Context‍

U.S. exchanges and equity market participants are required to produce reports for regulators like the SEC, detailed disclosures that analyze execution quality and order routing practices across massive volumes of trading activity.

For this stock exchange, generating these reports required analyzing more than 35TB of execution and market data spread across dozens of internal and external systems. Producing the required reports relied heavily on manual workflows and static reporting (e.g., large PDF and CSV files), with large teams of human analysts spending weeks aggregating data, validating calculations, and ensuring compliance.



The exchange needed a system capable of interpreting massive trading datasets rapidly while replicating the reasoning of its top regulatory analysts. The goal was to do this repeatedly, instantly, and with a high degree of accuracy.

Solution‍

They used Hyde's platform to build a specialist regulatory AI model capable of reasoning through historical data and generating future SEC reports.

Environment setup: The first step was collecting training data containing historical reports and common analysis workflows. These datasets captured the analytical reasoning and calculation patterns used by regulatory analysts when interpreting execution data. The system also connected with their data warehouse tables and analytical system APIs to understand the environment used by regulatory analyses.

Preparing training data: This data was synthetically multiplied using reinforcement learning techniques to 75,000+ unique model runs under a wide range of market conditions and analysis needs. The exchange's subject matter experts helped evaluate these synthetic trajectories to create a human-driven definition of ‘correctness’.



Model training: This data was used to post-train a model that functions as the exchange’s AI regulatory analyst. Because it understands the nuances of this exchange’s data landscape and analysis needs, it can outperform frontier models trained on generic open source information.

Continuous RL: Performance continually improves through an ongoing reinforcement learning loop, where usage patterns and analyst feedback automatically become new training signals, allowing the system to adapt as market conditions and regulatory requirements evolve.

Impact‍

The system transformed how the stock exchange manages regulatory analyses. The model now generates complex queries and detailed workflows with near complete accuracy, eliminating weeks of manual analysis work while minimizing operational risk of hallucination in the critical regulatory domain.



Improved visibility into execution performance has also strengthened commercial negotiations generating incremental revenue through improved payment-for-order-flow economics. Beyond internal use, the platform has become a strategic asset, providing the opportunity for the exchange to offer the product to equity market participants and brokerages. This represents an additional $20M+ potential revenue opportunity.

Takeaway‍

Regulatory AI agents have the ability to reason across large financial datasets and provide near-instant reports and analyses on demand. However, frontier models are trained to be generalists, not to understand the nuances of specific enterprise environments. For high-stakes functions like regulatory analysis and reporting, specialist AI models must be trained to meet the accuracy thresholds. 



Your regulatory AI must produce precise outputs every time.