AI in Finance
Generative AI, a technology capable of creating data samples resembling the input data, has unleashed a wave of innovation across industries. The remarkable success of ChatGPT has inspired enterprises to venture into the creation of their large language models, especially in the data-driven finance sector.
Every financial institution aspires to develop its finely-tuned LLMs, drawing from open-source models such as LLAMA 2 or Falcon. This pursuit is particularly fervent among legacy banks, which possess decades’ worth of financial data at their disposal.
Previously, limitations in computing resources and the prevalence of less complex, low-parameter models hindered the incorporation of vast datasets into a single model. However, the advent of open-source models boasting billions of parameters has changed the game. Data serves as the lifeblood of these models, with more data equating to better results.
Both data and LLMs hold the potential to save banks and financial services millions by enhancing automation, efficiency, and accuracy.
Recent estimates from McKinseysuggest that Generative AI could lead to annual savings of up to $340 billion in the banking sector alone.
BloombergGPT & the Economics of Generative AI In Finance
In March 2023, Bloomberg introduced BloombergGPT, a language model purpose-built with a staggering 50 billion parameters, finely tuned for financial data analysis.
However, the development of models like BloombergGPT comes at a cost. Training models such as Llama 2, with its colossal 70 billion parameters, demanded a whopping 1,700,000 GPU hours. On commercial cloud services, utilizing the Nvidia A100 GPU, as employed by Llama 2, can translate to a cost of $1 to $2 per GPU hour.
Calculating further, a 10 billion parameter model could incur around $150,000 in expenses, while a 100 billion parameter model could surpass the million-dollar mark.
Alternatively, organizations have the option to purchase GPUs outright. Yet, acquiring approximately 1000 A100 GPUs to construct a cluster might set one back by more than $10 million.
It’s worth noting that Bloomberg’s significant investment becomes noteworthy when compared to the rapid advancements in AI. Astonishingly, a model costing a mere $100 managed to surpass BloombergGPT’s performance within just six months.
While BloombergGPT leveraged proprietary data, a vast majority (99.30%) of its dataset was publicly accessible, giving rise to the emergence of FinGPT.
FinGPT: Leading the Way in Finance
FinGPT, a state-of-the-art financial fine-tuned large language model (FinLLM), is developed by the AI4Finance-Foundation. It currently boasts three versions, with the FinGPT v3 series standing out for their utilization of the LoRA method.
These models are trained on news and tweets for sentiment analysis, excelling in various financial sentiment tests. FinGPT v3.1 is based on the chatglm2-6B model, while FinGPT v3.2 draws from the Llama2-7bmodel.
- Data Sourcing and Engineering: FinGPT sources data from reputable outlets like Yahoo and Reuters, aggregating a wealth of financial news spanning from US stocks to CN stocks. This raw data undergoes extensive cleaning, tokenization, and prompt engineering to ensure relevance and accuracy.
- Large Language Models (LLMs): The curated data enables the fine-tuning of LLMs to create lightweight models tailored to specific needs. Existing models or APIs can also be adapted to support various applications.
- Fine-Tuning Strategies:
- Tensor Layers (LoRA): Recognizing the challenge of obtaining high-quality labeled data, FinGPT adopts an innovative approach. Instead of relying solely on traditional labeling, it leverages market-driven stock price fluctuations as labels, translating news sentiment into tangible categories like positive, negative, or neutral. This approach significantly enhances the model’s predictive capabilities, particularly in discerning sentiment.
- Reinforcement Learning from Human Feedback (RLHF): FinGPT utilizes RLHF to discern individual preferences, such as a user’s risk appetite or investment patterns. This technique, also found in ChatGPT and GPT4, enhances user experiences, tailoring them to individual needs.
Applications and Innovations
FinGPT finds applications across various financial domains:
- Robo Advisor: Functioning akin to seasoned financial advisors, FinGPT analyzes news sentiments and predicts market trends with precision.
- Quantitative Trading: By identifying sentiments from diverse sources, including news outlets and Twitter, FinGPT aids in formulating effective trading strategies, even when solely driven by Twitter sentiments.
FinGPT’s Trajectory and Future
In July 2023, FinGPT reached a significant milestone with the release of a research paper titled, “Instruct-FinGPT: Financial Sentiment Analysis by Instruction Tuning of General-Purpose Large Language Models.” This paper explores instruction tuning, allowing FinGPT to perform intricate financial sentiment analyses.
Beyond sentiment analysis, FinGPT boasts 19 diverse applications, promising novel ways to leverage LLMs in finance, from prompt engineering to deciphering complex financial contexts.
How Global Banks Embrace Generative AI
While the year 2023 witnessed some major financial players imposing constraints on the usage of OpenAI’s ChatGPT, others embraced the technology.
Morgan Stanley integrated OpenAI-powered chatbots as valuable tools for their financial advisors, tapping into internal research and data to enhance knowledge resources. Hedge fund Citadel explored securing an enterprise-wide ChatGPT license for software development and data analysis.
JPMorgan Chase focused on employing large language models for fraud detection, with plans to add up to $1.5 billion in value with AI by year-end. Goldman Sachs ventured into integrating generative AI to strengthen its software engineering domain.
In this rapidly evolving landscape, embracing generative AI is becoming a pivotal strategy for banks and financial institutions.
Use Cases of Generative AI in Banking and Finance
Generative AI is fundamentally reshaping banking and finance with its diverse applications:
- Fraud Prevention: Generative AI excels at detecting fraud by analyzing vast data pools, discerning intricate patterns, and reducing false positives.
- Credit Risk Assessment: It offers comprehensive risk profiles by considering various parameters, modernizing credit evaluations.
- Customer Interaction: Generative AI-powered NLP models enhance customer service by promptly addressing inquiries and automating routine tasks.
- Personalized Finance: Tailored financial planning based on individual needs and preferences is made possible through generative AI.
- Algorithmic Trading: It provides valuable insights for optimizing trading strategies and anticipating market shifts.
- Compliance Frameworks: Generative AI simplifies Anti-Money Laundering (AML) compliance by pinpointing suspicious activities.
- Cybersecurity: It enhances cybersecurity by enabling faster threat detection.
However, generative AI in finance presents challenges, including bias amplification, output reliability, data privacy, and the quality of input data.
In conclusion, while generative AI promises transformative potential in finance, it must be approached with caution, addressing ethical and privacy concerns. Institutions that effectively harness its capabilities while respecting its limitations will shape the future of the financial sector.
Explore our Breaking News section. Stay connected and engage in discussions by following us on Facebookand Instagram. To receive the top headlines and important stories directly in your email inbox, consider subscribing to our daily newsletter.
You can also explore the latest and best AI tools in the market by clicking here: Best AI Tools.
If you have any questions or comments, don’t hesitate to get in touch with us. We value your feedback and it’s important to us.