ai4finance-foundation/fingpt

FinGPT is an open-source project from the AI4Finance Foundation that provides financial large language models, AI systems specifically tuned to understand and reason about financial text, news, and market data. The project aims to democratize access to financial AI, which has historically been dominated by expensive, proprietary systems like BloombergGPT (which reportedly cost around $3 million to train). The core idea is to take existing powerful general-purpose open-source language models (like LLaMA 2) and fine-tune them on financial data using a technique called LoRA (Low-Rank Adaptation), which is computationally lightweight. This means a new specialized financial model can be created for under $300 in compute costs rather than millions, and can be updated frequently as new market data becomes available. Several specific applications are included. FinGPT-Sentiment performs financial sentiment analysis, reading news headlines or social media posts and determining whether the sentiment toward a stock or company is positive, negative, or neutral. The project claims its sentiment models outperform GPT-4 on financial benchmarks when run on a single consumer GPU (RTX 3090). FinGPT-Forecaster takes a stock ticker symbol and date, retrieves recent market news, and produces an analysis with a prediction about the following week's price movement, demonstrated via an interactive Hugging Face demo. The project uses reinforcement learning from human feedback (RLHF) to align model outputs with individual user preferences, enabling personalized financial advice applications. Trained models are released publicly on Hugging Face.