Performance Analysis Of Narx Neural Network Back Propagation Algorithm By Various Training Functions With Tracking Signal Approach For Time Series Data
Murugan Ashok Kumar · Zenodo (CERN European Organization for Nuclear Research) · 2017
This study proposed a novel Nonlinear Auto Regressive eXogenous Neural Network (NARXNN) with Tracking Signal (TS) approach and seeks to investigate the various training functions to forecast the closing index of the stock market. A novel approach strives to adjust the number of hidden neurons of a NARXNN model with different training functions. It uses the Tracking Signal (TS) and rejects all models which result in values outside the interval. The effectiveness of the proposed approach is seen to be a step ahead of Bombay Stock Exchange (BSE100) closing index of Indian stock market. This novel approach reduces the over-fitting problem, neural network structure, training time; fast at convergence speed and improves forecasting accuracy. In addition, the present approach has been tested with different training functions and identified the neuron counts in the hidden layer for every training function which leads to reduce over-fitting or under-fitting problem.