Impact of Learning Functions on Prediction of Stock Data in Neural Network

Digitization has made a vast impact on the modern society. Financial sector is one field where a huge revolution has been experienced because of digitization. Financial data especially time series data is being stored in the digital repositories where it can be used for prediction and analysis. One such data is a stock market data which is a time series data and is generated in a huge amount every second. The stock market data is of great importance as the proper analysis and prediction of data can transform the fate of the global market. Thus the companies and the individuals are looking forward for the development of the automated techniques that can predict stock market data accurately in a real time. In this regard, many researchers developed machine learning techniques such as use of neural network for prediction of stock data. The most common learning function used in neural network is sigmoid function. However, we found that there are many learning functions are available for building neural network. In this paper we are studying the impact of four different learning functions in estimating/predicting the stock value. From the experimental study we found that unipolar sigmoid learning function produced an accuracy of 95.65%, bipolar sigmoid produced an accuracy of 91.34%, tan hyperbolic equation produced an accuracy of 91.02%, and radial base equation produced an accuracy of 87.53%. Clearly, unipolar sigmoid function emerged as the best learning function to build stock data prediction model. The main reason behind its out-performance of unipolar sigmoid is its less complex structure and the 0 to 1 range.