Initialization of multilayered forecasting artificial neural networks

In this paper a method of initializing a neural network solving the problem of forecasting a time series is offered. The analogy with a linear prediction filter is used to select an initial weighting coefficient for neurons. Also the variants of decomposition of a transformation matrix corresponding to the linear prediction filter are offered to improve the initialization method quality. Through the neural nets forecasting of the Lorentz chaotic system's trajectory it is shown that the application of the given method allows significantly increasing the accuracy of the neural network prediction.