• Financial Generative Pre-trained Transformers (FinGPT)

Origins and research contributions

FinGPT began in 2023 as an open, data-centric approach to financial large language models. The original paper by Hongyang Yang, Xiao-Yang Liu, and Christina Dan Wang proposed accessible financial data pipelines and lightweight adaptation of general-purpose language models. At the time, proprietary financial models such as BloombergGPT highlighted the value of domain data, while access restrictions and the expense of large-scale training motivated an open research alternative. The project’s objective was to make financial language-model development more accessible and transparent.

The original research emphasized automatic data collection and curation, the time-sensitive and noisy nature of financial text, and low-rank adaptation (LoRA). The subsequent data-centric study by Xiao-Yang Liu, Guoxuan Wang, Hongyang Yang, and Daochen Zha described data collection from 34 Internet sources and explored market-derived supervision, termed Reinforcement Learning with Stock Prices (RLSP), alongside LoRA and QLoRA. Early papers outlined personalized assistants and alignment with individual preferences as research goals, alongside experiments on lightweight adaptation and market-derived supervision.

Instruct-FinGPT investigated instruction tuning for financial sentiment analysis, including challenges involving numerical values and financial context. Later work expanded the research in two complementary directions. FinGPT-HPC investigated low-rank structures, quantization, and GPU optimization for efficient model pretraining and fine-tuning. Customized FinGPT Search Agents studied retrieval-augmented workflows for individuals’ documents and selected sources, and institutional workflows using dynamic vector databases and proprietary data.

Together, these studies document a progression from open financial data and lightweight model adaptation toward efficient computing and customized financial agents. Their results describe particular datasets, models, and experimental settings, rather than universal performance guarantees. Historical cost comparisons also require context: the data-centric paper estimated one 64-A100-GPU-hour fine-tuning experiment at about $262 using July 2023 AWS prices; this is not a general cost per update and differs in scope from full model pretraining.

Foundational paper

Hongyang Yang, Xiao-Yang Liu, and Christina Dan Wang (2023). FinGPT: Open-Source Financial Large Language Models.

Historical development note

The earlier FinAI Lab project page recorded that the lab had been inactive on the AI4Finance FinGPT GitHub repository since November 2023. This records the lab’s involvement at that time; the community repository and subsequent research have separate development histories.

Community GitHub repository · Models and datasets · Model Openness Framework

Financial Generative Pre-trained Transformers (FinGPT)

High-Performance Computing (Optimized GPU Primitives) for FinGPT

Columbia Affiliations
School of Engineering