Improving GARCH model with supply and demand sentiments from news articles

Abstract

This project investigates the improvement of volatility modeling in crude oil markets by incorporating news-based sentiment scores into asymmetric GARCH-type models. Focusing on Brent Crude Oil returns, we extend traditional volatility frameworks by incorporating exogenous sentiment variables derived from oil-related news. Over 15 months of articles (January 2024 – April 2025) were collected from Oilprice.com. FinBERT was employed to quantify sentiment at the paragraph level, filtered using manually curated Supply and Demand keyword lists. These scores were aggregated, aligned with NYMEX trading days, and subsequently first-differenced to ensure stationarity for model input in the primary analysis period. After confirming ARCH effects, we fit various asymmetric models, specifically GJR-GARCH-X and eGARCH-X specifications, and incorporate the lagged differenced sentiment. Model performance was evaluated based on statistical significance, in-sample fit, residual diagnostics and out-of-sample forecast accuracy. Our results show that sentiment significantly improves model fit within the asymmetric frameworks. While in-sample fit favoured models with sentiment primarily in the variance equation, out-of-sample forecasting performance was best for an eGARCH-X model incorporating sentiment in both mean and variance, highlighting a fit-versus-forecast trade-off. However, forecast improvements were not universal across specifications, and the GJR-GARCH-X model exhibited instability during forecasting evaluations. This study contributes by demonstrating a pipeline for integrating granular sentiment, categorized by Supply and Demand themes, into asymmetric volatility models, providing evidence of its value for both model fit and forecasting in commodity markets, while underscoring the importance of comprehensive model evaluation.