AI+Wealth Management: Top international banks plan to use GPT in this way

Source: Bingjian Technology Research Institute

A recent list on AI and banking has been cited more and more. The list ranks the 23 largest banks in Europe and the United States, and the total assets of at least US$1 trillion are eligible for selection.

This list is called "Bank AI Index (Evident AI Index)", released by the consulting company Evident Insights, it is publicly available, and it is the first time to rank banks' AI maturity (AI maturity) list.

Top 10 Bank AI Index List:

Source: Evident Insights, compiled by Bingjian Technology Research Institute

In order to make this list, Evident Insights collected millions of data points, based on bank financial reports and public data from a series of third-party data sources, with the participation of more than 50 leading artificial intelligence and banking experts, to establish this list.

Each bank is assessed on 142 individual indicators across four dimensions: talent, innovation, leadership and transparency. Talent accounts for the highest weight, reaching 40%.

According to Evident Insights, the quantity and quality of AI talents will largely affect the future competitiveness of these top banks. JPMorgan Chase, which ranks first, has the most artificial intelligence employees in the banking industry, accounting for more than 10% of the total number of employees and is still accelerating recruitment. Between February and April 2023, JPMorgan Chase posted at least 20% of job advertisements for artificial intelligence and data core positions of all banks on the list.

After searching and analyzing the information on recruitment websites such as LinkedIn, Evident Insights also found an interesting phenomenon: at a time when generative artificial intelligence is so hot, these 23 banks only have less than 2% of AI-related job descriptions, Skills for generative AI such as large language models (LLMs) or ChatGPT are explicitly mentioned.

After analyzing the Evident-related reports and the listed banks, Bingjian Technology Research Institute also found that although GPT has become a well-known science, these large international banks do not think that it can cure all diseases. Due to the early investment in artificial intelligence technology and deep deployment, many banks** have already established deep learning wealth management systems that are quite mature, and they have not rushed to catch up with GPT hotspots. **

On the contrary, the "transparency" indicators mentioned in the list are valued by many big banks. come higher.

Morgan Stanley: Use GPT-4 for knowledge management

Morgan Stanley, which barely squeezed into the top 10 in the bank AI index list, is the most high-profile in terms of GPT applications, and its "innovation" sub-item ranks fourth. Even so, Morgan Stanley's application of GPT is still in the experimental stage and has not entered the production environment, and the experimental field is not extensive.

When OpenAI officially launched GPT-4 in March this year, it launched Morgan Stanley's wealth management application as a typical case.

Specifically, Morgan Stanley maintains a library of hundreds of thousands of pages of content covering investment strategies, market research and commentary, and analyst views—so much information spread across many internal sites, much of it in PDF form, Requiring a Financial Advisor (FA) to browse through a vast amount of information to find an answer to a particular question can be quite inefficient.

Starting last year, the company and OpenAI began working together to explore how to use GPT's embedding and retrieval capabilities to maximize its "intellectual capital" -- more than 100,000 documents.

GPT-4 will provide support for the company's internal chatbot** (note that it is not external)**, which can conduct a comprehensive search and integration of wealth management content, and then provide financial advisors with the answers they want.

Morgan Stanley, which has more than 15,000 financial advisors, might ask its internal chatbot these questions:

*Investment Advice (What is our take on Alphabet stock and is its future performance bullish or bearish?)

*Business as usual (Who are IBM's five main competitors?)

* Process question (How do I put an IRA into an irrevocable trust?).

Morgan Stanley "fine-tuned" GPT-4 on a similar problem using 100,000 documents as a training corpus.

According to Forbes, 300 of Morgan Stanley's FAs are helping the models with "reinforcement learning" -- when they get an answer from a chatbot, they can give an upvote or downvote, or provide more detail on demand. feedback.

One of the widely criticized problems of ChatGPT is that it often produces “hallucinations” content without factual basis, which is fatal for wealth management services. In response, Morgan Stanley is limiting the types of prompts/questions FAs can enter into the system, limiting topics to business-relevant questions, which ensures that the output comes from their existing knowledge documents.

If the FA finds that the content is wrong during use, you can also refer to the reason code - citing the underlying article linked to the source of the content - which is more complete and credible than most large language models.

