FinGPT is a large-scale pre-trained language model in the financial field. It can understand and generate financial news, analyze public sentiment on social media, interpret financial reports such as annual reports, quarterly earnings reports, etc., make market forecasts, and provide personalized investment advice by learning users’ personal preferences.

The training data for FinGPT comes from a variety of online sources, including financial news websites (such as Reuters, CNBC, Yahoo Finance, etc.), social media platforms, websites of financial regulators (such as the SEC in the United States), official websites of stock exchanges (such as NYSE, NASDAQ, SSE, etc.), financial blogs and forums, and academic datasets.

Core features:

• data-centric
• Use RLHF and Lora’s low-rank techniques to make training on new data fast and cheap
• Supports embedded commercial and open source LLM, the latter supports standalone deployment
• The application layer uses prompt engineer as analyst and trader

The FinGPT architecture looks like this:

Unlike other language models, FinGPT uses open-source large-scale language models and financial data for fine-tuning. Its key technology is reinforcement learning through human feedback (RLHF), enabling the model to learn personal preferences, such as risk aversion levels, investment habits, etc., thereby providing personalized robot advisory services. FinGPT has the ability to quickly fine-tune and adapt to new data , which makes it highly flexible in the financial sector.