AI Driven Stock Price Prediction Using Historical Data and Sentiment from News Sources with Deep Learning

Abstract

Stock market forecasting remains a challenging problem due to the inherently volatile, nonlinear, and sentiment-driven nature of financial markets. Traditional single-modality deep learning approaches fail to capture the full complexity of price dynamics, especially during periods of sudden market reactions triggered by news events and shifts in investor psychology. This paper proposes a novel hybrid framework that integrates a Temporal Convolutional Network (TCN) for time-series stock price forecasting with Bidirectional Encoder Representations from Transformers (BERT) for real-time financial news sentiment analysis. TCN serves as the core prediction engine, modeling both short-term and long-term temporal dependencies through causal and dilated convolutions while circumventing the vanishing gradient problem common in recurrent architectures. Real-time news is collected using the Newspaper3K library and analyzed by a fine-tuned BERT model that extracts contextual sentiment polarity from financial text. Sentiment scores are temporally synchronized with historical stock data and technical indicators before fusion into the TCN model. Experimental results demonstrate that the TCN-BERT framework achieves 97.86% prediction accuracy with significantly reduced forecasting error, substantially outperforming single-modality baselines and confirming that sentiment-aware multimodal fusion improves prediction robustness and reliability in dynamic financial environments.