This is a proof of concept (POC) for AlphaInsight. A co-pilot utilizing the agentic properties of large language models (LLMs) to help users with complex reasoning and planning involved in financial research.

"Alpha" in finance refers to the excess return on an investment strategy relative to benchmarks, and is often used as a measure of an investor's skills. Drivers of alpha include deep understanding of a domain, foresight and early recognition, of emerging issues, and unique insights. LLMs with their proven ability to store and recall information at super-human levels can empower investors in their research process if designed correctly.

Questions and research in finance often requires aggregation across sources, assumptions and estimations to work around information gaps. The typical Retrieval augmented generation (RAG) applications tend to fall short in these areas, performing only when answers to question/queries are found directly in documents. This model and co-pilot application has been specially trained and designed to "reason" and breakdown complex objectives\questions (to which answers may not be directly found from sources) into a series of sub problems (a plan), and with optional human guidance solve these complex research questions beyond the reach of typical RAG applications.

Usage:

• Begin by entering a question/objective for the co-pilot. E.g "Estimate the impact of Nvidia's delay in the launch of their H20 chips in the Chinese market on forward guidance." and hitting the "Query" button.


• The agent will develop a plan and start executing it to solve for the question/objective. Depending on the complexity of the task it may require user guidance. Refine the plan and provide guidance by editing the steps proposed by the agent.
  • Change the details of a step by editing the response from the agent by typing in the response bubble.
  • Insert additional steps by clicking on the "Insert Above" or "Insert Below" button and typing in the details of the new step.
  • Remove a step by clicking on its "Remove" button.


• Continue with further follow-up questions/objectives taking previous outputs into context by entering new queries and hitting the "Query" button. Rinse and repeat by editing the proposed plan from the agent.