Beyond Sentiment: Structured Information Extraction from Financial News
Combining LLM-extracted financial-news structure with FinBERT sentiment raised prediction performance to F1 0.600.
The paper tests 41,618 news-stock pairs from FNSPID and argues that sentiment leaves out event type, impact scope, time horizon, and confidence. Its LLaMA-3.1-70B extraction framework produced signals that often disagreed with sentiment, with a 53.5% systematic disagreement rate. FinBERT sentiment was strongest under nonlinear models, but adding the structured features significantly beat either source alone. Non-sentiment dimensions added +0.019 F1 beyond FinBERT in ablations. ArXiv · AI/CL/LG's note
The paper tests 41,618 news-stock pairs from FNSPID and argues that sentiment leaves out event type, impact scope, time horizon, and confidence. Its LLaMA-3.1-70B extraction framework produced signals that often disagreed with sentiment, with a 53.5% systematic disagreement rate. FinBERT sentiment was strongest under nonlinear models, but adding the structured features significantly beat either source alone. Non-sentiment dimensions added +0.019 F1 beyond FinBERT in ablations. ArXiv · AI/CL/LG's note
score 4