Polynomial Based Functional Link Artificial Recurrent Neural Network adaptive System for predicting Indian Stocks

D. K. Bebarta, Birendra Biswal, P. K. Dash
2015 International Journal of Computational Intelligence Systems  
A low complexity Polynomial Functional link Artificial Recurrent Neural Network (PFLARNN) has been proposed for the prediction of financial time series data. Although different types of polynomial functions have been used for low complexity neural network architectures earlier for stock market prediction, a comparative study is needed to choose the optimal combinations of the nonlinear functions for a reasonably accurate forecast. Further a recurrent version of the Functional link neural
more » ... is used to model more accurately a chaotic time series like stock market indices with a lesser number of nonlinear basis functions. The proposed PFLARNN model when trained with the well known gradient descent algorithm produces reasonable accuracy with a choice of range of weight parameters of the network. However, to improve the accuracy of the forecast further, the weight parameters of the recurrent functional neural network are optimized using an evolutionary learning algorithm like the differential evolution (DE). A comparison with other well known neural architectures shows that the proposed low complexity neural model can provide significant prediction accuracy for one day advance and speed of convergence using the International Business Machines Corp. (IBM) stock market indices. Keywords: PFLARNN, Polynomial functions, backpropagation learning algorithm, differential evolution, IBM stock indices, MAPE, AMAPE. from Table- 1, where all the performance metrics like MAPE, AMAPE, and variances have the least magnitudes. (a): Forecasting performance during the testing period with network weights -1.0 to +1.0 (b): Forecasting performance during the testing period with network weights -2.0 to +2.0. (c): Forecasting performance during the testing period with network weights -3.0 to +3.0. Fig.7: Effect of varying network weights on the forecasting performance using Laguerre Polynomial Function. (a): Forecasting performance during the testing period with network weights -1.0 to +1.0 (b): Forecasting performance during the testing period with network weights -2.0 to +2.0. (c): Forecasting performance during the testing period with network weights -3.0 to +3.0. Profit/loss using different PFLARNN methods Avg. actual value Avg. forecast value without DE Avg. forecast value with DE Legendre Polynomial Function with Range of network weights 1) -1.0 to +1.0 2) -2.0 to +2.0 3) -3.0 to +3.0 Chebyshev Polynomial Function with Range of network weights 1) -1.0 to +1.0 2) -2.0 to +2.0 3) -3.0 to +3.0 Lagurre Polynomial Function with Range of network weights 1) -1.0 to +1.0 2) -2.0 to +2.0 3) -3.0 to +3.0 Power Series Polynomial Function with Range of network weights 1) -1.0 to +1.0 2) -2.0 to +2.0 3) -3.0 to +3.0
doi:10.1080/18756891.2015.1099910 fatcat:lrh5jyc4dbcvrbrxdr3zktkhwq