Artificial Neural Networks or often called Neural Networks is a computation technique which has made significant progress in recent times. Neural networks have proven their capability of handling various problems in a number of scientific disciplines. Neural networks have a powerful ability called universal approximation, they can approximate all multivariate continue functions to every level of accuracy including for non-linear functions.

The ability of neural networks in universal approximation has been used by some researchers to forecast time series data in various kinds of data. The researches show that Neural Networks have a satisfactory performance in forecasting time series data.

Neural networks mechanisms imitate biological neural network mechanisms. Like biological neural networks, neural networks consists of neurons which are connected to each other and operate in parallel. The information processing mechanism in every neuron is adopted from the biological neuron.

Neurons in a neural network are grouped into several layers. Every layer can have one or more neurons. There are three layers in neural network architecture; they are the input layer, the output layer, and the hidden layer.

The function of the input layer is for data entry, data processing takes place in the hidden middle layer and the output layer functions as the data output result. The following illustration shows the architecture of neural networks.

Information processing in every neuron is done by summing the multiplication result of connection weights with input data. The result is transferred to the next neuron through the activation function. There are several kinds of activation functions, i.e. linear, semi linear, sigmoid, bipolar sigmoid and hyperbolic tangent.

In time series data forecasting, the input value for the input layer can be variable data of previous period (lagged variable) or the other variable used to help forecasting, can be qualitative or quantitative.

To forecast one variable (univariate), the input data for the input layer and output data in the output layer is similar to the autoregressive model AR(p). On certain point of t, forecasted data calculated by using observation from n previous point , where n shows the number of neuron inputs in a neural network.