FinGPT is a large language model (LLM) that is fine-tuned with financial data for the finance sector. It is more accessible, flexible, and cost-effective than models like BloombergGPT. The model is open-source and data-centric, meaning that it provides accessible and transparent resources for developing financial LLMs (FinLLMs).
FinGPT is a joint project by researchers from the Columbia University and NYU (Shanghai) who work with AI4Finance Foundation. Its codes are shared under the MIT education license for academic purposes.
You can use it for various financial tasks including robo-advising, quantitative trading, portfolio optimization, financial sentiment analysis, risk management, fraud detection, and low-code development.
How to update financial LLMs faster and cheaper
Successful LLMs such as ChatGPT require high-quality labeled data. Moreover, financial LLMs need frequent updates (retraining) with the latest financial data, which is expensive and time-consuming.
BloombergGPT used 1.3 million GPU-hours, costing $3 million per training on AWS cloud.
The team suggested a faster and cheaper way to update finance-specific LLMs, called lightweight fine-tuning. Instead of retraining the whole model, they used APIs to fine-tune it with the new data. Their method greatly reduced the time and money needed for updates. For example, using this way for FinGPT would lower the update cost to less than $300.
In summary, FinGPT improves the pre-trained open-source LLMs with updated financial data.
FinGPT framework
In the next diagram we can see the main components of the model’s framework.
There are four main parts:
1. Data Source: This component provides a collection of diverse and dynamic financial data from the online sources. They cover various aspects of the finance sector, such as market data, financial reports, news articles, social media posts, and research papers. For data collection and real-time data updates they used a set of APIs.
2. Data Engineering: This component prepares the data for LLMs. A real-time data engineering pipeline includes data cleaning, tokenization, text normalization, filtering, and prompt engineering.
3. LLMs: This component uses a lightweight adaptation technique to fine-tune pre-existing open-source LLMs with the financial data. The fine-tuning methods include Low-Rank Adaptation (LoRA) and Reinforcement Learning on Stock Prices (RLSP).
- LoRA reduces the number of trainable parameters from 6.17B to a mere 3.67M, by using low-rank decomposition.
- RLSP replaces Reinforcement Learning on Human Feedback (RLHF), as used by ChatGPT. The system learns from the market’s reaction to the stock price changes.
4. Applications: This component provides several potential applications of FinGPT in the finance sector, such as robo-advising, algorithmic trading, and low-code development.
The accuracy of FinGPT depends on the quality and quantity of the financial data it learns from. The researchers have not reported any quantitative evaluation of FinGPT’s accuracy on standard benchmarks or datasets, but they have showcased some examples of its potential use.
Example of using FinGPT
FinGPT may be widely applied in financial services. For example, if you are an investor who wants to buy some stocks, you can ask FinGPT to give you some suggestions based on your preferences and goals.
FinGPT will analyze the market data, news articles, and financial reports to generate a list of potential stocks that match your criteria. It will also explain why it chose those stocks and what are the risks and opportunities involved. You can then use FinGPT’s advice to make your own informed decision.
Conclusion
FinGPT is an innovative solution for financial language modeling that uses pre-existing LLMs and fine-tunes them for specific tasks. It is an open-source financial-specific variant of the GPT language model, developed as an alternative to proprietary models like BloombergGPT.
The authors recommend using FinGPT with proper caution and always seeking professional advice before trading or investing.
Learn more:
- Research paper: “FinGPT: Open-Source Financial Large Language Models” (on arXiv)
- Repositories on GitHub: FinGPT and FinNLP
- Project page