In the end, there are compliance auditors who check the content. In the company's normal knowledge management process, there are compliance personnel who review the content of investment research, not to mention the content that FAs want to provide to the outside world.

In fact, Morgan Stanley's wealth management department has spent many years researching the "Next Best Action" system (NBA), which is a realistic tool for arming 15,000 FAs with machine learning.

The NBA system discovers personalized investment ideas through machine learning and distributes them to specific customers through its CRM system. The NBA system has three distinct goal functions:

One is to provide customers with investment advice and help decision-making, not only provide passive investment, but also provide individual stock and bond investment options according to the wishes of customers;

The second is alerts for prompt operations, such as alerts for low cash balances and alerts for significant changes in the value of customer investment portfolios, etc.;

The third is life event planning. For example, if it is confirmed that the customer’s child is ill, the system can recommend the local hospital that is best at treating the disease and financial planning for treatment, so as to establish a value-added relationship with the customer.

Jeff McMillan, head of data and innovation at Morgan Stanley, who leads the GPT-4-related business, told Forbes that the NBA system's "push" approach may be as good as the "pull" approach based on GPT prompt answers to cooperate.

According to the latest report in July by AdvisorHub, a vertical website for the wealth advisory industry, Morgan Stanley expects to roll out generative AI tools to its more than 15,000 financial advisors in the third quarter of this year. Since March this year, only 900 FAs have been on trial. .

In the bank AI index list, Morgan Stanley's talent sub-item ranks only 11. Morgan Stanley has accelerated AI talent recruitment since the second half of the year. Its latest recruitment position is to recruit new wealth management executives for artificial intelligence and machine learning platforms. According to LinkedIn, the basic annual salary of this position is between 180,000 and 260,000 US dollars between.

AI champion and runner-up bank: Enhance the existing machine learning system

JPMorgan Chase, which ranks at the top of the list, has some declarative plans for GPT, but did not disclose too many application details; while Royal Bank of Canada (RBC), which appeared as a dark horse in the runner-up position, never mentioned GPT.

According to CNBC reports, JPMorgan Chase is developing a software service similar to ChatGPT. Documents submitted by JPMorgan Chase & Co show that the bank applied for trademark registration for a product called "IndexGPT" in May. IndexGPT will utilize "cloud computing software using artificial intelligence" to "analyze and select securities that suit clients' needs."

The trademark registration document pointed out that IndexGPT uses artificial intelligence technology represented by ChatGPT. Lori Beer, global technology director of JPMorgan Chase, said that the bank has hired 1,500 data scientists and machine learning engineers and is testing "multiple use cases" of GPT technology. ".

“This is going to be the holy grail of how people manage their assets,” Mary Callahan Erdoes, chief executive of the bank’s asset and wealth management division, said of AI at JPMorgan’s Investor Day conference on May 22.

“We’ve loaded 30 years of proprietary data on every company we’ve looked at,” Erdoes said, describing his department’s recent tool development, “and then we’ve combined it with the millions of data points we get every day. Matching, we've seen such a huge lift."

She further revealed that **JPMorgan Chase has its own internal asset management business, and the GPT-like model runs on its Spectrum portfolio management system. **

The runner-up of the bank AI index list is RBC from Canada. The bank has been using deep learning and reinforcement learning technology for wealth management for many years, especially in the top three rankings in the "innovation" and "transparency" sub-index rankings.

RBC has set up an artificial intelligence research center called Borealis AI, which not only serves the parent bank but also engages in cutting-edge research on artificial intelligence. In an interview with KPMG, Kathryn Hume, head of Borealis AI, detailed how her team applied reinforcement learning to banking customer service:

Borealis AI and RBC Capital Markets team launched a reinforcement learning-based trade execution system. “We wanted to understand how machine learning could be used to help clients with large or bulk orders better sequence trades for maximum returns. It turns out that the models we created were very dynamic, responding in real time with more flexibility than traditional trading algorithms Variations in volatility."

Borealis AI has also successfully helped retail and commercial banks transform yesterday's business processes into tomorrow's future products. For example, ** built a cash flow forecasting tool to help financial advisors proactively engage with clients, understand upcoming financial needs, and provide more targeted advice. ** Also helping retail customers manage their finances by creating apps and benefiting from the latest personalization machine learning technology.

