The following information is for those users who have previously used another neural network program, such as our own NeuroShell 2 Professional System. The neural net included in the NeuroShell Trader is unlike any other type of network you may have used. Unlike other back propagation networks, for example, it learns quickly and a test set is not required to improve the net’s ability to generalize (give reasonable answers on data it has never seen before). Because the net is designed to generalize well rather than to memorize training data, it may or may not do as well on the training set as it does on other data.
If you have worked with NeuroShell 2, you most likely used Calibration. Calibration requires the extraction of a test set from your original data in order for the network to generalize well on new data. The remainder of the original data was called the training set.
The training procedures in NeuroShell 2 actually used both the training and test sets in order to create a model. Therefore, if you want to compare the results of a NeuroShell 2 net with those in the NeuroShell Trader, you must use all of your original data (both the training and test sets) when you train your model with the NeuroShell Trader. It has a built-in method for making the model generalize without extracting a test set.
The nets in the NeuroShell Trader do not require you to set parameters such as learning rate and momentum, so they are easier to use than backpropagation nets. Additionally, the nets can only have one output, so if you want to predict more than one output, you need to create separate models for each output using the same set of inputs. Finally, the neural net works better with continuous value inputs than it does with binary inputs (1’s and 0’s).