Robo Advisor Powered by Generative AI

Background

The existence of advisors in the finance and insurance industry has been a tradition and they play an important role in society. It takes a huge amount of effort through education, training and shadowing to become a trusted advisor. They are always in demand by the customers and are huge revenue generators for the institutions. Their scarcity and scheduling has been a challenge. Access to the timely advice for investment and insurance is important for short-term financial goals, long-term financial goals and for wealth management of customers. It becomes crucial in the beginning or end of the financial year.This challenge could hardly got addressed by the invention of “Robo Advisor” in banking, finance and insurance industry. Even though Robo advisors were implemented by a huge number of SMEs to large enterprises but because it lacked the intelligence and expertise needed to give a real advice hence it is primarily being used as tool to the finance advisors.

Ideation

Being a passionate learner, I have been learning about the ground breaking inventions in GenerativeAI field. I learned how much amount of data samples is made available to the large language models for their training. I felt that it is natural for a human be a decently knowledgeable in the investment banking and insurance areas if trained with such a huge amount of data. But I never heard anyone mentioning about realizing financial advice as if the elephant in the room was being ignored. Hence I dreamt with open eyes about building financial and insurance advisors on the Generative AI platforms powered by the large language model. And my confidence kept on increasing as I learned and mastered the IBM’s watsonx? platform for generative AI and grantite models.

Tossing the idea | IBM watsonx challenge 2024

Despite having the required learning, a dream and confidence of building on innovative idea of a unique Robo Advisor - I was not able to prioritize the same while being busy in the client engagements to accelerate IBM deployments in my role as a CSM Architect in IBM. Then came the announcement of IBM watsonx challenge 2024 for the employees. I felt as if it is same opportunity I was looking for. I jumped into the challenge, created an internal team (chances were less to lead through my idea - had I joined another team). So I created the team as a lead named "magnificent advisors" invited my colleagues to join the and tossed my proposal of building a Robo advisor. I'm fortunate to have great colleagues they supported me, but asked tough questions and challenged me. That challenge encouraged me to bring the best in me, and I could prove my idea through a feasibility check.

Knowing the terrain | Tools needed

To realize the Robo Advisor powered by generative AI - I needed a set of large language models and a platform for prompt engineering and tuning to experiment on the models to get the desired advice through the right prompt. Being in IBM, working as a CSM architect and as a participant for the watsonx challenge, I had access to the platform - IBM watsonx.ai and a huge number of large language models.

IBM watsonx.ai is part of the IBM watsonx platform that provides foundation models and generative AI capabilities, plus traditional AI modelling. Watsonx is IBM's new integrated AI and data platform to enable the entire life cycle of foundation models (The details of IBM watsonx can be referred here https://www.ibm.com/watsonx ). There are three main components in watsonx . ai :

1. Foundation model libraries – these include many well-known models such as Codex, Generative Pre-trained transformer (GPT), Fine-tuned Language Net (FLAN), and more, as well as IBM’s own foundation models.
2. Prompt lab – Prompt engineering is a way of helping the model to be more accurate and precise in its response by providing some user input (called prompts). Prompt engineering is an art and requires practice. IBM provides samples of prompts, and the studio helps users to try different prompts and different models to get the best results.
3. Tuning studio – there are multiple ways that watsonx . ai lets clients fine-tune the foundation models:

• Changing various parameters

• Providing additional data to further train the model

Foundation models

All watsonx.ai models are trained on IBM’s curated, enterprise-focused data lake. IBM has collected the largest known repository for enterprise-relevant training data covering a variety of topics like legal, finance, human resources, and more. This data was then further cleansed and curated (removed duplicate dates, filtered out bias, copyrights, and more) to create a base pool of data. IBM foundation models have fully auditable data lineage and transparency – allowing clients to use them out of the box with trust.

IBM has established a training process – based on principles of trust and transparency – that starts with rigorous data collection and ends in control points for ensuring responsible deployments of models and applications for governance, risk assessment, bias mitigation, and compliance.

Using this trusted data, IBM is building different models that can be applied in different scenario

IBM watsonx foundation models?library?gives you the choice?and flexibility to choose the model that best fits your business needs, regional interests and risk profiles?from a library of proprietary, open-source and third-party models. Some of these models have been rebuilt on IBM’s data sets to reduce model sizes, mitigate problems (such as bias) and other common risks, that can be plugged into extra applications without migration.

In addition, IBM has a partnership with Hugging Face to provide clients access to open-source models. IBM works jointly with Hugging Face and surfaces the best models suited to clients’ business needs. These models are of different architectures (decoder, encoder, or encoder-decoder) and sizes (as signified by the number of parameters in their name). Different foundation models are optimized for different tasks. Depending on whether a client is trying to summarize, generate, or classify; different models may perform better.

Granite, developed by IBM Research

Granite is IBM's flagship series of LLM foundation models based on decoder-only transformer architecture. Granite language models are trained on trusted enterprise data spanning internet, academic, code, legal and finance. Two models in the Granite series I selected for my use-case of Robo Advisor.

1. Granite 13b chat: Chat model optimized for dialogue use cases and works well with virtual agent and chat applications
2. Granite 13b instruct: Instruct model trained on high-quality finance data to perform well in finance domain tasks

Step by step approach

Besides the platform and foundation model library, I needed customer details to generate the desired financial advice and insurance advice which included: customer profile, dependents, assets and liabilities, short term and long term financial goals.

