

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

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