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

The paper introduces FinBloom, a 7-billion-parameter knowledge-grounded large language model fine-tuned on extensive financial news and SEC filings, which functions as a real-time financial agent capable of dynamically retrieving and integrating up-to-date text and tabular data to support high-velocity decision-making and algorithmic trading.

The Gist

Imagine you are trying to make a smart decision about buying a house. You ask a very smart, well-read friend (an AI) for advice.

The Problem: Your friend is incredibly knowledgeable about history, literature, and general facts. However, they haven't read a newspaper in two years. If you ask, "What is the price of houses in my neighborhood right now?" or "Did a new factory just open nearby?", your friend will guess based on old data. They might tell you prices are low, when in reality, they skyrocketed last week. In the fast-moving world of finance, this "stale information" is dangerous.

The Solution: The paper introduces FinBloom, a system designed to fix this. Think of it not just as a smart friend, but as a smart friend who has a super-powered research assistant standing right next to them.

Here is how the system works, broken down into simple parts:

1. The Two-Step Dance (The Architecture)

Instead of trying to make the "smart friend" (the main AI) memorize every single stock price and news headline in the world (which is impossible and slow), the system splits the job:

• The Research Assistant (The Financial Agent): This is a specialized robot trained specifically on finance. When you ask a question like, "Is Google a good investment right now?", this assistant doesn't try to answer immediately. Instead, it acts like a detective. It instantly figures out exactly what facts are missing: "I need Google's stock price, their earnings from last quarter, and any news from the last 24 hours."
• The Data Library (The Data Module): The assistant runs to the library, grabs the exact fresh numbers and news articles, and hands them to the smart friend.
• The Smart Friend (The Large Language Model): Now, the smart friend has the fresh facts in hand. They combine their general knowledge with the new data to give you a perfect, up-to-the-minute answer.

2. The "Training School" (The Datasets)

To make this system work, the authors had to teach the "Research Assistant" how to be a detective.

• The Financial Context Dataset: They created a massive textbook of 50,000 practice questions. For every question (e.g., "How is Apple doing?"), they wrote down exactly what data was needed to answer it. It's like giving the assistant a cheat sheet that says, "When someone asks about Apple, look up their stock price and revenue."
• FinBloom 7B (The Specialized Brain): They took a standard, smart AI model (called Bloom) and fed it a diet of 14 million financial news articles and millions of official company reports. This turned the general smart AI into a Finance Expert. It learned the language of money, the jargon, and how to read between the lines of a financial report.

3. Why This is Better Than Just "Googling"

You might ask, "Why not just let the AI browse the internet?"
The paper argues that standard internet browsing is like sending a tourist to a library to find a specific book. They might get lost, pick the wrong book, or find an outdated edition.

• Precision: The FinBloom system is like a librarian who knows exactly which shelf the book is on and pulls the exact page you need.
• Speed: In finance, seconds matter. This system is built to be fast, grabbing data instantly without the AI getting "distracted" by irrelevant news.
• Accuracy: It avoids "hallucinations" (making things up). Because the assistant pulls real numbers from a database, the AI is forced to base its answer on facts, not guesses.

The Big Picture

Think of FinBloom as a Financial Co-Pilot.

• Without it: You are flying a plane with a pilot who has a great map but hasn't seen the weather radar in years.
• With it: You have a pilot who is an expert, and a co-pilot who is constantly watching the live weather radar, feeding the pilot real-time updates so they can navigate safely through the storm.

The authors have released their "textbook" (the dataset) and their "co-pilot" (the AI model) to the public, hoping to help investors, analysts, and regular people make smarter decisions in a world where money moves faster than ever.

Technical Summary

1. Problem Statement

Large Language Models (LLMs) excel at natural language generation but struggle with interactive tasks requiring real-time information, particularly in high-velocity domains like finance.

• Static Knowledge Limitation: Standard LLMs are trained on static datasets. In finance, where asset prices, news, and regulatory filings change continuously, relying on "frozen" models leads to decisions based on stale data.
• Fine-Tuning Challenges: Continuously fine-tuning LLMs on real-time data is computationally expensive, time-consuming, and creates latency issues. Furthermore, it is difficult to verify what specific knowledge the model has learned or to ensure it hasn't hallucinated facts.
• Limitations of Traditional RAG: Standard Retrieval-Augmented Generation (RAG) systems often rely on dense vector embeddings of unstructured text. This approach is ill-suited for financial data, which is predominantly structured (tabular, time-series, numeric metrics). Converting structured tables into text chunks for vector search often leads to redundancy, loss of schema relationships, and retrieval errors (e.g., missing exact cell values).

