Motivation

The motivation for developing the Investment Research Assistant stems from the need for Relationship Managers (RMs) in banks to efficiently query a stock universe and provide comprehensive investment advice to customers. Traditionally, RMs must navigate through a cumbersome click interface to find relevant stocks, particularly when specific criteria such as sector interest or geographical focus are involved. Furthermore, once stocks are identified, RMs need to delve into fact sheets to provide detailed rationales for each recommendation, explaining why a customer should invest in specific stocks or bonds. This process is time-consuming and labor-intensive. The Investment Research Assistant leverages GenAI to simplify this workflow, allowing RMs to use natural language queries and automate the extraction of relevant stock data and fact sheet insights.

Goal

The goal of the Investment Research Assistant, developed by Unique, is to streamline the process of querying a stock universe and preparing detailed investment recommendations. By integrating GenAI capabilities, it enables RMs to efficiently generate queries, extract pertinent stock data, and produce comprehensive rationales from fact sheets. This reduces the time and effort required to provide high-quality investment advice, enhancing the overall efficiency and effectiveness of RMs.

Structure and Logic of Assistant

Step 1: Extraction of Stock Data

The process begins when the RM inputs a natural language query specifying the customer's interests and criteria (e.g., "My client is interested in stocks with a 5 star Sustainability Rating"). The system then generates a query to an in-house or CSV-based stock database. The system extracts the most relevant stocks based on the given criteria, considering factors such as sustainability ratings or buy signals.

→ The extracted list of stocks is then prepared for further analysis.

Open

Step 2: Extraction of Fact Sheet Rationales

In this step, the system loads the fact sheets of the stocks identified in Step 1 into an LLM (Large Language Model). The LLM processes each fact sheet to extract the most relevant rationales, providing detailed insights and justifications for why each stock is a suitable investment. The extracted rationales are then compiled into a coherent narrative.

→ The compiled rationales are presented in a chat interface, allowing the RM to review and refine the recommendations.

Open

Step 3: Generating Follow-up Email

Once satisfied with the extracted stock data and rationales, the RM can generate a follow-up email to the customer. This email includes a detailed investment story, attaching relevant fact sheets and any additional necessary documents. The RM can customize the email content and attachments as needed.

→ The final email, complete with fact sheets and recommendations, is ready to be sent to the customer.

Possible Adaption of Assistant

The Investment Research Assistant can be adapted to various scenarios, enhancing its flexibility and applicability:

• Custom Data Integration: Users can upload their own stock universe as a CSV file and attach their own fact sheets for analysis.

• Tailored Outputs: The assistant can be customized to generate different types of outputs based on specific user needs, such as creating tailored investment stories.

• Custom Query Optimization: The system allows for the integration of custom few-shot learning examples to improve the accuracy of query generation.

Required and optional modules

The following modules are required/optional for this assistant:

Example AI Assistant Configuration:

Download the InvestmentRecommendation.zip file and unpack it. Upload the documents from the ZIP file to the Knowledge Center. Next, download the provided investment_research_assistant.txt file and upload it into a new space as an AI Assistant configuration.

Prompt Engineering Guide

By following these guidelines, you can enhance the effectiveness of your prompts and queries, leading to more accurate and useful responses from the Investment Research Assistant.

Extraction of Stock Data

If your initial query leads to no results, consider adjusting your search criteria. Here are some strategies:

• Refer to Your Stock Universe: Ensure your query aligns with the available data. For instance, if your stock universe includes features like sustainability rating and stock price, focus on those features. You can start by querying, "Give me one stock," to see all the features available.

• Be Specific in Your Queries: Provide clear and detailed context to avoid assumptions. Specify all relevant details to get accurate responses.

  • Original Query: "My client likes to have stocks in Germany with a good rating."

  • Modified Query: "My client likes to have tech stock in Germany with a buy rating."

Extraction of Fact Sheet Rationales

To effectively extract fact sheet rationales, be specific about what you need. If results are not as expected, adjust your query for more precision.

• Reference Previous Results: When referencing previous results, remember that you can only go back one iteration. Use the last message from the assistant to guide your next query.

• Identify the Correct Stock: If the investment rationale for the wrong stock is extracted, be specific. Instead of saying “Give me the rationale for the first stock,” say “Give me the rationale for AstraZeneca.”

Generating Follow-up Email

To generate effective follow-up emails, provide detailed instructions on content and tone. Follow these steps:

• Produce Fact Sheet Rationales First: First, produce the extraction of fact sheet rationales so that the LLM can see what can be used in the email text.

• Specify Tone and Length: Indicate whether the email should be long or short and the desired tone (formal, friendly, etc.). Example: "Generate a follow-up email with a formal tone, including detailed rationales for the above stocks."

General Prompt Engineering Tips

• Use Table Format: If you want a table format and the output is incorrectly a rational of a factsheet, explicitly state your requirement. "I want to see these stocks in a table format."

• Detail Extraction Requirements: When extracting from fact sheets but a table is created, specify that you want to extract from the fact sheet. “Extract the sustainability rating and growth outlook from the fact sheet of Microsoft.”

By adhering to these guidelines, you can optimize your interactions with the Investment Research Assistant, ensuring efficient and accurate data extraction, rationale generation, and follow-up communication.

Key Differentiators Compared to General Chat Solutions

These unique features position the Investment Research Assistant as a superior solution for Relationship Managers, offering unmatched compliance, accuracy, sector-specific insights, and integrated data handling.