Tehran Stock Exchange Price Movement Prediction using Daily News with Hierarchical Attention Network Plus BERT

Articles in Press

Document Type : Original Article

Authors

Computer Engineering Department, Yazd university, Yazd, Iran

Abstract

The stock market’s significance in the global economy necessitates demands more accurate prediction methods. This paper introduces a novel hierarchical attention mechanism aimed at enhancing the performance of predicting stock price movements. Hierarchical attention networks assume that not all news segments hold equal relevance in forecasting stock market trends. Furthermore, we assert that not all daily news carries an equivalent significance in predicting market trends. To tackle this challenge, we suggest a hierarchical attention network plus BERT that emulates the news hierarchy and assigns weights to news items based on their significance, and also the most informative news articles in each trading day in stock market prediction. Our HAN+BERT method incorporates three levels of attention mechanisms, operating at the word, sentence, and news level. This allows the model to identify the most significant news stories of the day and select the most informative sentences and words within these articles. Using BERT as the word embedding approach has resulted in better performance for our stock trend prediction model. Empirical results on Persian financial news and three stock market indices reveal the effectiveness of our HAN+BERT model, with a peak accuracy of 65.49%, which is 3% higher than the best baseline model.

Keywords

Tabriz Journal of Electrical Engineering

Articles in Press, Corrected Proof
Available Online from 21 April 2025

Files

Share

How to cite

Statistics