The transformational integration of large language models (LLMS) into the financial domain presents unique complexities and enormous opportunities. Efficient solutions are required to address the challenges of high time sensitivity, dynamic financial landscape, and low signal-to-noise ratio in financial data. FinGPT responds innovatively by leveraging pre-existing LLMs and fine-tuning them for specific financial applications. Compared with models such as BloombergGPT, this method greatly reduces the adaptation cost and computational requirements, and provides a more convenient, flexible and economical solution for financial language modeling. As such, it enables consistent updates to ensure model accuracy and relevance, a critical aspect in the dynamic and time-sensitive world of finance.
Why fine-tuning instead of retraining
Although BloombergGPT is very capable in finance, it is very computationally intensive. It took about 1.3 million GPU hours to train, which equates to about $3 million per training at AWScloud's $2.3 rate. In contrast to the high computational cost of models such as BloombergGPT, FinGPT provides a more accessible solution by focusing on lightweight adaptation of top open-source LLMs. Adaptation costs have dropped significantly, estimated at less than $300 per session. This approach ensures timely updates and adaptability, especially in the dynamic financial domain. The open source FinGPT not only improves transparency, but also allows user customization, catering to the rising trend of personalized financial consulting services. Ultimately, FinGPT's cost-effective, flexible framework has the potential to make financial language modeling irrelevant and facilitate user-centric financial services.
Fine-tuning via Low-Rank Adaptation (LoRA)
In FinGPT, we leverage a novel financial dataset to fine-tune a pretrained LLM. It is well known that high-quality labeled data is a key determinant for many successful LLMs, including ChatGPT. However, obtaining such top-level labeled data is often costly in terms of time and resources, and often requires the expertise of financial professionals.
If our goal is to use LLMs to analyze financially relevant textual data and assist in quantifying transactions, it seems sensible to leverage the market's inherent labeling capabilities. Therefore, we use the relative stock price change percentage for each news item as the output label . We classify these labels into three categories based on the sentiment of news items : positive, negative, and neutral .
In the corresponding hint engineering step, we also hint the model to choose one from positive, negative and neutral outputs. This strategy ensures optimal utilization of pre-trained information. By using Low-Rank Adaptation (LoRA) with LLMs [Hu et al., 2021; Dettmers et al., 2023], we managed to reduce the number of trainable parameters from 6.17 billion to just 3.67 million.
Fine-tuning stock prices with reinforcement learning (RLSP)
Likewise, we can replace stock price-based reinforcement learning (RLSP) with the human feedback-based reinforcement learning used by ChatGPT. The reason for this substitution is that stock prices provide a quantifiable, objective measure of market sentiment about news and events. This makes it a powerful real-time feedback mechanism for training models.
Reinforcement learning (RL) allows models to learn through interaction and feedback with the environment. In the case of RLSP, the environment is the stock market and the feedback comes in the form of changes in stock prices. This approach enables FinGPT to refine its understanding and interpretation of financial texts, improving its ability to predict market responses to various financial events.
By linking news sentiment to the subsequent performance of related stocks, RLSP provides an efficient way to fine-tune FinGPT. Essentially, RLSP allows models to infer market reactions to different news events and adjust their understanding and forecasts accordingly.
Therefore, integrating RLSP into the fine-tuning process of FinGPT provides a powerful tool to improve the model's understanding of financial markets and predictive accuracy. By using actual stock price movements as feedback, we directly leverage market intelligence to make our models more effective.
end-to-end architecture
The whole framework is divided into five levels
- _Data source layer：确保全面的市场数据覆盖和时间敏感性问题 _The starting point of the FinGPT pipeline is the Data Source Layer, which orchestrates the acquisition of extensive financial data from a wide array of online sources. This layer ensures comprehensive market coverage by integrating data from news websites, social media platforms, financial statements, market trends, and more. The goal is to capture every nuance of the market, thereby addressing the inherent temporal sensitivity of financial data.
- _Data engineering layer: preparing for real-time NLP data processing, dealing with high temporal sensitivity and noise issues in financial data_This layer focuses on the realtime processing of NLP data to tackle the challenges of _high temporal sensitivity _and _low signal-to-noise ratio _inherent in financial data. It incorporate s state-of-the-art NLP techniques to filter noise and highlight the most salient pieces of information.
- _LLMs layer: focus on a series of fine-tuning methodologies, which alleviate the highly dynamic nature of financial data and ensure the relevance and accuracy of the model_Lying at the heart, it encompasses various fine-tuning methodologies, with a priority on lightweight adaptation, to keep the model updated and pertinent. By maintaining an updated model, FinGPT can deal with the highly dynamic nature of financial data, ensuring its responses are in sync with the current financial climate
- _Application layer：实际应用和demo示例，_The final component of FinGPT is the Applications Layer, designed to demonstrate the practical applicability of FinGPT. It offers hands-on tutorials and demo applications for financial tasks, including roboadvisory services, quantitative trading, and low-code development. These practical demonstrations not only serve as a guide to potential users but also underscore the transformative potential of LLMs in finance.