FinGPT: Open-Source Financial Large Language Models

In one sentence

Summary

The paper addresses the critical need for accessible, transparent, and cost-effective financial large language models (FinLLMs) by introducing FinGPT, an open-source framework. It highlights the unique challenges of the financial domain, including high temporal sensitivity, constant dynamism, and low signal-to-noise ratios, which make proprietary models like BloombergGPT inaccessible to many researchers. FinGPT democratizes access by providing a full-stack, data-centric pipeline that emphasizes rigorous data curation, real-time processing, and lightweight adaptation techniques, aiming to foster innovation within the open-source AI4Finance community. The methodology centers on a five-layer architecture: data source, data engineering, LLMs, tasks, and applications. The data engineering layer implements a real-time pipeline for cleaning, tokenization, and embedding financial text from news, social media, and filings. For model adaptation, FinGPT employs Low-Rank Adaptation (LoRA) to fine-tune base models like Llama-3.1-8B with minimal trainable parameters, significantly reducing costs. It further introduces Reinforcement Learning on Stock Prices (RLSP), using short-term stock price movements as objective feedback signals to align model outputs with actual market reactions, replacing traditional human feedback with market-driven rewards. Experiments on financial sentiment analysis demonstrate that FinGPT outperforms baselines such as FinBERT, ChatGPT, and zero-shot Llama-3. The model achieves an accuracy of 82.1% and a Macro-F1 of 80.9%, significantly higher than FinBERT's 71.2% accuracy. Ablation studies confirm that LoRA provides substantial gains over base models, while RLSP further enhances market alignment. The paper showcases applications including robo-advising, quantitative trading signals, and risk management, while noting limitations in real-time deployment latency and the need for continuous data pipeline maintenance to handle market dynamism.

Research problem

The paper addresses the lack of accessible, transparent, and cost-effective financial large language models (FinLLMs). It highlights the challenges of high temporal sensitivity, market dynamism, and low signal-to-noise ratios in financial data, which make proprietary models like BloombergGPT inaccessible to most researchers and practitioners. The goal is to democratize FinLLMs through an open-source, data-centric framework.

Core contributions

• Introduction of FinGPT, an end-to-end open-source framework for building and customizing FinLLMs.
• Development of a real-time data curation pipeline that handles diverse financial data sources (news, social media, filings) with noise reduction and entity-aware embedding.
• Implementation of lightweight fine-tuning using Low-Rank Adaptation (LoRA) to drastically reduce computational costs compared to full pre-training.
• Proposal of Reinforcement Learning on Stock Prices (RLSP) as a novel alignment technique using market price movements as objective feedback signals.
• Creation of a comprehensive benchmarking suite of fundamental financial NLP tasks, including sentiment analysis, NER, and numerical reasoning.

Method

• Data Source Layer: Aggregates data from news APIs, social media, SEC filings, and trend platforms.
• Data Engineering: Real-time cleaning, tokenization, and entity-aware vector embedding into a vector database.
• LLM Layer: Utilizes base models (e.g., Llama-3.1-8B) with LoRA/QLoRA for parameter-efficient fine-tuning.
• RLSP Alignment: Uses stock price changes as reward signals in a reinforcement learning loop to align sentiment predictions with market impact.
• RAG Integration: Incorporates Retrieval-Augmented Generation for context-aware responses.
• Task Layer: Executes sentiment analysis, NER, summarization, and numerical reasoning.

Results & findings

• FinGPT (SFT+RLSP) achieves 82.1% accuracy and 80.9% Macro-F1, outperforming FinBERT (71.2% Acc, 69.9% Macro-F1).
• LoRA fine-tuning reduces trainable parameters from 6.17 billion to ~8.3 million, cutting adaptation costs by approximately 1000x compared to full fine-tuning.
• RLSP alignment improves Macro-F1 from 77.3% (LoRA only) to 80.9%, demonstrating the value of market-driven feedback.
• Case study shows FinGPT correctly identifies negative sentiment in 'Tesla cuts prices' news, whereas base Llama-3 predicts neutral.
• The self-labeled dataset aligns sentiment with true market reactions, avoiding costly manual annotation.
• FinGPT supports multiple base models including Llama, ChatGPT, and Claude via APIs.

Data & experiments

• Dataset: Over 620,000 cleaned financial news headlines from sources like CNBC, Reuters, and Yahoo Finance (2016-2024).
• Labeling: Market-driven self-labeling using short-term stock price changes to determine positive, negative, or neutral sentiment.
• Baselines: Comparison against FinBERT, BloombergGPT, ChatGPT (zero-shot), and Llama-3.1-8B (zero-shot).
• Metrics: Accuracy, Precision, Recall, F1-score, Macro-F1, and AUC for binary subsets.
• Ablation Study: Isolating the impact of LoRA fine-tuning and RLSP alignment on model performance.

Limitations & caveats

• Reliance on market price movements for labeling may introduce noise if prices are driven by non-news factors.
• Real-time pipeline requires significant infrastructure for low-latency data ingestion and processing.
• Performance is evaluated primarily on sentiment analysis; other tasks like numerical reasoning are less detailed.
• Open-source models may lack the proprietary data depth of BloombergGPT.
• Disclaimer notes that the code is for academic purposes and not financial advice.

Curator notes

Emphasizes data-centric pipelines and Low-Rank Adaptation fine-tuning.