Forecasting oil volatility direction with news: A language model and regime-aware SHAP approach

This paper investigates whether financial news can forecast the direction of crude oil price volatility. While most existing studies focus on predicting volatility levels based on historical econometric dynamics, we examine the economically relevant task of anticipating directional shifts. Using approximately 590,000 Reuters oil-market headlines from 2014–2023, we construct daily textual predictors based on news volume and language-model embeddings and benchmark them against a directional heterogeneous autoregressive (HAR) model. We find that news volume and subword-level FastText embeddings significantly outperform the benchmark in out-of-sample forecasts. In contrast, complex transformer-based models (e.g., Gemini) do not yield statistically significant improvements, suggesting that lexical intensity is more informative than deep semantic context in short financial headlines. Structural breaks are identified using the Pruned Exact Linear Time (PELT) algorithm, and SHAP (SHapley Additive exPlanations) values are anchored to these regimes to reveal time-varying volatility drivers. Overall, our results show that news-based indicators provide forward-looking signals of volatility direction and that the importance of these signals varies across market regimes.