Source-linked AI summary
FinGPT: Open-Source Financial Large Language Models
Hongyang Yang, Xiao-Yang Liu, Christina Dan Wang
TL;DR
Financial LLM development is constrained by access to high-quality data and by proprietary models’ limited transparency. FinGPT responds with an open-source, data-centric framework using automated curation and lightweight adaptation, and reports a scalable foundation for financial sentiment analysis.
Problem
High-quality financial data is difficult to access, while proprietary models restrict transparency and accessibility for developing FinLLMs.
Method
FinGPT uses an open-source, data-centric framework centered on financial-data curation, preprocessing, real-time processing, and lightweight LoRA adaptation.
Results
FinGPT provides a scalable and effective foundation for financial sentiment analysis.
Takeaways & Limitations
FinGPT aims to democratize financial data and FinLLMs through accessible resources, customization, and open-source collaboration.
Takeaways & Limitations
The released code is for academic purposes and is not financial advice or a recommendation to trade real money.
Abstract
from arXiv · showhide
Large language models (LLMs) have shown the potential of revolutionizing natural language processing tasks in diverse domains, sparking great interest in finance. Accessing high-quality financial data is the first challenge for financial LLMs (FinLLMs). While proprietary models like BloombergGPT have taken advantage of their unique data accumulation, such privileged access calls for an open-source alternative to democratize Internet-scale financial data. In this paper, we present an open-source large language model, FinGPT, for the finance sector. Unlike proprietary models, FinGPT takes a data-centric approach, providing researchers and practitioners with accessible and transparent resources to develop their FinLLMs. We highlight the importance of an automatic data curation pipeline and the lightweight low-rank adaptation technique in building FinGPT. Furthermore, we showcase several potential applications as stepping stones for users, such as robo-advising, algorithmic trading, and low-code development. Through collaborative efforts within the open-source AI4Finance community, FinGPT aims to stimulate innovation, democratize FinLLMs, and unlock new opportunities in open finance. Two associated code repos are https://github.com/AI4Finance-Foundation/FinGPT and https://github.com/AI4Finance-Foundation/FinNLP
1 Introduction
Financial LLMs require high-quality, current data, but proprietary models limit access and transparency. FinGPT addresses this gap with an open-source, data-centric, full-stack framework for financial applications.
- High-quality, relevant, and up-to-date data is critical for developing effective financial large language models.
- Financial data is highly time-sensitive and dynamic because market-moving information creates narrow opportunities and arrives continuously.
- Proprietary models such as BloombergGPT rely on exclusive data access, motivating a more open and transparent alternative.
- FinGPT introduces an end-to-end open-source framework that emphasizes financial-data acquisition, cleaning, and preprocessing.
- Its five-layer framework spans data sources, data engineering, LLMs, tasks, and applications, supporting real-time processing, evaluation, and practical demonstrations.
- FinGPT aims to democratize financial data and FinLLMs while fostering collaboration, customization, and innovation in open finance.
2 Related Work
Prior financial LLM work includes broad financial text applications and proprietary systems such as BloombergGPT. FinGPT offers an open-source alternative built around adaptation, personalization, and community access.
- Financial LLMs have been applied to predictive modeling and financial text analysis using news, earnings calls, and social-media data.
- BloombergGPT demonstrated a financial LLM trained on mixed financial and general data, but access limitations and training costs motivate low-cost domain adaptation.
- FinGPT uses reinforcement learning from human feedback to adapt financial assistants to individual preferences.
- Open-source FinLLMs promote broad access, transparency, research progress, and education through publicly available technology and code.
- Community collaboration can strengthen infrastructure for financial text data and support the long-term durability and effectiveness of models.
3 Overview of FinGPT: An Open-Source Framework for FinLLMs
FinGPT is an open-source framework that connects financial data collection, processing, model adaptation, tasks, and applications. It combines real-time data engineering with lightweight adaptation and market-informed techniques.
- FinGPT is organized around data source, data engineering, LLM, and application components that maintain framework functionality and adaptability.
- The data source layer gathers extensive financial information from news, social media, financial statements, market trends, and other online sources.
- Real-time data engineering cleans and processes financial text to address temporal sensitivity and low signal-to-noise ratio.
- The tasks layer executes sentiment analysis, summarization, and numerical reasoning while providing standardized metrics for evaluation and comparison.
- LoRA fine-tunes pretrained LLMs on financial data using lightweight adaptation, reducing the cost of updating models for dynamic financial information.
