FinGPT is an open-source financial large language model project developed and maintained by the AI4Finance Foundation.
In terms of purpose: FinGPT is aimed at natural language processing tasks in the financial domain, with the goal of providing an accessible and quickly updatable open-source alternative in the context of highly dynamic financial data and where internal data and APIs of professional institutions are not open to the public. The project notes that one should not expect Wall Street to open-source large language models or open APIs, so FinGPT emphasizes lightweight adaptation and can quickly incorporate new data at a lower cost.
Main capabilities: FinGPT provides financial sentiment analysis models, a financial multi-task large language model, the financial forecaster FinGPT-Forecaster, and related instruction fine-tuning datasets and models. The project has released multiple versions of sentiment analysis models, including FinGPT v3.3 (based on llama2-13b), FinGPT v3.2 (based on llama2-7b), and FinGPT v3.1 (based on chatglm2-6B), and has published benchmark results on financial sentiment analysis datasets such as FPB, FiQA-SA, TFNS, and NWGI. FinGPT-Forecaster supports inputting a stock ticker, the forecast start date, the number of past weeks for retrieving market news, and whether to include the latest fundamental financial information, and then outputs company analysis and a stock price trend forecast for the next week. The project also includes the FinGPT-RAG retrieval-augmented framework, FinGPT-FinNLP, and the FinGPT-Benchmark instruction fine-tuning benchmark.
How to use: Users can try FinGPT-Forecaster online through the HuggingFace Space without installation. For local or Replit execution, refer to the SETUP guide. The project provides a notebook for fine-tuning the FinGPT v3 model on a single RTX 3090 using the LoRA method, supporting 8bit and int4 (QLoRA) approaches. Datasets and the multi-task financial large language model are published on the FinGPT organization page on HuggingFace. FinGPT supports inference using multiple cloud large language model providers by setting the environment variable FINGPT_LLM_PROVIDER.
Applicable scenarios: Suitable for scenarios such as financial sentiment analysis, financial text processing, financial forecasting research, and fine-tuning and evaluation of open-source financial large language models. FinGPT-Forecaster is currently based on U.S. market data (DowJones 30) for training. If support for other markets (including the Chinese stock market) is needed, corresponding market data must be collected, the model must be fine-tuned on the target market data, and consideration can be given to using the FinGPT v1.x model designed for the Chinese market, along with adapting the data pipeline and model architecture.
Limitations: The materials do not provide FinGPT's commercial charging model, complete hardware requirement details, complete explanations of all installation steps, or guarantees of model performance in real trading. FinGPT-Forecaster's market support scope is limited, and other markets require self-fine-tuning. Some tutorials and detailed explanations are marked as "coming soon."
Project achievements: FinGPT-Forecaster was released in November 2023; the financial multi-task large language model was released in October 2023; the financial sentiment analysis model was released in August 2023; multiple related papers were accepted by NeurIPS 2023 Instruction Workshop, ACM ICAIF-23, IJCAI 2023 FinLLM, and others.