In April of this year, RBC won the Best Artificial Intelligence Customer Experience Award from Digital Banker magazine for the NOMI Forecast system jointly developed by the bank and Borealis AI.

The NOMI Forecast system utilizes deep learning to provide timely and accurate forecasts of clients' cash flow. Powered by the bank's unique dataset, models are trained to personalize experiences for RBC customers, including bill payments, electronic transfers, investments and payroll payments.

Vertical large model: suitable is the best

Whether it is the NBA system of Morgan Stanley, the Spectrum system of Morgan Stanley, or the NOMI Forecast system independently developed by RGB, they are all combinations of various models trained by the bank's own data accumulated. After grafting GPT for fine-tuning training, enhancing the general interaction ability is a similar choice of these top international banks.

Regardless of foreign countries or domestically, with the increasing number of open source large models and the decline in model training costs, the obsession with general large language models has gradually faded. From the just-concluded Shanghai World Artificial Intelligence Conference, it can be found that the new narrative is: industry model, vertical model and "large model empowers thousands of industries".

The most typical example is BloombergGPT launched by Bloomberg. Bloomberg made the model smaller, with about 50 billion parameters, which is much smaller than the 175 billion parameters of GPT-3. Although it weakens the versatility, its performance in the financial field Significantly better than general-purpose large models.

The strong supervision and professionalism of the financial industry determines that on the basis of Know-How, professional data training accumulated by financial institutions can be used to create a vertical model suitable for the needs of the industry. **For example, the Origin One large model launched by Bingjian Technology, relying on years of experience in algorithm models serving banks and insurance customers, is making efforts in intelligent customer service, financial document processing, and analysis of foreign investment products. **

The professional advantages of the vertical model will become more and more obvious in the wealth management industry. The emergence of GPT will change the concept that robo-advisors are "exclusive services" for high-net-worth families in the past, and enable the rapid development of the long-tail market. With the help of GPT-4, how many times will the number and efficiency of Morgan Stanley's 15,000 financial advisors serve customers be increased?

In addition to these top banks, the vertical large model gives startups more opportunities to serve more customers at the application layer. Technology companies attract more sinking customers through low barriers to entry and independence, while traditional large banks use their own advantages to target existing customers and promote various product portfolios. Institutions of different sizes determine different strategies in the field of robo-advisory .

From the perspective of the US robo-advisory market, it mainly includes three types of participants:

First, start-up companies represented by Wealthfront and Betterment use their own technological advantages and low threshold requirements to tap the value of long-tail customers;

The second is that large-scale financial institutions represented by Vanguard and Charles Schwab leverage their own capital advantages, existing customer advantages, brand advantages and competitive barriers to launch intelligent investment advisory products;

The third is to acquire third-party companies to quickly deploy the smart investment advisory market, such as BlackRock’s acquisition of Future Advisor.

According to the calculation of the credit rating company CRISIL GR&A, the application of large-scale models in the field of intelligent investment research is expected to save 22.5% of the cost, which will make wealth management benefit more people.

** Greater inclusiveness also means more risk. **It is worth mentioning that RBC’s ability to surpass many major European and American banks and rank second in the bank AI index list is also due to its responsible AI (Responsible AI) action. Kathryn Hume believes that people are increasingly aware of the moral hazard that AI may exacerbate. Around the world, the debate surrounding the ethical and responsible use of artificial intelligence is intensifying.

Explainable artificial intelligence (XAI) will be an emerging technology, perhaps its counterpart to GPT. While ChatGPT's algorithmic insights are completely "black boxes," XAI allows users and regulators to scrutinize the fundamentals of how AI works, and urges developers to hone algorithms so they work as intended. **XAI enables wealth managers and investment advisors to monitor and justify AI-derived financial advice and align it with regulatory requirements and clients’ best interests. **

Reference materials:

• Evident Insights: Evident AI Index for Banks Key Findings Report
• FORBES: How Morgan Stanley Is Training GPT To Help Financial Advisors
• AdvisorHub: Morgan Stanley to Roll Out AI Software to All Brokers in Third Quarter
• Ping An Securities: * "A View of Digital Transformation of Commercial Banks from AIGC"*