2. Methodology

The authors propose a modular, agent-based architecture that keeps the core LLM frozen while using specialized modules to handle real-time data retrieval and context generation. The system consists of three main components:

A. The Data Module

This serves as the repository for real-time financial data, maintaining two distinct backends:

1. Tabular Database (Relational Store): Stores structured data (prices, financial metrics, statements) in a PostgreSQL database. It supports schema-aligned queries for precise retrieval of specific metrics for specific companies and date ranges.
2. Text Database (Vector Store): Stores the latest news articles in an Elasticsearch index. It uses a domain-adapted FinRoBERTa model (fine-tuned on financial news) to generate dense embeddings for semantic search, filtering by ticker, sector, and date.

B. The Financial Agent

This is the core innovation, a custom 7-billion parameter LLM (FinBloom 7B) fine-tuned to act as an intermediary.

• Function: It interprets natural language user queries and converts them into Structured Data Requests.
• Process: Instead of relying on vague semantic similarity, the Agent generates a formal request specifying:
  • Target Companies (normalized to internal IDs).
  • Required Metrics and Related Metrics (e.g., if asking for P/E ratio, it requests Share Price and EPS).
  • Date Ranges.
• Execution: The Agent sends this structured request to the Data Module, which retrieves exact data points. The Agent then linearizes this data into a textual context ($c$) and appends it to the original query ($x$) to form an enriched input $(c, x)$ for the final LLM.

C. The Large Language Model (LLM)

A larger, general-purpose LLM (e.g., GPT-3.5 or GPT-4) receives the enriched query $(c, x)$. Because the context is grounded in accurate, real-time data retrieved by the Agent, the LLM can generate precise, hallucination-free financial answers.

3. Key Contributions

1. Financial Context Dataset

The authors created a custom dataset containing 50,000 financial queries paired with the necessary contextual data.

• Construction: Derived from real-world queries via technology channels, financial advisor interviews, and retail investor interviews.
• Structure: Each entry includes the user query, "Required Data" (companies, metrics, dates), and a "Structured Data Request" (machine-readable format).
• Innovation: It includes "Related Metrics" to ensure robustness (e.g., providing raw data needed to calculate a ratio if the ratio itself isn't directly available).

2. FinBloom 7B (Domain-Specific LLM)

A custom 7B parameter model built by fine-tuning Bloom 7B on a massive, high-quality financial corpus:

• Training Data: Over 14 million financial news articles (Reuters and Deutsche Presse-Agentur) and a random 25% sample of 12 million SEC filings.
• Methodology: Used QLoRA (Quantized Low-Rank Adapters) for efficient fine-tuning.
• Performance: Outperformed other financial LLMs (FinMA, FinGPT, CFGPT) on the FinBen benchmark across 25 diverse tasks (NER, sentiment analysis, causal detection, etc.) without supervised fine-tuning on the benchmark datasets.

3. The Financial Agent System

The authors fine-tuned FinBloom 7B on the Financial Context Dataset to create the Financial Agent.

• Capability: The Agent successfully translates natural language queries into structured data requests with high accuracy (BLEU score of 0.96 and ROUGE-L of 0.97 on test sets).
• Efficiency: This approach eliminates the need for users to manually find data and significantly reduces latency compared to iterative web searches by general LLMs.

4. Results and Evaluation

• Benchmark Performance: On the FinBen benchmark, FinBloom 7B achieved the highest Average F1 Score (0.3464) among comparable open-source 7B models, outperforming FinMA 7B (0.3224) and FinGPT 7B-lora (0.2136).
• Agent Accuracy: The Financial Agent demonstrated superior ability to identify required data compared to a baseline Bloom 7B model fine-tuned on the same dataset, proving the value of pre-training on financial text corpora.
• Qualitative Comparison (Appendix D):
  • Scenario 1 (Success): When asked to compare Pepsi vs. Coca-Cola based on net income, the proposed system retrieved exact quarterly data, calculated profit margins, and provided a nuanced investment analysis.
  • Scenario 2 (Failure of Web-Enabled LLMs): When asked for 12 quarters of revenue for Reliance Industries or General Motors, standard ChatGPT (even with web access) provided hallucinated or approximate numbers, admitted inability to find data, or required manual user intervention.
  • Scenario 3 (Success with Agent): When asked about Magnus Retail's EBITDA, the Agent retrieved exact negative/positive values for the last 4 quarters, allowing the LLM to correctly identify a volatile trend and advise against immediate investment.

5. Significance and Conclusion

The paper addresses a critical gap in applying LLMs to high-stakes, data-driven fields like finance.

• Real-Time Decision Making: By decoupling the "reasoning" engine (LLM) from the "data retrieval" engine (Agent + Data Module), the system enables low-latency, high-accuracy financial decision-making.
• Handling Structured Data: The approach solves the specific challenge of retrieving structured financial data (tables/time-series) which traditional RAG systems fail to handle accurately due to reliance on unstructured text embeddings.
• Scalability: The modular design allows for updates to the data repository without retraining the LLM, making it suitable for high-velocity data flows.
• Future Impact: This framework paves the way for automated algorithmic trading, personalized financial advisory, and regulatory compliance tools that rely on up-to-the-minute data, transforming financial services by making them more efficient and accessible.