🤖Webinar - From Theory to Practice: Real-World Applications of AI in Asset-Based Finance - Register here

Exploring the fine-tuning of LLMs for enhanced performance in ABF & future use cases

After delving into ethical and regulatory considerations regarding the use of Large Language Models (LLMs) in financial services, it’s time to shift our attention to their strategic implementation in Asset-Based Finance (ABF). While this implementation promises a significant advancement for every market participant, its success hinges on factors such as the sophistication of implementation, the availability of high-quality data, and the expertise of users in interpreting and applying the model’s output.

What does it mean to fine-tune LLMs for ABF?

Fine-tuning LLMs for ABF means adjusting these sophisticated models to perform specific tasks within the sector. It’s not just about making them more accurate; it’s about matching AI skills with the specific needs of asset financing. Models like FinBERT and InvestLM show how fine-tuning with industry data can really make a difference in tasks like credit scoring and understanding market feelings.

How is Cardo AI pioneering custom LLM use cases in ABF?

Cardo AI stands at the forefront of integrating LLMs into Asset-Based Finance, demonstrating the practical benefits and innovative potential of these technologies. Our focused approach showcases how custom-built LLMs can solve complex financial problems with precision:

Automatic categorization of companies’ economic sectors

This use case focuses on categorizing small and medium-sized businesses into appropriate economic sectors as defined by the North American Industry Classification System (NAICS).

This process is especially valuable for lending institutions that manage extensive loan portfolios with limited business data from web and official sources. Our innovative tool is designed to identify the right NAICS category quickly and accurately for each business.

Our approach is the following:

1. Harnessing Transformer Language Model

We have developed a method that utilizes a transformer language model to link a company and its description to the financial sectors it operates in. One of the standout features of this model is its ability to function across 24 different languages, leveraging the multilingual capabilities of the foundational pre-trained model.

2. Creating a Reliable Data-Driven Solution

To ensure the model’s effectiveness, we combined the LLM with our analysts’ expertise to design specific prompts that helped generate the necessary dataset for training. We rigorously tested the model against manually annotated data, meticulously prepared in-house at Cardo AI. This testing was crucial in confirming the model’s accuracy in sector assignment.

3. Optimizing Performance with Fine-Tuning

Our infrastructure enabled us to evaluate various language models, selecting the most effective one as our baseline. We further refined this model with task-specific fine-tuning to enhance its performance. By integrating tailored prompts with data from the NAICS sector taxonomy, we created approximately 24,000 data points (evenly distributed across the 24 languages) for training purposes.

4. Human-Generated Data for Precision

Cardo AI’s business analysts dedicated weeks to labeling additional company description-sector pairs and maintaining best practices in model evaluation. This human-generated data ensured that the model’s performance was grounded in real-world accuracy.

Future use cases: Expanding LLM and GenAI applications in ABF

Below, we have described some of the initiatives we are considering as possible enablers for faster answers, higher transparency, and more accurate information using various sources and data structures in the ABF space.

  • Calendar reminders for structured Deals

Utilizing RAG frameworks and LLMs to automate the extraction of critical dates from ABF transaction documents promises to streamline operations significantly. This application aids all parties involved, from buyers and sellers to servicers, ensuring timely compliance and operational efficiency.

  • Covenant quality indicator for ABF transactions

Drawing inspiration from Moody’s Covenant Quality Indicator, this proposed use case involves using GenAI to analyze and categorize covenants in ABS transactions. It aims to provide a benchmark for covenant quality, offering a transparent and standardized assessment tool that can foster trust and reliability in the market.

  • Data Chatbots

Envisioned to build ad-hoc analytics and provide explainable narratives through natural language requests, these chatbots could revolutionize data accessibility in ABF. By integrating with user databases, they would enable users to swiftly access complex data across multiple platforms, simplifying the analysis and decision-making process.

What’s next for LLMs in Asset-Based Finance?

The integration of LLMs into the financial and ABF industries highlights a transformative potential moderated by challenges. Success stories from both sectors underscore the critical role of domain-specific fine-tuning, ethical considerations, and regulatory compliance in leveraging LLM capabilities.

However, as the landscape evolves, continuous innovation, adherence to ethical standards, and robust risk management strategies will be essential prerequisites to realize the full potential of LLMs in these industries. Future endeavors must prioritize precise data handling, model transparency, and the mitigation of inherent risks to ensure the sustainable and responsible growth of LLM applications in financial services and asset-based finance.


To read the full version of our e-book “Exploring the Role and Potential of Large Language Models in the Asset-Based Finance Sector”, claim your free copy HERE.