AI
FinBERT and LSTM networks combine to predict stock market trends from financial sentiment
A new deep learning approach fuses specialized financial language models with sentiment analysis to forecast market movements more reliably than traditional methods.
Researchers have developed a hybrid machine learning system that applies sentiment analysis to financial news in order to improve stock market forecasting. The work, published on arXiv, combines FinBERT—a version of the popular BERT language model trained specifically on financial text—with Long Short-Term Memory (LSTM) neural networks to capture both market sentiment and temporal dependencies in stock price movements.
The approach#
The model leverages FinBERT's ability to understand financial language nuances, extracting sentiment signals from news and other textual market data. These sentiment scores feed into an LSTM network, which is designed to recognize patterns in sequential time-series data. This layered architecture allows the system to process both what people are saying about markets and how that sentiment evolves over time.
Comparative results#
The team evaluated their integrated system against several baselines: a standard BERT model without financial specialization, a standalone LSTM network, and ARIMA—a classical statistical forecasting method. The results showed that combining sentiment analysis with deep learning substantially improved prediction accuracy relative to each of these alternatives. The findings suggest that market sentiment extracted from financial text provides genuine predictive value beyond what price history or generic language understanding alone can offer.
Path forward#
The authors also discuss optimization techniques aimed at further refining model performance, suggesting the framework could serve as a foundation for more reliable AI-driven financial prediction tools in the future.
