equity-analyst

equity-analyst makes the hard to grasp financial reports to, easy to understand actionable items to a common person.

Comment

Inspiration

We believe that the profession of equity research is extremely time-intensive and manual. With the arrival of LLM techniques, the industry is due for a huge disruption. With the US stock market valued at $109T, equity research reports can inform and move large segments of the market. It is also extremely important for the reporting to be correct because analysts (or their employers) are liable for mistakes.

What it does

Creates an equity research report by following an analyst workflow defined using prompt templates. Interacts with complex PDF documents using a VectorStore with a recursive retrieval approach. Generates an interactive report that is tailored to audience expertise and enables the user to chat to ask more questions.

How we built it

  • Used LlamaIndex and LlamaParse to generate an index using the PDF documents
  • Performed extensive prompt engineering to imitate an analyst's workflow
  • Designed a UI based on categories that we know are common components of equity reports

Challenges we ran into

  • LlamaParse was not initially working well on the PDF document with tabular data
  • We were wondering if it would be possible to get the right numbers at all until we tried the recursive retriever
  • We originally iterated on multiple ideas before landing on this one.

Accomplishments that we're proud of

  • Found high-quality data sources with PDFs that are guaranteed to be correct
  • Did better than regular ChatGPT, better than a ChatGPT plugin, and better than the basic version of LlamaParse
  • Tailored the approach to an expert and a novice audience

What we learned

  • The choice of retriever engine matters a lot for results
  • LlamaParse still has some potential improvements, especially for parsing of tabular data, though you all are making great progress

What's next for equity-analyst

  • Provide portfolio strategies based on evaluations of many different stocks
  • Add news data from DiffBot into our RAG index
  • Include CSV input data
  • Visualizations and citations from PDF sources
  • Provide suggested questions for the user to interact with the documents => startup!?

Built With

Share this project:

Updates