International Conference on Computer and Knowledge Engineering

Home / 15th International Conference on Computer and Knowledge Engineering

FinTNet: From Tweets to Trades

Authors :

Dorsa Tavakoli¹ Saman Haratizadeh²

1- College of Interdisciplinary Science and Technology, University of Tehran, Tehran, Iran 2- College of Interdisciplinary Science and Technology, University of Tehran, Tehran, Iran

Keywords :

Stock Market Prediction،Deep Learning،Text Analysis،Graph Neural Networks،Large Language Models

Abstract :

Recent research indicates a correlation between textual information, media sources, and stock price movements. Additionally, studies have highlighted the influence of inter-stock relationships on their respective behaviors. In this study, FinTNet is introduced as a novel graph neural network model that integrates financial text data, price information, and stock relationships for enhanced stock price prediction. Three text-based graphs, derived from Twitter text analysis and the exploration of stock relationships, are presented for prediction using a semi-supervised graph convolutional model. The predictive model concurrently incorporates textual data, stock prices, and indicators. For in-depth tweet content analysis, Large Language Models (LLMs), including Financial Bidirectional Encoder Representations from Transformers (FinBERT) and Gemini Pro, are employed. FinBERT transforms tweets into embedding vectors, yielding a dataset of approximately three million tweets with vectors of size 768. Additionally, about 85,000 are getting four different labels (trend direction, discussed time, amount of change in price, and sentiment), using the Gemini Pro model, forming an accessible labeled dataset. Experiments demonstrate that FinTNet achieves superior accuracy compared to baseline models by leveraging LLM-extracted textual features within carefully designed graph structures.