More than words: valuation of words for stock price by using the combination of natural language processing, time-series panel and gradient boosting

Purpose This study investigates whether consumer complaints and sentiment can predict abnormal stock returns in the U.S. financial sector. It explores the financial impact of behavioral signals derived from complaint narratives, aiming to integrate consumer voice into financial forecasting. Design/methodology/approach Using monthly complaint data from the Consumer Financial Protection Bureau and financial data from 261 publicly traded financial firms between 2018 and 2023, the study applies Latent Dirichlet Allocation to extract complaint topics and VADER sentiment analysis to quantify emotional tone. These variables are incorporated into fixed-effects panel path models and machine learning to evaluate their predictive value for stock price movements. Findings Higher complaint volume and stronger negative sentiment are significantly associated with short-term stock price declines. Topic-specific complaint trends also contribute to prediction accuracy, suggesting that investors may interpret consumer complaints as early signals of reputational or financial risk. Practical implications Monitoring consumer complaints can help financial firms detect emerging risks and manage reputational threats. For investors, sentiment and topic data from complaints offer complementary behavioral indicators to enhance forecasting models. Originality/value This study introduces a novel integrated framework to quantify behavioral signals from large-scale consumer complaint narratives using natural language processing (NLP). By incorporating these textual features into financial econometric models, it advances behavioral finance and demonstrates the predictive value of consumer voice in explaining abnormal stock returns.