The predictive ability of neural networks falls into the forecasting area and can be best related to the auto-associative memory of certain neural networks. The method used for neural network prediction is call generalization (Dutta and Shekhar, 1989). Generalization is different from auto-associative memory in that once the network has been trained, new data is input for the network to predict the output. Neural networks typically consist of six primary components; processing elements (neurons), connections between neurons, adaptive coefficients of connections, transfer functions, learning laws and the overall transformation process. Structural and mathematical details of neural networks can be found in several books and/or papers, e.g. Rumelhart and McClelland (1986).
Development of a particular neural network involves determining the number of perceptrons to be used. This refers to the number of layers that will be contained in the network. Most generally, neural network models are comprised of an input layer, an output layer, and at least one hidden layer. The number of nodes that are available at each layer also needs to be decided. This is most generally based on the particular problem being addressed. Once the number of layers, and nodes within each layer, are determined, the neural network is trained on a set of data. The training phase involves providing data to the neural network along with the known outcome that resulted. After many iterations through the training data set, the neural network will learn the patterns that are present in the data. The neural network will converge on a pattern which is based on the least number of errors during prediction of the training data outcomes. Once training is complete, the neural network is ready for use.