Beyond Polarity: Multi-Dimensional LLM Sentiment Signals for WTI Crude Oil Futures Return Prediction

Dehao Dai, Ding Ma, Dou Liu, Kerui Geng, Yiqing Wang · 2026

Read Beyond Polarity: Multi-Dimensional LLM Sentiment Signals for WTI Crude Oil Futures Return Prediction by Dehao Dai, Ding Ma, Dou Liu, Kerui Geng, Yiqing Wang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-11408.txt.

Introduction

Dai and co-authors test whether richer news sentiment can help predict weekly moves in WTI crude oil futures. Simple positive or negative scores miss much of what energy news carries. The team uses large language models to score five traits in each story, from relevance to uncertainty. News from 2020 to 2025 feeds a weekly model. Pairing GPT-4o with FinBERT works best, and the intensity and uncertainty scores prove most useful.

Catalog record

CREATORDehao Dai, Ding Ma, Dou Liu, Kerui Geng, Yiqing Wang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE38514
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.11408

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Beyond Polarity: Multi-Dimensional LLM Sentiment Signals for WTI Crude Oil Futures Return Prediction Dehao Dai, Ding Ma, Dou Liu, Kerui Geng, Yiqing Wang Abstract Forecasting crude oil prices remains challenging because market-relevant information is embedded in large volumes of unstructured news and is not fully captured by traditional polarity-based sentiment measures. This paper examines whether multi-dimensional sentiment signals extracted by large language models improve the prediction of weekly wti crude oil futures returns. Using energy-sector news articles from 2020 to 2025, we construct five sentiment dimensions covering relevance, polarity, intensity, uncertainty, and forwardness based on GPT-4o , Llama 3.2-3b , and two benchmark models, FinBERT and AlphaVantage. We aggregate article-level signals to the weekly level and evaluate their predictive performance in a classification framework. The best results are achieved by combining GPT-4o and FinBERT , suggesting that LLM-based and conventional financial sentiment models provide complementary predictive information. SHAP analysis further shows that intensity- and uncertainty-related features are among the most important predictors, indicating that the predictive value of news sentiment extends beyond simple polarity. Overall, the results suggest that multi-dimensional LLM-based sentiment measures can improve commodity return forecasting and support energy-market risk monitoring. 1 Introduction Crude oil remains a cornerstone strategic resource for modern economies and plays a fundamental role in sustaining economic growth and social stability ( 30 ; 12 ; 45 ) . As the most actively traded energy commodity, crude oil prices influence inflation, industrial production, transportation costs, and financial market dynamics. Accordingly, crude oil futures markets have become the primary venues for price discovery, and accurate forecasting of crude oil price remains a persistent challenge for market participants. There is a considerable literature to address this challenge including traditional statistical frameworks used in high-dimensional time series like ARIMA and GARCH ( 72 ; 2 ) , feature selection ( 76 ; 59 ) and modern machine learning techniques such as LSTM networks and deep learning models ( 77 ; 29 ) . These models emphasize quantitative analysis of structured data to provide well-performed crude oil price forecasting and have also been widely applied to broader equity markets. However, their forecasting accuracy is constrained due to lack of the crude oil markets’ intrinsic features. Different from equity markets, which are largely driven by firm-level fundamentals, crude oil futures markets are highly sensitive to expectation-driven risk factors, including supply disruptions, geopolitical conflicts, policy interventions, and global demand uncertainty ( 5 ) . In contrast to numerical data, text data like news, reports and social media can provide a new framework to forecast crude oil prices by extracting market signals. These resources encode timely sentiment, capture shifts in expectations, and often contain policy or geopolitical cues that influence oil price dynamics. Recent work on effective crude oil price forecasting using text-based and big-data-driven models further supports this direction ( 71 ) . There is a substantial literature evaluating the role of media tone and news sentiment in asset return prediction. Early studies on textual sentiment rely on dictionary-based approaches that quantify the frequency of positive and negative words to measure tone ( 64 ; 18 ) . More recent work employs transformer-based models such as FinBERT to classify financial text polarity ( 73 ) and forecast both returns and volatility ( 40 ) . Although these tools have