AI news ·
FinDialogLens uses GPT-4o to extract trade events from financial chatrooms with 92.1% accuracy
FinDialogLens hits 92.1% accuracy on final price and 94.3% on trade outcome extracting trade events from financial chatrooms. Its difficulty-aware router slashes LLM calls by 85%, saving over $300 daily at 70,000 RFQs.
FinDialogLens, a hybrid LLM pipeline built to extract trade events from multi-party financial chatrooms, hits 92.1% accuracy on final price and 94.3% on trade outcome using GPT-4o. For sales and trading desks where missed signals in chat channels mean real P&L impact, the system identifies overlooked RFQs and fills argument roles through a dedicated Trade Engine without requiring full model runs on every message.
How the pipeline processes chatroom noise
The system uses compact fine-tuned classifiers as scaffolds. These lightweight models first detect RFQ triggers and price or trade outcome metadata, then segment RFQ windows per event. Only the segmented windows get passed to the larger model for argument role filling. This architecture keeps compute costs manageable at scale.
Fine-tuned open-source LLMs with as few as 3 billion parameters achieve comparable performance to much larger models on the task. The research team tested multiple model sizes and found that targeted fine-tuning on the specific event extraction domain closed the gap between small and frontier models.
Cost savings from difficulty-aware routing
A difficulty-aware router sits between the classifiers and the large language model. It decides whether a given extraction task requires GPT-4o or can be handled by a smaller model. This routing cuts LLM calls by 85% on final price extraction.
At a volume of 70,000 RFQs per day, the reduction saves over $300 daily in inference costs. For a trading floor processing high-frequency chat traffic across multiple desks, the annual savings become material without sacrificing the accuracy that compliance and trade reconciliation demand.
Open-source release and underlying tech
The paper, published on arXiv under cs.CL by Chin-Lun Fu, Hong Ni, and Behrouz Madahian, details the full pipeline. The code is available as open source. FinDialogLens targets a persistent problem in institutional trading: unstructured chat messages contain actionable trade intent that structured systems miss, and manual review does not scale.
GPT-4o serves as the primary reasoning engine, but the scaffold-and-route design means organizations can swap in different models as costs and capabilities shift. The modular approach also makes it easier to update individual components without retraining the entire system.
Why this matters for finance and sales professionals
Traders and salespeople who work chat channels for client orders know the pain of scrolling back through hundreds of messages to reconstruct a trade. FinDialogLens automates that extraction at high accuracy. For sales coverage teams, it surfaces missed opportunities directly from existing communication flows. For compliance officers, it creates an auditable, structured record of trade events that originated in informal channels - without requiring everyone to change how they communicate. Understanding how these extraction pipelines work is becoming part of the modern finance toolkit, and structured learning in Generative AI Courses or AI for Finance Courses can help teams evaluate where similar architectures fit their own workflows.