FinBERT + Claude: Multi-Model Pipelines for Sentiment Analysis

General-purpose LLMs are remarkably capable, but in financial NLP, specialized models still have a role. FinBERT — fine-tuned on financial text — brings domain-specific calibration that general models lack. Claude brings reasoning, synthesis, and instruction-following that FinBERT can't do. The combination is more robust than either alone.

The Pipeline

Our sentiment analysis pipeline runs in two stages. First, FinBERT scores each financial text segment (earnings call transcripts, news, filings) on a positive/neutral/negative axis with confidence scores. Second, Claude receives the raw text plus FinBERT's scores as context, and synthesizes a structured analysis: key themes, sentiment drivers, and flags for unusual language patterns.

Why Not Just Claude?

Cost and latency at scale. Processing thousands of earnings call segments per day with a frontier model is expensive. FinBERT runs locally, processes fast, and provides a strong prior. Claude handles the small fraction of segments where FinBERT confidence is low or where deeper synthesis is needed.

Why Not Just FinBERT?

FinBERT gives you a score, not an explanation. When a risk committee asks "why did our system flag this transcript as negative?", a confidence score isn't an answer. Claude provides the reasoning layer.

Results

On our internal benchmark, the hybrid pipeline outperforms either model alone on precision for high-conviction signals — which is the metric that matters in production financial systems.