Integrating Classical Classifiers, Explainable AI and Transformer Models for Financial Headline Sentiment and Summarisation

The ability for Natural Language Processing (NLP) systems to transform textual data into quantifiable insights for financial news sentiment analysis for investment strategies, risk assessment and market trend forecasting is of predominant interest to financial market stakeholders. Unfortunately, the availability of large-scale labelled financial data required by these NLP systems for efficiency and effective performance remains limited due to confidentiality and proprietary issues. This challenge necessitates the implementation of new, transparent systems for sentiment analysis that involves less data and interpretable. This study implements a hybrid NLP system for summarizing financial news sentiment, which involves selecting the best embeddings (Word2Vec, GloVe, Sentence Transformers) for machine learning classifiers – Support Vector Classifier (SVC), Logistic Regression, Random Forest, Extreme Gradient Boosting – alongside explainable artificial intelligence methods like SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME). A transformer-based language model (Mistral-7B-Instruct) was employed for context-aware summarization, balancing predictive accuracy and interpretive depth through stratified K-fold cross-validation, few-shot examples and sentiment lexicon guidance to address data scarcity. The results show that SVC with Word2Vec embedding outperformed the other classifiers. It also revealed that even with limited data, Mistral-7B integrated with SHAP and LIME can generate understandable and actionable summaries, offering a transparent solution for financial sentiment analysis