Financial sentiment analysis with FUNNEL: filtered UNion for NER-based ensemble labeling

Abstract

This paper introduces FUNNEL (Filtered UNion for NER-based Ensemble Labeling), a novel ensemble-based framework for labeling financial news that enhances the reliability of stock-specific sentiment signals. The framework integrates weak keyword-based heuristics with a transformer-based named entity recognition model (spaCy) through a weighted voting scheme, balancing precision and recall. Manual evaluation of 1,400 article–label pairs demonstrates that FUNNEL outperforms the original FNSPID labels in both accuracy and coverage. Applied across seven major companies, the framework reveals systematic differences in sentiment signals produced by FinBERT, RoBERTa, and VADER. These results indicate that integrating different labeling strategies enhances dataset reliability, coverage, and stability, providing a scalable framework for financial sentiment analysis.