Introduction
The field of natural language processing (NLP) has witnessed remarkable advances in recent years, thanks to the development of large-scale pre-trained language models (LMs) such as BERT, GPT, and T5. These LMs are trained on massive amounts of general-domain text data, such as Wikipedia, news articles, and web pages, and can capture rich linguistic and semantic knowledge. However, when it comes to domain-specific applications, such as finance, these LMs may not perform well due to the lack of domain-relevant data and vocabulary.
A team of researchers from Columbia University, and New York University (Shanghai) has developed a financial large language model that leverages internet-scale data for pre-training. Their goal is to democratize the internet-scale Data for Financial Large Language Models by introducing an open-sourced and data-centric framework that automates the collection and curation of real-time financial data from diverse sources on the Internet. This new AI model is called 'FinGPT'.
What is FinGPT?
FinGPT stands for Financial Generative Pre-trained Transformer. It is a data-centric framework that introduces a simple yet effective strategy for fine-tuning FinLLM using the inherent feedback from the market, dubbed Reinforcement Learning with Stock Prices (RLSP).
Key Features of FinGPT
FinGPT has several key features that make it unique and powerful for financial NLP applications:
- FinGPT democratizes access to high-quality financial data from various online sources, such as company websites, financial news, blogs, forums, reports, and social media. This enables FinGPT to capture diverse and rich financial knowledge and vocabulary.
- FinGPT uses an automatic data curation pipeline that can update the financial data on a regular basis, such as monthly or weekly. This ensures that FinGPT stays up-to-date with the latest trends and developments in the financial domain.
- FinGPT leverages the strengths of some of the best available open-source large language models (LLMs) and fine-tunes them for financial language modeling. This allows FinGPT to inherit the general linguistic and semantic knowledge from the LLMs and enhance it with domain-specific knowledge from the financial data.
Capabilities/Use Case of FinGPT
FinGPT can enable various applications that require financial domain knowledge and natural language understanding, such as:
- Robo-advisor: FinGPT can provide personalized and automated financial advice to users based on their goals, preferences, and risk profiles. FinGPT can also generate natural language explanations and recommendations for the users to understand the rationale behind the advice.
- Sentiment analysis for algorithmic trading: FinGPT can analyze the sentiment of financial text data, such as news articles, tweets, or reports, and infer the market sentiment and trends. FinGPT can also generate trading signals or strategies based on the sentiment analysis results.
- Low-code development: FinGPT can facilitate the development of financial applications or services by using natural language instructions or queries. FinGPT can also generate code or scripts based on the natural language inputs and execute them accordingly.
How does FinGPT work?
The FinGPT framework is designed to gather a comprehensive amount of accessible financial data from the internet and provide a unified data interface for developers. It incorporates data curation pipelines to ensure that only high-quality data is used in training. Additionally, FinGPT employs reinforcement learning to instruct large language models (LLMs) with market feedback and adapt the model with Low-rank Adaptation (LoRA). This lightweight adaptation approach can significantly reduce costs.
FinGPT consists of four layers, each with its own unique function. The first layer is the data source layer, which offers unified data APIs. The second layer is the data curation layer, which is responsible for cleaning and processing the fine-tuning data. The third layer is the LLM layer, which is capable of accommodating any pre-trained LLM. The fourth and final layer is the application layer, which applies the fine-tuned model to diverse financial applications. This four-layer design makes FinGPT highly extensible and adaptable to a wide range of financial applications.
Performance Evaluation with Other Model
The researchers conducted few experiments to showcase distinct fine-tuning methodologies for FinGPT. The first experiment used Reinforcement Learning with Stock Prices (RLSP) for labeling, leveraging market feedback. The second experiment harnessed an external LLM, such as GPT-4, for labeling, enabling the model to distill knowledge from an already potent LLM. The third experiment involved full-shot fine-tuning, leveraging the entirety of the training data to refine the model. The results of these experiments showed significant enhancements over prevailing LLMs, underscoring the promise of crafting financial large language models (FinLLMs) through fine-tuning.
source - https://arxiv.org/pdf/2307.10485.pdf
FinGPT was compared with other models such as LLaMA and BloombergGPT (as shown in above Table) on various financial tasks. The results showed that FinGPT had a consistent advantage over these models and exhibited substantial improvement in tasks such as sentiment classification and quantitative trading.
source - https://arxiv.org/pdf/2307.10485.pdf
In another experiment as shown in above table, the researchers used instruction fine-tuning on the same datasets as in the previous experiment. The results with full-shot fine-tuning showed that FinGPT outperformed BloombergGPT, which has not released its model or API. Three models were trained in this experiment, including FinGPT, FinGPT-8bit, and FinGPT-4bit. FinGPT is the LoRA model based on ChatGLM which was fine-tuned under the instructions made by the training set. FinGPT-8bit is the 8-bit training version of the FinGPT model and FinGPT-4bit is the 4-bit training version. Thanks to LoRA, the researchers were able to reduce the training cost of LLMs for sentiment analysis from $2.67 million to $262 with high-quality instruction datasets.
For more detailed information on cost estimations and F1 score calculations etc., please refer to the research paper.
How to access and use this model?
FinGPT is an open-source framework that can be accessed through its code repositories on GitHub. The model can also be accessed through the Hugging Face website. FinGPT is designed to be data-centric and open-source for open finance, allowing developers to access its code and use it for their own purposes.
To use FinGPT, developers can visit the GitHub repositories and follow the instructions provided in the README files. The repositories contain detailed information on how to install and use the framework, as well as examples of how to fine-tune the model for specific financial tasks. NOTE That currently codes are shared for academic purposes under the MIT education license.
If you are interested to know more about FinGPT model, all relevant links are provided under 'source' section at the end of this article.
Future Work
The research team in collaboration with The AI4Finance Foundation aims to improve access to high-quality financial data and expand to other markets.
- They plan to investigate more parameter-efficient fine-tuning methods and fine-tune pre-trained FinLLMs for better performance.
- They intend to explore advanced prompting strategies, support longer context windows, and use retrieval-augmented generation techniques.
- They plan to evaluate and mitigate potential bias and fairness issues and enhance the experience of low-code development.
- FinLLMs could potentially be applied to SWAPs for decision support and risk analysis, as well as event detection and outlier detection for predicting market trends and managing risk.
The community is encouraged to use the FinGPT framework to train their own LoRA weights, utilizing a variety of data sources, to expedite advancements in the field of FinLLMs.
Conclusion
FinGPT is a breakthrough in financial NLP that leverages internet-scale data for pre-training. It enables various applications that require financial domain knowledge and natural language understanding. FinGPT is a milestone in the journey of artificial intelligence across industries, especially finance, as it democratizes access to high-quality FinLLMs and stimulates innovation and opportunities in open finance.
Source
research Paper - https://arxiv.org/abs/2307.10485
research document- https://arxiv.org/pdf/2307.10485.pdf
code repo1 - https://github.com/AI4Finance-Foundation/FinGPT
code repo2 - https://github.com/AI4Finance-Foundation/FinNLP
Model - https://huggingface.co/oliverwang15/FinGPT_ChatGLM2_Sentiment_Instruction_LoRA_FT
Website - https://ai4finance-foundation.github.io/FinNLP/
License - https://github.com/AI4Finance-Foundation/FinNLP/blob/main/LICENSE
No comments:
Post a Comment