What is FinGPT
FinGPT is an open-source large pre-trained language model for the financial domain, driving financial innovation based on natural language processing technology. Through reinforcement learning with human feedback (RLHF) technology, it can learn personal preferences and provide personalized investment advice. FinGPT supports a variety of financial tasks, such as sentiment analysis, relation extraction, headline classification, and named entity recognition, with broad data sources covering financial news websites, social media, financial regulatory agency websites, and more.
FinGPT's main features
- Financial sentiment analysis: Through natural language processing technology, it analyzes the sentiment tendency of financial texts (such as financial news, social media comments, etc.) to determine whether they are positive, negative, or neutral.
- Financial relation extraction: Extracts relationships between financial entities from text, such as cooperative relationships and merger and acquisition relationships between companies.
- Financial headline classification: Classifies financial news headlines to determine the financial topic they belong to (such as the stock market, monetary policy, industry trends, etc.).
- Financial named entity recognition: Identifies financial entities in text, such as company names, stock tickers, and financial product names.
- Market prediction: Combines historical data and real-time information to predict market trends and stock price movements.
- Personalized investment advice: Provides personalized investment advice by learning users' investment preferences and risk tolerance.
- Data-driven model training: Supports low-rank adaptation (LoRA) and reinforcement learning technology to quickly adapt to new data and reduce training costs.
- Multilingual support: Supports financial text processing in multiple languages, covering global financial markets.
FinGPT's technical principles
- End-to-end framework: FinGPT adopts a complete end-to-end framework, containing four layers:
- Data source layer: Obtains real-time financial data from multiple channels to ensure comprehensive market coverage.
- Data engineering layer: Cleans, preprocesses, and extracts features from real-time data to address the high time sensitivity and low signal-to-noise ratio of financial data.
- LLMs layer: Based on pre-trained LLMs, fine-tuned through low-rank adaptation (LoRA) and stock price-based reinforcement learning (RLSP) to adapt to the dynamic changes in the financial domain.
- Application layer: Provides a variety of financial applications, such as sentiment analysis, market prediction, investment advice, etc.
- Pre-trained models and fine-tuning: FinGPT is based on a pre-trained Transformer architecture and learns rich financial knowledge and language patterns by training on large-scale financial text data. Fine-tuning techniques include:
- Low-rank adaptation (LoRA): Significantly reduces training costs by reducing the number of trainable parameters. For example, it reduces trainable parameters from 6.17 billion to 3.67 million while maintaining model performance.
- Stock price-based reinforcement learning (RLSP): Uses stock price changes as feedback signals instead of traditional reinforcement learning based on human feedback. This method enables the model to learn in real time how the market reacts to news and events, improving prediction accuracy.
- Real-time data processing: FinGPT's data engineering layer supports real-time data processing, including:
- Data cleaning: removing irrelevant data, handling missing values, text normalization, etc.
- Feature extraction: using techniques such as TF-IDF, bag-of-words models, or Word2Vec to convert raw data into input that machine learning models can understand.
- Sentiment analysis: performing sentiment classification on the cleaned data to determine whether it is positive, negative, or neutral.
FinGPT's project address
- GitHub repository：https://github.com/AI4Finance-Foundation/FinGPT
- arXiv technical paper：https://arxiv.org/pdf/2306.06031
FinGPT's application scenarios
- Investment analysis: FinGPT can help analyze sentiment and fundamental information for investment targets such as stocks and funds, providing investors with a more comprehensive basis for decision-making.
- Market research: by quickly processing unstructured data such as news and social media, FinGPT can identify market sentiment and help investors and financial institutions grasp market trends.
- Automated quantitative trading: FinGPT's predictive capabilities can be combined with quantitative trading strategies to optimize AI-based trading strategies and improve trading efficiency and returns.
- Intelligent forecasting: based on the FinGPT-Forecaster module, users can obtain predictions of stock price trends, and the functionality supports easy access through HuggingFace Spaces.
- Sentiment analysis: FinGPT can perform accurate sentiment analysis on information in financial news and social media to assist decision-making.