Does sentiment analysis bring more responsive and comprehensive commodity price forecasting: Case of machine learning models?

Accurate forecasting of commodity prices is vital for effective decision-making in trading, investment, and policy. This study combines sentiment analysis with Machine Learning (ML) and Natural Language Processing (NLP) techniques to enhance predictive performance. Sentiment data are extracted from Reddit discussions, Google News articles and Twitter(X) and incorporated alongside historical price data for eleven key commodities across the energy, agriculture, and metals sectors. To better understand the underlying price dynamics, Local Lyapunov Exponent (LLE) and Hurst Exponent (HE) are applied to high-frequency data, identifying chaotic or persistent patterns that guide model selection. The predictive performance of several models is evaluated, including Linear Regression (LR), Random Forest (RF), Gradient Boosting Regressor (GBR), Deep Neural Networks (DNN), and Long Short-Term Memory (LSTM) networks. The results indicate that sentiment integration improves forecasting accuracy in selected markets rather than uniformly across all commodities. RF and GBR maintain strong performance on structured data, and sentiment features further reduce MSE in specific cases. LSTM and DNN models show the largest relative gains when enriched with sentiment features, particularly in news-sensitive markets. To enhance transparency and confidence in model results, SHapley Additive exPlanations (SHAP) values are used to interpret the influence of input characteristics, including sentiment signals, on individual predictions. These findings highlight the value of hybrid forecasting models that integrate both quantitative and qualitative inputs using ML and NLP tools to provide more adaptive and robust price predictions in dynamic commodity markets. This offers valuable insights into the role of market sentiment and historical trends in shaping price dynamics.