Jesse Ellis
June 4, 2025 16:03
Explore the integration of AI and rules -based error modifications in the trade capture workflow to increase the accuracy and efficiency of financial analysis.
Integrating LLMS (Large Language Model) into business process automation, especially in the sector, which requires free form of natural language contents. According to NVIDIA, this workflow has achieved a challenge while achieving the reliability of the human level, but there is a significant development to improve accuracy and efficiency.
AI of trade entry
Trade entry forms an important part of the ‘What-IF’ analysis, where potential transactions are evaluated on the effects of risks and capital requirements. Traditionally, trade descriptions are free and diverse, making automation difficult. AI models, such as NVIDIA’s NIM, interpret these explanations and convert them into structured data compatible with trading systems.
For example, the trade description describes the interest rate swap and says, “We pay 5Y fixed 3% to 100m from January 10.” The challenge requires subtle understanding by the AI model because the same deal can be explained in many ways when there is no pre -defined format.
AI hallucinations resolution
During NVIDIA’s tradeness.ai Hackathon, LLM can reach high accuracy with simple trade text, but it is difficult with complex inputs and leads to hallucinations that make the model wrong. Notable errors emphasized the importance of processing that AI recognized the situation by adding a year wrongly on the start of the trading.
To cope with this problem, NVIDIA proposes its own conquest approach to create a string template with the data dictionary that accurately reflects the input. This method is treated with additional logic such as date interpretation, which greatly reduces errors.
AI model distribution
NVIDIA’s NIM provides a platform that supports a variety of models by distributing low -waiting time and high -handling AI models. This flexibility allows users to balance accuracy and speed, and their own modified workflows have reduced errors by 20-25% and improved F1-score.
The performance is further improved through a small number of learning that provides example input and output to the model. The models trained specifically for reasoning such as DeepSeek-R1 show excellent accuracy, especially with rich prompt contexts.
conclusion
Integrating its own adjustment workflow in the AI -based trade capture system leads to significant developments to reduce errors and improve accuracy. NVIDIA is recommended to adopt this approach from the Financial Workflow to use the model API for local deployment.
To gain insight into AI applications in the financial service sector, NVIDIA invites industry experts to attend the GTC Paris event to provide sessions for deployment in the creation AI and production environment.
Image Source: Shutter Stock