Opening article
Capital Fund Management used Hugging Face’s LLM ecosystem to improve financial named entity recognition. By generating synthetic labels with Llama 3.1‑70B‑Instruct via Inference Endpoints and refining them in Argilla, the team increased NER accuracy by up to 6.4% and cut costs to as low as 80× cheaper than using large LLMs alone. The process involved selecting Benzinga headlines, creating a prompt that instructs the model to extract company names, and processing 900k samples in about eight hours for roughly $70.
Overview: This article presents a deep dive into Capital Fund Management’s (CFM) use of open-source large language models (LLMs) and the Hugging Face (HF) ecosystem to optimize Named Entity Recognition (NER) for financial data. By leveraging LLM-assisted labeling with HF Inference Endpoints and refining data with Argilla , the team improved accuracy by up to 6.4% and reduced operational costs, achieving solutions up to 80x cheaper than large LLMs alone.
This structured approach combines accuracy and cost-effectiveness, making it ideal for real-world financial applications.
Capital Fund Management (CFM) is an alternative investment management firm headquartered in Paris, also has teams in New York City and London currently overseeing assets totaling 15.5 billion dollars.
CFM has been working with Hugging Face's Expert Support to stay updated on the latest advancements in machine learning and harness the power of open-source technology for their wide range of financial applications. One of the collaboration's main goals has been to explore how CFM can benefit from efficiently using open-source Large Language Models (LLMs) to enhance their existing machine learning use cases. Quantitative hedge funds rely on massive amounts of data to inform decisions about whether to buy or sell specific financial products. In addition to standard data sources from financial markets (e.g., prices), hedge funds are increasingly extracting insights from alternative data, such as textual information from news articles. One major challenge in incorporating news into fully automated trading strategies is accurately identifying the products or entities (e.g., companies, stocks, currencies) mentioned in the articles. While CFM’s data providers supply these tags, they can be incomplete and require further validation.
CFM explored several approaches to improve financial entity recognition, including zero-shot NER using LLMs and smaller models, LLM-assisted data labeling with Hugging Face Inference Endpoints and Argilla, and fine-tuning smaller models on curated datasets. These approaches not only leverage the versatility of large models but also address the challenges of cost and scalability in real-world financial applications.
Few-shot Examples : A series of input-output examples to demonstrate the expected behavior and ensure consistency in performance across varied inputs. These examples help the model understand how to distinguish relevant entities and format its output correctly. Each example illustrates different headline structures to prepare the model for a range of real-world cases.
Employing a scientific approach to finance, CFM leverages quantitative and systematic methods to devise superior investment strategies.
Among open-source models, the Llama 3.1 series by Meta stood out due to its strong performance across benchmarks, making it a top choice for generating synthetic annotations. These LLMs were pivotal in creating high-quality labeled datasets, combining automation and human expertise to streamline the labeling process and enhance model performance in financial NER tasks.
Our focus on this use case was to extract company names from news headlines from the Financial News and Stock Price Integration Dataset (FNSPID) . It is composed of news headlines and articles associated with corresponding stock symbols coming from several sources such as Bloomberg, Reuters, Benzinga and others. After analyzing the various news sources, we found that news from Benzinga had no missing stock symbol values. This subset of the dataset contains ~900k samples. As a result, we decided to reduce our dataset to Benzinga headlines for a more consistent and reliable analysis.
Demonstrates that open‑source LLMs can efficiently generate high‑quality training data for domain‑specific tasks, lowering costs and improving performance in finance and other data‑intensive sectors.