FinDialogLens extracts missed trades from financial chatrooms with 94.3% accuracy

Researchers propose FinDialogLens, a hybrid pipeline that combines compact fine-tuned classifiers and an LLM-powered Trade Engine to extract Request-for-Quote (RFQ) events and their final prices and trade outcomes from multi-party financial chatrooms.

The system first detects RFQ-trigger messages and metadata using fine-tuned classifiers, then segments per-RFQ windows with an RFQ-Level Module, and finally fills argument roles via the Trade Engine, which uses GPT-4o as its backbone.

On a held-out evaluation, FinDialogLens achieved 92.1% accuracy on final price extraction and 94.3% accuracy on trade outcome extraction, outperforming full-chatroom chain-of-thought prompting baselines.

Fine-tuned open-source LLMs as small as 3 billion parameters matched FinDialogLens performance when trained on modest in-domain data, indicating feasibility for on-premise deployment.

To reduce cost at scale, a difficulty-aware router allocates RFQs between a low-cost rule-based engine and the LLM-powered Trade Engine, cutting LLM calls by 85% on final price extraction while recovering roughly half the accuracy gap to the full LLM pipeline.

At a deployment scale of 70,000 RFQs per day, the router saves over $300 per day in LLM inference costs while maintaining high extraction accuracy.