Integrating Economic Theory, Domain Knowledge, and Social Knowledge into Hybrid Sentiment Models for Predicting Crude...

Area of Science:

• Financial sentiment analysis
• Commodity market prediction
• Natural Language Processing (NLP) for finance

Background:

• Traditional sentiment analysis models lack domain-specific knowledge for accurate commodity market prediction.
• Predicting crude oil prices requires understanding supply/demand dynamics beyond standard affective models.

Purpose of the Study:

• To introduce CrudeBERT and CrudeBERT+, domain-specific affective models for the crude oil market.
• To enhance sentiment analysis by incorporating economic theory and external knowledge sources.
• To evaluate the predictive performance of these models against existing benchmarks.

Main Methods:

• Developed CrudeBERT and CrudeBERT+ by integrating economic theory, entity knowledge, and Google Trends data.
• Utilized dynamic time warping to compare model predictions with WTI crude oil futures prices.
• Conducted experiments using news headlines and price data from January 2012 to April 2021.

Main Results:

• CrudeBERT+ demonstrated superior performance in approximating WTI crude oil futures price movements.
• Outperformed established models like RavenPack, BERT, and FinBERT over a 9-year period.
• Achieved better results in most individual years analyzed, highlighting consistent predictive power.

Conclusions:

• Domain-specific affective models integrating economic theory and external knowledge significantly improve financial sentiment analysis.
• CrudeBERT+ shows strong potential for enhancing decision-making in the volatile crude oil market.
• The study validates the benefits of combining NLP with economic principles for market forecasting.