FinGPT: Open-Source Financial LLM

Updated 15 October 2025

FinGPT is an open-source financial language model designed to democratize access to high-quality financial data and adaptable NLP tools for applications like robo-advising and trading.
It employs a data-centric pipeline with automated real-time data curation and advanced low-rank adaptation techniques to enhance model efficiency and scalability.
Its modular architecture supports rapid prototyping and community-driven innovation, addressing challenges in complex reasoning and numerical tasks.

FinGPT is an open-source financial LLM (FinLLM) developed to democratize access to high-quality financial data and provide researchers and practitioners with a suite of adaptable, transparent, and efficient tools for financial natural language processing. Built on a data-centric philosophy, FinGPT incorporates advanced model adaptation techniques, an automated real-time data curation pipeline, and a strong open-source ethos, facilitating applications in robo-advising, algorithmic trading, and low-code development. Its layered architecture enables researchers to rapidly construct financial LLMs without dependence on proprietary data silos, positioning FinGPT as a pivotal resource in open finance.

1. Data-Centric Architecture and Curation Pipeline

FinGPT operationalizes a data-centric approach, wherein the ingestion, preprocessing, and structuring of financial data are foundational. The system automates collection from a wide array of sources, including financial news sites, social media, regulatory filings, and academic datasets (Liu et al., 2023). The real-time pipeline is engineered to solve three primary challenges intrinsic to financial NLP:

High temporal sensitivity (data validity may rapidly expire),
High dynamism (market regimes and language evolve quickly),
Low signal-to-noise ratio (irrelevant, duplicate, or noisy text is prevalent).

The multi-stage process encompasses:

Tailored web crawling and API integration to support heterogeneous data sources,
Standardization and rigorous cleaning (URL removal, whitespace normalization, language filtering, n-gram de-duplication, perplexity scoring),
Layered architecture partitioned into data source aggregation, data engineering, LLM adaptation, and application modules.

This pipeline is continually updated, handling structural variation between sources such as stream-based vs. date-ranged APIs, and implements document-level filtering for real-time downstream analysis. The architecture is formally structured to support rapid prototyping and iterative improvement.

2. Model Adaptation: Low-Rank Techniques and Parameter Efficiency

Central to FinGPT is its use of lightweight Low-Rank Adaptation (LoRA, QLoRA) for efficient fine-tuning of large pre-trained LLMs (Yang et al., 2023, Liu et al., 2023). Rather than retraining the full weight matrix $W \in \mathbb{R}^{d \times k}$ , LoRA decomposes updates as:

$\Delta W = A B$

with $A \in \mathbb{R}^{d \times r}$ , $B \in \mathbb{R}^{r \times k}$ , and $r \ll \min(d,k)$ , such that only a small subset of model parameters (often a few million out of billions) require updating. This drastically reduces compute and memory cost of adaptation—from multi-million dollar training expenses (e.g., BloombergGPT) to sub-\$300 per run, with LoRA and QLoRA supporting efficient customizations and frequent updates for time-sensitive financial data.

FinGPT-HPC extends this paradigm by replacing the standard transformer’s large linear layers with compositions of two narrow layers, yielding parameter counts of $2nr$ rather than $n^2$ and enabling quantization to 8-bit or 4-bit precision (Liu et al., 21 Feb 2024). These advances result in speedups ( $\sim$ 1.3 $\times$ ), significant compression ratios ( $\sim$ 2.64 $\times$ ), and high memory efficiency (up to $6.3\times$ ), with model sizes suitable for mobile inference.

3. Instruction Tuning and Market-Driven Training

FinGPT leverages advanced instruction tuning methods for adapting general-purpose LLMs to financial tasks, converting classification datasets into prompt-response pairs that align model outputs with nuanced contextual understanding (Zhang et al., 2023, Wang et al., 2023). Using instruction-formatted sequences, e.g.:

Human: [instruction] + [input text]
Assistant: [sentiment label],

the approach enables sequence-to-sequence learning and exploits financial numeracy, outperforming traditional supervised sentiment models in key metrics (accuracy up to 0.88, F1 up to 0.841).

Market-driven methods such as Reinforcement Learning with Stock Prices (RLSP) further automate label generation by using post-publication stock returns as ground-truth sentiment:

$\text{Label} = \begin{cases} \text{Positive,} & \text{if } (P_1/P_0 - 1) > 0.02 \ \text{Negative,} & \text{if } (P_1/P_0 - 1) < -0.02 \ \text{Neutral,} & \text{otherwise} \end{cases}$

where $P_1$ is the future price and $P_0$ is the reference price (Liu et al., 2023).

