💰 quantitative finance

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

This paper demonstrates that multi-dimensional sentiment signals extracted by large language models from energy news, particularly when combined with conventional financial sentiment models, significantly improve the prediction of weekly WTI crude oil futures returns by capturing nuances like intensity and uncertainty beyond simple polarity.

Original authors: Dehao Dai, Ding Ma, Dou Liu, Kerui Geng, Yiqing Wang

Published 2026-03-13

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Dehao Dai, Ding Ma, Dou Liu, Kerui Geng, Yiqing Wang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict the weather for next week.

For decades, meteorologists (and in this case, oil traders) have relied on thermometers and barometers. In the financial world, these are the "traditional" tools: looking at past prices, supply numbers, and simple "good news vs. bad news" counts. They tell you if the temperature is rising or falling, but they often miss the subtle shifts in the wind that signal a storm is coming.

This paper is about installing a super-powered, multi-sensory weather station to predict the price of WTI Crude Oil (the global standard for oil).

Here is the breakdown of what the researchers did, using simple analogies:

1. The Problem: The "Good/Bad" Trap

Traditional tools look at news articles and ask a simple question: "Is this article happy (bullish) or sad (bearish) about oil?"

The Flaw: Imagine a news headline: "The war in the Middle East might cause supply issues, but experts aren't sure yet."
- A traditional tool might say, "This is neutral," or "This is slightly bad."
- But the real story is: The uncertainty is high, and the future outlook is scary. The market hates uncertainty. Traditional tools miss the "shaking hands" and "sweaty palms" of the news; they only look at the "smiling face" or "frowning face."

2. The Solution: The "Super-Reader" (LLMs)

The authors used Large Language Models (LLMs) like GPT-4o and Llama 3.2. Think of these not as calculators, but as super-intelligent literary critics who read every news article and can detect five specific "flavors" of sentiment, not just good or bad:

Relevance: Is this story actually about oil, or is it just noise? (Like checking if a weather report is about your city).
Polarity: Is it good news or bad news? (The traditional "smile vs. frown").
Intensity: How strong is the feeling? Is it a mild suggestion or a screaming headline? (A whisper vs. a siren).
Uncertainty: How confused or ambiguous is the story? (Is the author saying "maybe" or "definitely"?).
Forwardness: Is the story looking at the past or predicting the future? (Looking in the rearview mirror vs. looking through the windshield).

3. The Experiment: The "Taste Test"

The researchers gathered thousands of energy news articles from 2020 to 2025. They fed them into three different "tasters":

The Old School: FinBERT (a specialized AI trained only on finance).
The New School: GPT-4o and Llama 3.2 (the super-readers).
The Data Provider: AlphaVantage (a standard industry tool).

They then tried to predict whether oil prices would go UP or DOWN the following week.

4. The Results: The "Secret Sauce"

Here is what they found, which is the most exciting part:

Mixing is Magic: The best prediction didn't come from just the super-readers or just the old school. It came from combining them.
- Analogy: It's like cooking a stew. The FinBERT provides the solid "meat" (reliable financial data), while the LLMs provide the "spices" (nuance, uncertainty, and future outlook). Together, they make a much tastier (more accurate) dish than either could alone.
It's Not About "Good vs. Bad": When they looked at which part of the news mattered most, they found that Intensity and Uncertainty were the real winners.
- Analogy: Knowing the temperature is 70°F (Polarity) isn't as important as knowing if the wind is picking up speed (Intensity) or if the barometer is dropping rapidly (Uncertainty). The market reacts more to how nervous people are than just whether the news is "positive."

5. The Takeaway

This paper proves that to predict oil prices, you can't just count "happy" and "sad" words anymore. You need to understand the texture of the news.

For Investors: Don't just look at the headline. Look at how confident the writer is and how strong the emotion is.
For the Future: We are moving from a world where AI just "reads" the news to a world where AI "understands the mood" of the news. This helps traders spot risks (like a coming storm) before the price actually crashes.

In a nutshell: The authors built a smarter radar that doesn't just see if it's raining, but tells you how hard the wind is blowing and how likely a hurricane is, giving oil traders a much better chance of staying dry.

1. Problem Statement

Forecasting crude oil prices is a persistent challenge due to the market's sensitivity to unstructured information (news, geopolitical events, policy shifts) that traditional quantitative models often fail to capture.

Limitations of Current Methods: Existing approaches rely heavily on structured time-series data (ARIMA, GARCH) or traditional sentiment analysis that focuses primarily on polarity (positive vs. negative) and intensity.
The Gap: Oil futures pricing is driven not just by directional sentiment but also by forward-looking uncertainty, expectation revisions, and the degree of ambiguity in market narratives. Simple polarity measures are structurally limited in capturing these multi-dimensional economic drivers.
Objective: To investigate whether multi-dimensional sentiment signals extracted by Large Language Models (LLMs) provide incremental predictive power for weekly WTI crude oil futures returns beyond traditional polarity-based measures.

2. Methodology

Data Collection and Preprocessing

Source: Energy-sector news articles from AlphaVantage (Jan 1, 2020 – Dec 31, 2025).
Sample: A stratified random sample of 8,639 articles (20% of the full corpus of 29,153) was used to ensure uniform temporal coverage.
Target Variable: Weekly binary classification of WTI crude oil futures returns ( $y_t = 1$ if $r_{t+1} > 0$ , else $0$).
Prevention of Bias: All sentiment features were constructed using news published no later than the end of week $t$ to avoid look-ahead bias.

