مهرشاد متانی

Abstract

Forecasting stock trends, due to their inherent dynamics and complexities, has been one of the most challenging topics for which researchers have developed various machine learning and deep learning models. These studies have generally not considered the impact of other stocks on price trend prediction. Therefore, this research aims to evaluate the performance of a circuit model based on convolutional graph networks and convolutional neural networks. The proposed method employs a convolutional graph network based on the features of convolutional neural networks to predict stock trends, taking into account both stock market information and the specific target stock information. Subsequently, this data is transformed into images and analyzed using a dual convolutional neural network to extract the target stock's features. In summary, this model employs a combination of a convolutional graph network and a dual convolutional neural network to predict stock trends, capable of simultaneously capturing the features of the Tehran Stock Exchange market and the target stocks.

The statistical population of this research includes all companies listed on the Tehran Stock Exchange over a period of ۷ years and ۳ months (from the beginning of ۲۰۱۷ to the end of June ۲۰۲۴). Target stocks are selected through systematic sampling from industries with the highest trading value and liquidity. Additionally, after selecting the target stocks, ۳ related stocks with the highest correlation (Spearman rank correlation) from each stock industry are chosen.

The results indicated that this model generally performed acceptably under various market conditions and provided reasonable returns. On a relative scale, the model outperformed the convolutional neural network in terms of returns for all ۵ selected stocks. Furthermore, the model demonstrated better performance than the recurrent neural network for ۲ stocks and showed superior performance compared to the long short-term memory network for ۲ stocks in terms of returns. However, it should be noted that the circuit model based on convolutional graph networks and neural networks has a higher execution time and complexity compared to other models under review.