Can News Predict the Direction of Oil Price Volatility? A Language Model Approach with SHAP Explanations (2508.20707v1)

Abstract: Financial markets can be highly sensitive to news, investor sentiment, and economic indicators, leading to important asset price fluctuations. In this study we focus on crude oil, due to its crucial role in commodity markets and the global economy. Specifically, we are interested in understanding the directional changes of oil price volatility, and for this purpose we investigate whether news alone -- without incorporating traditional market data -- can effectively predict the direction of oil price movements. Using a decade-long dataset from Eikon (2014-2024), we develop an ensemble learning framework to extract predictive signals from financial news. Our approach leverages diverse sentiment analysis techniques and modern LLMs, including FastText, FinBERT, Gemini, and LLaMA, to capture market sentiment and textual patterns. We benchmark our model against the Heterogeneous Autoregressive (HAR) model and assess statistical significance using the McNemar test. While most sentiment-based indicators do not consistently outperform HAR, the raw news count emerges as a robust predictor. Among embedding techniques, FastText proves most effective for forecasting directional movements. Furthermore, SHAP-based interpretation at the word level reveals evolving predictive drivers across market regimes: pre-pandemic emphasis on supply-demand and economic terms; early pandemic focus on uncertainty and macroeconomic instability; post-shock attention to long-term recovery indicators; and war-period sensitivity to geopolitical and regional oil market disruptions. These findings highlight the predictive power of news-driven features and the value of explainable NLP in financial forecasting.