- RLSP uses stock-price changes as feedback to associate news sentiment with subsequent stock performance and refine market-response predictions.
- RAG combines context retrieval with LLM generation so responses are informed by retrieved documents and relevant input context.
- FinGPT demonstrates applications including sentiment analysis, robo-advising, quantitative trading, and portfolio optimization.
4 Data-Centric Approach for FinLLMs
FinGPT prioritizes financial-data collection, preparation, and processing because diverse sources differ in timeliness, reliability, variability, and market relevance. Integrating these sources supports broader financial-market understanding and decision-making.
- FinGPT’s data-centric approach treats financial-data collection, preparation, and processing as central to successful FinLLMs.
- Financial news provides timely, dynamic, and market-influential information about economies, industries, companies, and financial events.
- Company filings and announcements offer granular, regulated, periodic information about corporate finances and strategy.
- Financial social-media discussions capture real-time public sentiment but vary widely in tone, quality, and volatility.
- Finance-oriented websites and forums provide analyst perspectives, market sentiment, and broad coverage across securities and market segments.
- Integrating diverse financial data types can facilitate comprehensive market understanding and effective financial decision-making.
5 Experiments: Financial Sentiment Analysis
FinGPT’s sentiment-analysis experiment combines market-driven labeling with lightweight LoRA adaptation and reinforcement learning on stock prices. It outperforms the listed baselines and shows stronger market-aligned financial reasoning.
- 5.1 Dataset: Over 620,000 cleaned financial news headlines from 2016–2024 receive sentiment labels based on subsequent stock-price movements.Sources include CNBC, Reuters, Yahoo Finance, and MarketWatch, collected through the FinNLP pipeline.
- 5.2 Model and Training Setup: FinGPT uses a two-stage adaptation process combining LoRA-based supervised fine-tuning with reinforcement learning on stock prices.RLSP uses the financial market as the environment and post-news stock-price reactions as rewards.
- 5.2 Model and Training Setup: LoRA introduces approximately 8.3M trainable parameters, well below 0.1% of the original 8B-parameter model.The configuration uses rank r = 8 and scaling factor α = 16.
- 5.5 Results: FinGPT outperforms FinBERT, BloombergGPT, ChatGPT zero-shot, and Llama3.1-8B zero-shot on the reported sentiment-analysis evaluation.The evaluation includes accuracy, class-level precision, recall, F1-score, Macro-F1, and binary-subset AUC.
- 5.5 Results: The ablation study finds that LoRA performs most of the heavy lifting, while RLSP further improves alignment with market behavior.The reported discussion attributes the overall gains to data-centric labeling and RLSP reinforcement alignment.
- 5.5 Results: In a Tesla pricing case, FinGPT predicts negative sentiment like human annotation, while base Llama3 predicts neutral.RLSP produces stronger alignment with the subsequent price reaction than the supervised-fine-tuned model.
- 5.6 Discussion: Market-driven labels improve real-world applicability, LoRA reduces adaptation cost by ∼1000×, and RLSP incorporates financial-market feedback.The experiment is presented as a scalable and effective foundation for financial sentiment analysis.
6 Conclusion
FinGPT addresses financial-sector LLM challenges by adapting existing LLMs to specific financial applications. The conclusion emphasizes reduced adaptation costs alongside the sector’s temporal sensitivity and noisy data.
- 6 Conclusion: Financial applications of LLMs must address high temporal sensitivity, a dynamic financial landscape, and a low signal-to-noise ratio in financial data.The conclusion frames these conditions as requiring efficient solutions.
- 6 Conclusion: FinGPT leverages pre-existing LLMs and fine-tunes them for specific financial applications.The conclusion presents this strategy as its response to the identified financial-sector challenges.
- 6 Conclusion: This approach significantly reduces adaptation costs while supporting financial-sector applications.
7 Future Work
Future FinLLM development will emphasize open standards, efficient customization, unified data curation, and transparent workflows for reliable financial AI systems. The paper also limits its shared code to academic purposes and explicitly disclaims financial advice.
- Future FinLLM development will focus on establishing open, industry-level standards for financial large language models.
- LoRA and QLoRA are identified as parameter-efficient methods for low-cost, domain-specific customization across diverse financial institutions.
- A unified data curation pipeline is intended to promote high-quality, standardized financial datasets for streamlined training and evaluation.
- Open-source tooling, reproducible benchmarks, and transparent workflows aim to support reliable, scalable, and interoperable financial AI systems.
- The shared code is provided for academic purposes, and the paper states that it is not financial advice or a recommendation to trade real money.