FinBloom: Knowledge Grounding Large Language Model with Real-time Financial Data

In one sentence

Summary

The paper addresses the critical limitation of static Large Language Models (LLMs) in finance, where reliance on pre-trained weights leads to hallucinations and outdated information when handling high-velocity, real-time data. The authors argue that traditional Retrieval-Augmented Generation (RAG) systems, which embed unstructured text into vector spaces, are ill-suited for financial data due to its structured, tabular nature and the precision required for metric-specific queries. To solve this, they propose a knowledge-grounding architecture that keeps the generator LLM frozen while using a specialized agent to handle data retrieval and context construction. The methodology centers on three key components: a custom 50,000-sample Financial Context Dataset, the FinBloom 7B model, and a Financial Agent pipeline. FinBloom 7B is fine-tuned on Bloom 7B using 14 million news articles from Reuters and DPA, plus a 25% sample of 12 million SEC filings. The Financial Agent, derived from FinBloom 7B, interprets user queries to generate structured data requests (SQL-like) against a Data Module containing real-time tabular and textual repositories. This approach avoids embedding tables into vectors, instead using schema-aligned query prediction to fetch exact values, which are then linearized and appended to the query for a larger LLM (e.g., ChatGPT) to synthesize the final answer. Experiments demonstrate that this agent-based grounding significantly outperforms standard LLMs with web access, which often provide approximate or incorrect figures for specific financial metrics. The system reduces latency by eliminating manual data provision and enhances accuracy by ensuring the generator LLM receives precise, up-to-date context. The authors release the Financial Context Dataset, the FinBloom 7B model, and the Financial Agent weights. Limitations include the inability to release raw news data due to contractual restrictions and the reliance on a separate large LLM for final synthesis, which may introduce its own latency or cost overheads compared to a fully self-contained model.

Research problem

Static LLMs fail in finance because they lack access to real-time, high-velocity data, leading to hallucinations and outdated answers. Traditional RAG systems struggle with structured financial data (tables, time series) because dense-vector embedding over textified tables is imprecise and computationally expensive for frequent updates. The problem is to ground LLMs in real-time financial data efficiently without retraining the main model.

Core contributions

• Development of a 50,000-sample Financial Context Dataset with queries paired with required structured and textual context.
• Creation of FinBloom 7B, a domain-specific LLM fine-tuned on 14 million news articles and 12 million SEC filings.
• Design of a Financial Agent that generates structured data requests to retrieve precise metrics from a Data Module.
• Proposal of a knowledge-grounding framework that avoids dense-vector RAG for structured data, using schema-aligned query prediction instead.
• Demonstration that agent-based grounding improves accuracy and reduces latency compared to LLMs with web access.

Method

• Template-based generation of 50,000 financial queries with associated required data and structured requests.
• Fine-tuning Bloom 7B on 14M news articles and 12M SEC filings to create FinBloom 7B.
• Further fine-tuning FinBloom 7B on the Financial Context Dataset to create the Financial Agent.
• Implementation of a Data Module with real-time tabular and textual databases.
• Agent uses schema-aligned query prediction to fetch exact data from the Data Module.
• Linearization of retrieved structured data into text for concatenation with the user query.
• Final answer generation by a larger, frozen LLM (e.g., ChatGPT) using the enriched input.

Results & findings

• FinBloom 7B is fine-tuned on massive financial text corpora to understand domain-specific language.
• The Financial Agent successfully translates natural language queries into structured data requests.
• Retrieved data is linearized and appended to the query, enabling the generator LLM to provide precise answers.
• Standard LLMs with web access often provide approximate or incorrect figures for specific quarterly metrics.
• The proposed method eliminates the need for users to manually provide accurate data, reducing latency.
• The approach is designed to handle high-velocity data flows, making it suitable for real-time decision support.

Data & experiments

• Financial Context Dataset: 50,000 queries generated from templates based on retail investor and advisor interactions.
• FinBloom Training Data: 14 million news articles (Reuters, DPA) and 25% sample of 12 million SEC filings.
• Data Module: Real-time repositories of tabular financial metrics and textual news.
• Evaluation: Comparison of ChatGPT with web access vs. ChatGPT enriched with FinBloom-generated context.
• Metrics: Qualitative assessment of accuracy, hallucination reduction, and latency in answering specific financial queries.

Limitations & caveats

• Raw news data (14M articles) cannot be released due to contractual restrictions.
• Relies on a separate large LLM (e.g., ChatGPT) for final synthesis, adding cost/latency.
• Performance depends on the quality and coverage of the Data Module's real-time feeds.
• The Financial Agent's ability to generate correct structured queries may fail on ambiguous or novel queries.
• Evaluation is primarily qualitative; quantitative metrics on hallucination rates are not extensively reported.

Curator notes

Discovery status: round1_promoted_seed_for_round2. Source links: A Multimodal Foundation Agent for Financial Trading: Tool-Augmented, Diversified, and Generalist; BloombergGPT: A Large Language Model for Finance; CFGPT: Chinese Financial Assistant with Large Language Model; ConvFinQA: Exploring the Chain of Numerical Reasoning in Conversational Finance Question Answering; Empowering Many, Biasing a Few: Generalist Credit Scoring through Large Language Models; FinBen: A Holistic Financial Benchmark for Large Language Models; FinGPT: Open-Source Financial Large