Instruction tuning paradigms are benchmarked using end-to-end schemes with standardized prompts and phased evaluation (task-specific, multi-task, zero-shot), supporting systematic LLM assessment (Wang et al., 2023).

4. Applications

FinGPT’s open-ended modular approach enables deployment across diverse financial use cases (Yang et al., 2023, Liu et al., 2023, Talazadeh et al., 22 Sep 2024):

Robo-advising: Automated, customizable financial advice that incorporates user-specific risk profiles and real-time market signals.
Sentiment analysis and algorithmic trading: Extraction of granular sentiment signals from financial news and social media for downstream trading strategies; empirical improvements in accuracy, F1, and simulated cumulative return rates.
Low-code financial development: Code-generating capabilities facilitating rapid creation of factor libraries, financial analytics engines, and modeling prototypes for non-programmers.
Search agents and retrieval-augmented generation (RAG): Custom FinGPT agents for individuals and institutions leverage web search and local file integration for personalized, privacy-preserving insights, with accuracy substantially exceeding base LLMs (Tian et al., 20 Oct 2024).

5. Benchmarking, Evaluation, and Comparative Performance

FinGPT has been extensively benchmarked, including on the Golden Touchstone bilingual evaluation suite spanning eight core financial NLP tasks (Wu et al., 9 Nov 2024, Djagba et al., 6 Jul 2025). Performance details include:

High scores in sentiment analysis and headline classification (F1 up to 87.62% and 95.50%, respectively, comparable to GPT-4),
Moderate performance in stock movement prediction (accuracy/F1 45–53%),
Underperformance in complex reasoning and generation tasks (e.g., financial QA Exact Match 28.47% vs. GPT-4's 76%, and summarization ROUGE-1 0% vs. GPT-4's 30%). Notably, relation extraction and summarization scores are frequently lower, indicating architectural and data limitations in current domain adaptation.

FinGPT’s strengths are most pronounced in structured classification tasks, whereas its weaknesses are rooted in numerical reasoning and summarization—primarily due to decoder-only architectural constraints and limited bidirectional context (Djagba et al., 6 Jul 2025).

6. Community, Open Source, and Collaborative Impact

FinGPT’s evolution is driven by an active AI4Finance open-source community (Yang et al., 2023). Key resources are provided via public code repositories:

These platforms foster innovation, code sharing, and reproducibility, enabling transparent customization and lowering barriers for academic and industry adoption. Projects such as FinAgents operationalize multimodal models for domain-specific tasks (search, tutoring, credit scoring, trading), leveraging standardized protocols and agentic architectures (Yanglet et al., 15 May 2025). Collaborative efforts are essential for continuous refinement, transparency, and robust deployment in financial environments.

7. Limitations and Future Directions

FinGPT’s domain-centric adaptation, while effective for several tasks, exhibits notable deficiencies in complex reasoning, numerical arithmetic, relation extraction, and summarization, as evidenced by multiple benchmarks (Wu et al., 9 Nov 2024, Djagba et al., 6 Jul 2025). These gaps are attributed to limitations in model architecture (decoder-only), context window length, instruction quality, and training data alignment.

Proposed future research avenues include:

Architectural improvements: Integration of hybrid encoder-decoder or retrieval-augmented generation frameworks to enhance reasoning and global contextual processing.
Enhanced instruction tuning: Development of richer, domain-specific instruction templates and dynamic prompt engineering.
Multimodal fusion: Expansion into visual, tabular, and structured financial data modalities for improved downstream task performance (Yanglet et al., 15 May 2025).
Numerical reasoning and symbolic integration: Integration of symbolic modules, chain-of-thought prompting, or program-of-thought routines to address shortcomings in arithmetic and logic-heavy applications.
Localization and regulatory adaptation: Extension to localized domains (e.g., Thai financial legal frameworks) using tailored data augmentation and parameter-efficient fine-tuning strategies (Labs et al., 27 Nov 2024).

In sum, FinGPT represents a foundational step in open financial NLP, with ongoing research aimed at bridging gaps in reasoning, multimodal understanding, and complex financial analytics, paving the way for robust, equitable, and interpretable financial AI systems.