Paper: arXiv 2306.06031
Authors: Hongyang Yang, Xiao-Yang Liu, Christina Dan Wang
Abstract
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
Complexity vs Empirical Score
- Math Complexity: 3.0/10
- Empirical Rigor: 8.0/10
- Quadrant: Street Traders — empirical and applied, lower math complexity
Why this score: The paper focuses on a practical, end-to-end framework with extensive implementation details, code repositories, and real-world applications like robo-advising and algorithmic trading, indicating high empirical rigor. However, the mathematical content is relatively light, primarily citing existing LLM architectures like Transformers and LoRA without deep derivations or novel complex mathematics.
Research Flowchart
flowchart TD
A["Research Goal: Democratize FinLLMs with an Open-Source Alternative"] --> B["Data-Centric Approach"]
B --> C["Automatic Data Curation Pipeline"]
C --> D{"Inputs: Internet-Scale<br>Financial Data"}
D --> E["Low-Rank Adaptation<br>LoRA Technique"]
E --> F["Computational Process:<br>Lightweight Finetuning"]
F --> G["Key Outcomes:<br>FinGPT Model"]
G --> H{"Applications:<br>Robo-Advising, Trading, Low-Code"}
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes FinGPT, an open-source large language model for finance built with a data‑centric approach and lightweight low‑rank adaptation.
Source sentence
we present an open-source large language model, FinGPT, for the finance sector.
Assumes 4
- Sentiment labels derived from short‑term price movements accurately reflect market sentiment. data
Source sentence
Market-driven labels (self-labeled data) strongly improve real‑world applicability;
- Low‑rank adaptation (LoRA) can replace full fine‑tuning with comparable performance. model
Source sentence
LoRA reduces adaptation cost by ∼1000× compared to full fine‑tuning;
- Reinforcement Learning on Stock Prices (RLSP) aligns model outputs with actual market reactions. model
Source sentence
RLSP incorporates financial market feedback, distinguishing FinGPT from traditional supervised models.
- The financial news dataset collected via the FinGPT pipeline is clean and representative of market information. data
Source sentence
We utilize a large‑scale financial news sentiment dataset curated through the FinGPT real‑time data pipeline.
Data
- large‑scale financial news sentiment dataset text-news public
Source sentence
The dataset contains: • Over 620,000 cleaned financial news headlines; • Sources including CNBC, Reuters, Yahoo Finance, MarketWatch, etc., collected through the FinNLP pipeline; • Time span from 2016–2024;
Yahoo Finance
Costs
Not applicable
Finds 4
- FinGPT (SFT+RLSP) achieves 82.1% accuracy and 80.9% macro‑F1, outperforming all baselines. performance
Source sentence
FinGPT (SFT+RLSP) 82.1 80.9 83.4 81.5 77.8
- FinGPT (LoRA‑SFT) reaches 78.8% accuracy and 77.3% macro‑F1, beating FinBERT and other models. performance
Source sentence
FinGPT (LoRA‑SFT) 78.8 77.3 79.6 76.8 75.4
- LoRA provides most of the performance gain, while RLSP adds further improvement. performance
Source sentence
Ablation Study LoRA performs most of the heavy lifting, while RLSP further improves market alignment.
- Market‑driven self‑labeled sentiment data improves real‑world applicability of the model. other
Source sentence
Market-driven labels (self‑labeled data) strongly improve real‑world applicability;
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 38 related papers). Coverage is partial, so this is neither replication nor refutation.
78 papers in the archive : Ploutos: Towards interpretable stock movement prediction … ; HedgeAgents: A Balanced-aware Multi-agent Financial Trading … ; RealFin: How Well Do LLMs Reason About Finance When Users … ; FinBloom: Knowledge Grounding Large Language Model with … ; Learning to Generate Explainable Stock Predictions using …
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.