Modifying Hidden Layer in Neural Network Models to Improve Prediction Accuracy: A Combined Model for Estimating Stock Price

Abbas Mahmoudabadi,Fatemeh Pourhossein Ghazimahalleh,Mehdi Kanaani

doi:10.28991/hij-2022-03-01-05

Abbas Mahmoudabadi, Fatemeh Pourhossein Ghazimahalleh + Show 1 more

Open Access

https://doi.org/10.28991/hij-2022-03-01-05

Copy DOI

Abstract

Investment experts, who deal with stock price estimation, commonly look for the most accurate and appropriate statistical techniques to make decisions on investment. The aim of this study is to improve the accuracy of stock price prediction models through modifying the structure of a combined neural network model with time-series data, in which the main contribution is to insert the time-series analysis prediction into the hidden layer of the neural network. The proposed structure is made up of neural networks and time-series analysis, with variable reduction used to remove attributes with inter-correlations. Data has been collected over six years (72 months) from the Iranian stock market, including the number of trades, new-coin price, gold-18 price, US Dollar and Euro equivalent currencies, oil-index price, Brent-oil price, industry index, and balanced stock index, followed by developing the prediction models. Comparing the performance criteria of the proposed structure to the traditional ones in terms of the mean square and mean absolute errors revealed that inserting time-series estimated variables into hidden layers would improve the performance of neural network models to estimate stock prices for making investment decisions. Doi: 10.28991/HIJ-2022-03-01-05 Full Text: PDF

Highlights

IntroductionMany models have been proposed utilizing a variety of fundamental, technical, and time-series forecasting techniques to gain accurate predictions in this field
Investment experts, who deal with stock price estimation, commonly look for the most accurate and appropriate statistical techniques to make decisions on investment
A notable success of the proposed models was achieving prediction accuracies of over 80% based on the Dow Jones monthly industrial index predictions, and the results demonstrated that Radial Basis Function (RBF) neural networks are preferred to Multi-Layer Perceptron (MLP) networks

Summary

Introduction

Many models have been proposed utilizing a variety of fundamental, technical, and time-series forecasting techniques to gain accurate predictions in this field. One of the main concerns in stock market investment is to gain an overall view of the future and predict the trend of stock prices as well as illustrative graphs to make the right decisions and affordable plans for the future. Precise forecasting of markets may be impossible, the researchers intend to tackle this problem by proposing methods that are more accurate and comparing the accuracy of the prediction models to select the best method. Some characteristics of financial time-series data, such as non-stationarity, nonlinearity, and high volatility, prompt investment professionals to create more accurate and fitted models [3].

Objectives

Results

Conclusion