In a nutshell, first a model is selected, a customer profile is created, the relevant financial goals to be supplied followed by the prompt engineering and tuning. But the complexity is that all the these are variable. So I had to do my research on multiple permutations and combinations of prompts on different models to get the desired financial advice. I had to do the same for seeking the insurance advice as well.

I observed that if all the inputs were supplied in one shot and if all the advices were sought together, then the generated advices were not as expected. In this case IBM watsonx.ai platform came very handy as it remembered the history of the information supplied in a given prompt session.

Note: The financial advice did not include on the specific portfolio or funds - this article points towards the possibility for the banking and financial institutions to innovate by building on a foundation model and training it on their organization specific funds to generate the specific financial advice best for them and their customers. Similarly the Insurance advice did not include the Insurance company specific advices. The details of the advice varies on the prompts supplied.

Through my research through all the models, I was astonished to see how Granite Model is able to provide an advice which is of an immense value and will easily take from months to years for a person to reach to that level of understanding. Although in the challenge we were focussed to see how the solution is better for an IBMer, what we arrived at is looking like a huge time-saver for a non financial domain person. Granite 13b Chat model was able to generate the best advice with much lesser effort in prompt engineering.

The example of the basic prompt and financial advice is the following which was experimented on IBM watsonx.ai platform is below:

Nitin: 03:29 PM        

John Doe is a Software Engineer working in IBM Canada and is earning a medium salary range in Ontario, Canada. He is looking for investment advice to save taxes, secure his retirement, create tax free savings as well as have funds for his children education. What are investments options available for John?        

watsonx 03:29 PM        

Greetings! I am Granite Chat, created by IBM. I'm here to provide information and answer questions to the best of my ability. However, I should clarify that I cannot provide personalized financial advice or recommendations. I can share general information about various investment options available in Canada.        

For John Doe, considering his location in Ontario, Canada, he has several investment options to consider:        

1.???? Tax-Free Savings Account (TFSA): A TFSA allows Canadians to contribute up to $6,000 per year and any unused contribution room can be carried forward to future years. The money in a TFSA is tax-sheltered, meaning it grows tax-free and can be withdrawn tax-free. This account is suitable for short-term and long-term savings goals, including education and retirement.        

2.???? Registered Retirement Savings Plan (RRSP): An RRSP is a tax-deferred account designed for retirement savings. Contributions made to an RRSP are tax-deductible, reducing John's taxable income in the current year. Withdrawals from an RRSP are taxed as income when withdrawn. John can contribute up to 18% of his earned income, subject to certain limitations.        

3.???? Taxable Investment Account: A taxable investment account is not tax-sheltered, but it offers flexibility in terms of withdrawals. John can invest in various assets such as stocks, bonds, and mutual funds without worrying about tax implications until the funds are withdrawn. This account is suitable for long-term growth and wealth accumulation.        

4.???? Guaranteed Investment Certificates (GICs): GICs are low-risk investments that provide a guaranteed return over a specified period. They are suitable for conservative investors who prioritize capital preservation over growth.        

5.???? Mutual Funds: Mutual funds allow John to pool his money with other investors to purchase a diversified portfolio of securities, such as stocks, bonds, and commodities. They are managed by professional fund managers and offer various risk levels and investment objectives.        

6.???? Exchange-Traded Funds (ETFs): ETFs are similar to mutual funds in that they provide exposure to a diversified portfolio of securities. However, ETFs are traded on an exchange like individual stocks, allowing for more flexibility in buying and selling.        

7.???? Real Estate: Real estate can be an excellent investment option for John, providing both income and potential for capital appreciation. He can consider purchasing a residential property, investing in rental properties, or even real estate investment trusts (REITs).        

8.???? Education Savings Plans: There are various education savings plans available in Canada, such as Registered Education Savings Plans (RESPs) and Education Savings Accounts (ESAs). These plans allow John to save for his children's post-secondary education while benefiting from tax incentives.        

It's essential for John to conduct thorough research and consider his financial goals, risk tolerance, and investment horizon before making any investment decisions. Additionally, he may want to consult with a financial advisor to ensure that he makes informed choices tailored to his unique circumstances.        

Keep in mind that this information is accurate as of the time of my last update, and regulations and offerings may have changed. It's always a good idea to check the latest information from reliable sources for the most current details.        

There were many more examples of much better prompts and more precise financial advice I came up with. And similarly I got the insurance advice as well

Conclusion

IBM Granite 13b is a market leader in foundation model and a game changer for the finance and insurance industries. It reduces the amount of training needed in making an expert,?seasoned and easily available financial advisor and insurance advisor. It is a huge time-saver and a revenue generator for the banks, Investment banking organizations, insurance sector companies, wealth management organizations.

Without Granite model and?watsonx.ai ?platform it would have taken at least 2 hours each of discussion back and forth between financial advisor, insurance advisor and John Doe that would have been a loss of 4-5 productive hours from his work at IBM. With the tuned prompt it provided financial and insurance advice to John in less than 5 minutes. That is huge time saver for IBMers. It saved more than 98% of time.

The idea of creating an advisor is replicable across the organizations though it was experimented through the prompts for IBMers. There is no specific PII information of any IBMer was passed to generate the advice. The prompt engineering done on the watsonx.ai platform can be done in the background keeping space for creating an application around this idea. Also IBM's other offering watsonx.orchestrate can make it simpler to create such advisors.