Enhancing Trade Capture with Self-Correcting AI Workflows

The integration of large language models (LLMs) into business process automation is igniting high expectations, particularly in sectors requiring the handling of free-form, natural language content. According to NVIDIA, while achieving human-level reliability in these workflows has posed challenges, significant advancements are being made to enhance accuracy and efficiency.

AI in Trade Entry

Trade entry forms a critical part of financial 'what-if' analysis, where potential trades are evaluated for their impact on risk and capital requirements. Traditionally, trade descriptions are free-form and varied, making automation difficult. AI models like NVIDIA's NIM are being employed to interpret these descriptions and convert them into structured data compatible with trading systems.

For instance, a trade description might state, "We pay 5y fixed 3% vs. SOFR on 100m, effective Jan 10," describing an interest rate swap. The challenge lies in the absence of a predefined format, as the same trade can be described in multiple ways, necessitating a nuanced understanding by AI models.

Addressing AI Hallucinations

During NVIDIA's TradeEntry.ai hackathon, it was observed that LLMs can reach high accuracy with simple trade texts but struggle with complex inputs, leading to hallucinations where the model makes incorrect assumptions. A notable error involved the AI incorrectly adding a year to a trade's start date, highlighting the importance of context-aware processing.

To counteract these issues, NVIDIA proposes a self-correction approach, prompting the AI to produce a string template alongside a data dictionary that accurately reflects the input. This method ensures any additional logic, such as date interpretation, is handled in post-processing, significantly reducing errors.

Deploying AI Models

NVIDIA's NIM offers a platform for deploying AI models with low latency and high throughput, supporting a variety of model sizes. This flexibility allows users to balance accuracy and speed, with the self-correcting workflow demonstrating a 20-25% reduction in errors and improved F1-scores.

Through few-shot learning, where models are provided with example inputs and outputs, performance is further enhanced. Models specifically trained for reasoning, like DeepSeek-R1, show superior accuracy, particularly with richer prompting contexts.

Conclusion

The integration of self-correcting workflows in AI-based trade capture systems marks a significant advancement, reducing errors and enhancing accuracy. NVIDIA encourages the adoption of this approach in financial workflows, leveraging their model APIs for local deployment.

For more insights into AI applications in financial services, NVIDIA invites industry professionals to attend the GTC Paris event, offering sessions on generative AI and its deployment in production environments.