Multi-Dimensional Sentiment Extraction

The authors constructed five sentiment dimensions using structured prompts for three distinct models:

GPT-4o: State-of-the-art instruction-tuned LLM.
Llama 3.2-3b: Lightweight open-source LLM.
FinBERT: A domain-specific transformer fine-tuned on financial text (serving as a supervised baseline).
AlphaVantage: Proprietary baseline sentiment score.

The Five Dimensions:

Relevance: Alignment with WTI/energy markets (0–1).
Polarity: Directional sentiment (Bullish/Bearish) (-1 to 1).
Intensity: Strength of conviction (0–1).
Uncertainty: Degree of ambiguity or risk (0–1).
Forwardness: Extent of future-looking language (0–1).

Note: FinBERT and AlphaVantage provided only Polarity and Intensity, while LLMs provided all five dimensions.

Feature Engineering

Article-level scores were aggregated to the weekly level using relevance-weighted means. Additionally, the authors constructed:

Dispersion Features: Standard deviation of polarity and uncertainty within a week (capturing market disagreement).
Momentum Features: Week-on-week changes (first differences) in Polarity, Uncertainty, and Forwardness (capturing sentiment trends).
Total Features: 31 candidate independent variables.

Prediction Model

Algorithm: LightGBM (Gradient Boosting), chosen for its robustness with tabular data and ability to handle feature scale differences.
Hyperparameter Tuning: Optimized using Optuna (Tree-structured Parzen Estimator) via 5-fold time-series cross-validation.
Evaluation Metrics:
- AUROC: Area Under the Receiver Operating Characteristic Curve.
- Accuracy: Proportion of correct directional predictions.
- Information Coefficient (IC): Rank correlation between predicted probabilities and realized returns.

Interpretability

SHAP (SHapley Additive exPlanations): Used to quantify the marginal contribution of each feature to the model's output, identifying global feature importance and local directional effects.

3. Key Results

Model Performance

Baseline: The AlphaVantage-only model performed poorly (AUROC: 0.569), confirming that simple proprietary polarity scores have limited predictive power.
Traditional vs. LLM: Adding FinBERT to the baseline improved performance (AUROC: 0.620).
LLM Performance:
- GPT-4o alone achieved the highest IC (0.249) and strong AUROC (0.634), outperforming Llama 3.2-3b.
- Llama 3.2-3b showed less stable scoring and lower predictive power when used alone.
- Combination (GPT-4o + FinBERT): This hybrid approach yielded the best overall AUROC (0.6515). This suggests that LLMs and traditional financial sentiment models provide complementary information; FinBERT captures nuances that GPT-4o misses, and vice versa.
- LLM Ensemble (GPT + Llama): Combining the two LLMs did not improve upon GPT-4o alone, suggesting Llama 3.2-3b introduced noise rather than signal.

Feature Importance (SHAP Analysis)

Beyond Polarity: The most important predictors were not simple polarity.
- Top Feature: gpt_intensity_mean (Sentiment Intensity).
- Second Feature: finbert_polarity_std (Dispersion of polarity across articles).
- Third Feature: gpt_uncertainty_momentum (Trend in uncertainty).
Key Insight: The predictive value of news extends beyond whether the news is "good" or "bad." Intensity (how strongly the market feels) and Uncertainty (ambiguity levels) are critical drivers.
Directional Effects:
- High intensity is associated with positive returns (up-weeks).
- High levels of uncertainty push predictions toward down-weeks, but rising uncertainty (momentum) is associated with up-weeks, indicating that the change in ambiguity carries distinct information.

4. Key Contributions

Multi-Dimensional Framework: Demonstrates that decomposing sentiment into Relevance, Intensity, Uncertainty, and Forwardness provides superior predictive power compared to traditional binary polarity.
Hybrid Model Superiority: Establishes that combining LLM-based semantic extraction (GPT-4o) with domain-specific supervised models (FinBERT) creates a more robust forecasting system than using either alone.
Uncertainty as a Signal: Provides empirical evidence that uncertainty and intensity are distinct, high-value predictors for commodity returns, challenging the dominance of polarity-based metrics in financial literature.
Practical Application: Offers a workflow for integrating unstructured news data into commodity risk monitoring and trading strategies.

5. Significance and Limitations

Significance: The paper bridges the gap between NLP advancements and financial economics, showing that LLMs can extract economically relevant, non-linear signals from text that traditional econometric models miss. It supports the use of multi-dimensional sentiment for better risk management in energy markets.
Limitations:
- Scope: Focused on a single commodity (WTI) and a single news source.
- Frequency: The weekly horizon may not generalize to high-frequency trading.
- Model Selection: The choice of GPT-4o and Llama 3.2-3b was based on a cost-performance tradeoff; future work should explore other model architectures.
- Data: Did not incorporate alternative data sources like satellite imagery or shipping data.

Conclusion: The study concludes that multi-dimensional LLM-based sentiment measures significantly improve the forecasting of commodity returns, with intensity and uncertainty being the most critical dimensions, validating the need to move "Beyond Polarity" in financial sentiment analysis.