The Influence of Financial News Sentiment on Investor Behaviour and Market Dynamics: A Machine Learning Approach to Sentiment Classification

Financial news sentiment influences investor decisions, but most research focuses on market-level outcomes rather than how individual investors with different personalities interpret the same news. We combined a Random Forest sentiment classifier (80.2% accuracy on FinancialPhraseBank) with a multi-agent simulation using FinBERT on 2,263 expert-labeled headlines, producing 226,300 decisions across 100 agents with randomized perception, portfolio memory, and herding dynamics. The clearest result was about neutral headlines: agents agreed on only 89.4% of neutral headlines compared to 98-99% for positive and negative, which mirrors the classifier's 26% neutral error rate. Herding cascades appeared in 32.8% (buy) and 16.0% (sell) of headlines, but ablation tests showed cascades barely changed when the herding mechanism was turned off, meaning agents were independently reaching the same conclusion rather than copying each other. A small shift toward caution appeared but was too small to be meaningful. Overall, neutral sentiment appears to be where both classification and investor agreement break down.