Enhancing Stock Market Prediction Accuracy Using Advanced Sentiment Analytics and Hybrid Machine Learning Frameworks

Abstract

The growing complexity of financial markets and the rapid influence of information-driven events have heightened the demand for more accurate and context-aware stock market prediction models. Traditional forecasting approaches, which depend largely on numerical historical patterns, often overlook the critical role of news sentiment, market psychology, and narrative-driven shifts that shape investor behaviour. This study investigates the effectiveness of integrating advanced sentiment analytics with hybrid machine learning frameworks to improve stock market prediction accuracy. A comprehensive dataset of large-scale financial news articles and market-related textual information is analysed using a multi-stage sentiment extraction pipeline that incorporates lexicon-based tools, machine-learning sentiment classifiers, and domain-specific contextual models such as FinBERT. These enriched sentiment features are then combined within deep learning and transformer-based hybrid architectures, including LSTM, Bi-LSTM, and attention-driven models.

Evaluation using RMSE, MAE, MAPE, R², and Directional Accuracy demonstrates that contextual sentiment embeddings lead to substantial improvements over both traditional sentiment methods and sentiment-free models. Among all tested approaches, Transformer-based hybrid models deliver the strongest performance, highlighting the advantage of jointly modelling semantic patterns and temporal dependencies in financial narratives. The study underscores the pivotal role of sophisticated sentiment analytics in capturing real-time market emotions and enhancing the reliability of predictive systems. These findings offer valuable insights for quantitative analysts, institutional investors, and algorithmic traders seeking more adaptive and information-rich forecasting strategies.