The IUP Journal of Applied Finance
Predictive Accuracy of Neural Network Model with Multiple Train Functions for Stochastic Stock Indices

Article Details
Pub. Date : Oct, 2020
Product Name : The IUP Journal of Applied Finance
Product Type : Article
Product Code : IJAF01020
Author Name : Vijay Shankar Pandey and Jitendra Kumar Sharma
Availability : YES
Subject/Domain : Finance
Download Format : PDF Format
No. of Pages : 25

Price

Download
Abstract

Two popular stock indices, i.e., BSE and NSE, are modeled through Artificial Neural Network (ANN) to identify superior combinations for predictive accuracy. Ten-year daily data of closing index along with open, high and low indices are modeled through NARX modeling. The statistical tools AAE, RMSE, MAPE and MSPE are used to find out the predictive accuracy of ANN model. The results indicate ANN (4-10-1) with trainfunction GDX as the best predictor within the final outcome. The high predictive accuracy of the model to predict stock indices suggests a relook at the EMH for long-term data series.


Description

The models used for forecasting stock markets are either deterministic or stochastic in nature. These models for forecasting time series data are based on associated components of tendency and cyclical and seasonal trends which are deterministic in nature, along with stochastic random changes (Rusu and Rusu, 2003). The stock market being dynamic generates white noise which defines prediction by deterministic models. Therefore, research has mostly focused on random walk models, prominent amongst which are autoregressive category of models considered as better representative of stochastic models for predicting stock market behavior.

The linear models, due to inherent limitation in learning the real-world problems of forecasting, gave rise to development of nonlinear models such as nonlinear autoregressive, threshold model (Tong, 1983) and autoregressive bilinear model (Granger and Anderson, 1978), smooth transition regression model (Bacon and Watts, 1971), autoregressive heteroskedastic model (Engle, 1982) Exponential Smooth Transition Autoregressive (ESTAR) model (Sarno and Taylor, 2002) and Markov switching model (Hamilton, 1996). Though these models have provided better results than the linear models, their universal acceptability is questionable due to differences in the nature of markets, length and period of study (Zhang, 2001).

Modeling of stochastic time series data by researchers (Hornik et al., 1989; Tang et al., 1991; Tang and Fishwick, 1993; Hill et al., 1996; Zhang, 2003; and Chiang et al., 2007) using


Keywords