Advanced Deep Learning System for Financial News-Based Sentiment Analysis in Stock Price Prediction

Abstract

The complex connection between news-driven mood and macroeconomic factors, as well as the intrinsic volatility of financial markets, makes stock price prediction a tough undertaking. This study recommends a Bi-direcstional Long Short-Term Memory (Bi-LSTM) model for analyzing the tone of financial news as a means to improve stock price prediction. The Financial Phrase Bank dataset, which contains 4,848 news headlines categorized as positive, negative, or neutral financial, is a good place to start when developing the approach. Before dividing the dataset into training and testing subsets, data must undergo preprocessing operations such as text cleaning, tokenization, normalization, and label encoding. From what can tell from the experiments, the Bi-LSTM gets an F1-score of 90.62%, a recall of 90.11%, and a precision of 91.13%. When compared to standard models like FinBERT and Random Forest, the Bi-LSTM performs far better. The results show that the suggested method is suitable for practical uses in financial forecasting since it combines sentiment analysis with sophisticated deep learning architectures to give a solid foundation for dealing with market unpredictability.