FinGPT

FinGPT is a financial LLM framework for researchers and developers building tools that interpret market news and company information. It includes trained models, financial datasets, and tools for adapting models to specific finance tasks.

• Sentiment analysis: Classify financial news and tweets as positive, neutral, or negative, with a retrieval-based framework that adds external context.
• Financial text extraction: Identify people, organizations, and locations, extract relationships, and classify financial headlines with models trained for multiple tasks.
• Stock forecasting: FinGPT-Forecaster uses market news and optional company financials to produce company analysis and predict the following week's stock price direction.
• Training and evaluation: Fine-tune base models with LoRA using financial instruction datasets, and compare results through finance-specific benchmarks.

Base models include Llama-2, Falcon, MPT, Bloom, ChatGLM2, Qwen, and InternLM. Trained models and datasets are available on Hugging Face. Local inference can use NVIDIA GPUs, Apple Silicon through MPS, or a slower CPU fallback. Some sentiment models support training and inference on a single NVIDIA RTX 3090.

The code is open source under the MIT license. Finogrid, the included platform, can run inference locally or send requests to OpenAI or MiniMax cloud services without a local GPU. A hosted Forecaster demo runs on Hugging Face Spaces. Forecaster is trained on US Dow 30 market data; other markets require suitable data and model adaptation. The project is intended for academic use, and its outputs aren't financial advice.