improved sentiment measurement accuracy, they primarily focus on directional tone (positive versus negative) and sentiment intensity. However, oil futures pricing is driven not only by directional sentiment but also by forward-looking uncertainty and expectation revision. A news article may be neutral in tone yet signal elevated uncertainty or significant forward guidance regarding future supply conditions. Polarity-based measures are therefore structurally limited in capturing these economically relevant dimensions. Recent advances in large language models (LLMs) provide an alternative framework for semantic decomposition. Unlike traditional tools, LLMs can extract multi-dimensional attributes from text, including relevance, polarity, intensity, uncertainty, and forward-looking orientation. These dimensions may better align with the theoretical drivers of commodity pricing, ( 38 ; 69 ; 33 ) particularly in markets where geopolitical risk and expectation formation play central roles. Building on this insight, we examine whether LLM-derived signals offer incremental predictive value for wti crude oil futures. In this papeinvestigates whether multi-dimensional sentiment signals extracted by LLMs provide incremental predictive power for weekly returns of wti crude oil futures. Using energy-related news articles from 2020 to 2025 from AlphaVantage, we construct weekly aggregated sentiment features generated by three models: GPT-4o , Llama 3.2-3b , and FinBERT . We compare their predictive performance using a LightGBM classifier and evaluate feature contributions through SHAP-based explainability analysis. Our findings suggest that LLM-derived uncertainty and forward-looking dimensions exhibit statistically and economically significant predictive power beyond traditional polarity measures. Multi-dimensional LLM-based sentiment measures can improve commodity return forecasting and support energy-market risk monitoring.Our findings contribute to the emerging area of LLM applications in financial economics and provide new evidence on the importance of multi-dimensional textual signals in commodity futures forecasting. The remainder of the paper is organized as follows. Section 2 reviews related work. Section 3 describes the end-to-end workflow pipeline from data collection to prediction modeling evaluation. Section 4 reports empirical results, and Section 5 discusses the results and future directions. 2 Related Literature 2.1 Crude Oil Futures Forecasting Predicting crude oil price movements has drawn scholars’ interest due to the oil market’s economic importance and the naturally high volatility of crude oil prices. Early economic approaches rely on traditional econometric and statistical models, including ARIMA, volatility models, and market-based predictors. Short-term forecasting with ARIMA and linear time series models remains a common benchmark ( 72 ; 50 ) . Earlier volatility-oriented studies, such as 20 , highlighted the strong macro-financial linkages surrounding oil price dynamics, and 2 showed that conditional heteroskedastic models provide important insights into the time-varying uncertainty of crude oil futures markets. There is also considerable literature that emphasizes machine learning and deep learning methods to capture the strong nonlinearity, regime dependence, and complex maturity structure of oil futures markets. For example, 13 modeled the crude oil futures curve through a dynamic Nelson–Siegel framework enhanced by neural networks, showing that nonlinear methods can improve forecasts of the term structure. Focusing on the Chinese market, 29 and 74 further developed the framework of deep learning to exploit historical prices, volatility, and other nonlinear features to enhance crude oil futures forecasting accuracy, while 77 and 27 proposed deep neural and hybrid learning models in extracting latent patterns from highly volatile oil price series. Moreover, recent studies adopt hybrid and multi-step forecasting frameworks to improve predictive performance over longer horizons. For instance, 22 developed a novel crude oil futures forecasting approach that combines multiple modeling stages, and 19 showed that hybrid designs can achieve superior performance in multi-step-ahead prediction by integrating complementary information from different model classes. A consistent finding across this literature is that energy markets are highly sensitive to expectation-driven shocks, like geopolitical events, OPEC+ policy announcements, and demand uncertainty, which has motivated researchers to integrate news-based information into forecasting models. For example, 17 adopted news media and attention spillovers across energy markets to possess significant predictive content for crude oil futures prices; 41 proposed a text-based deep learning approach that exploits

Read the complete text