Training GPT3 to be an Investment Banker

Hey IH community,

We have been experimenting with using GPT3's reasoning capabilities to enhance our finclout.io platform to perform the work of an equity analyst in an investment bank.

We started by prompting GPT3 to do this:

(1) Summarize relevant news of the last 30 days,
(2) Write a SWOT analysis,
(3) Identify competitors,
(4) Identify Key Risks, and
(5) Write three investment theses (bull, neutral, bear)

Then we combined these datapoints with our existing data on finclout.io into visually pleasing dashboards and presented this to a selected group of test users. (We currently have about 7K unique monthly users)

Here are some examples how these dashboards look like:

https://app.finclout.io/tp/MSFT
https://app.finclout.io/tp/LYFT
https://app.finclout.io/tp/AAPL
https://app.finclout.io/tp/UBER
https://app.finclout.io/tp/BBBY

From an architectural perspective, we embedded the GPT3 API calls in
well-engineered prompts that return json as datatypes in a pre-defined format. That helps parsing the information in the right buckets. The returned data especially around news summarization, competitor analysis, and key risk was really impressive.

One notable downside from this integration was surprisingly API rate limits. Even though we pay for the service, have waiting periods between calls, and only allow certain users to trigger the call we frequently didn't get data back because of rate limits.

In some cases we observed obvious errors.
"Uber Technologies, Inc. reported a surge in fourth-quarter revenues and a strong outlook for 2021." We are not in 2021 anymore for a long while.

Clearly, GPT3's knowledge of the Yahoo finance dataset has been cut off in 2021 and applied the provided news effectively.

In conclusion, while we need to further investigate if the actual reasoning of GPT3 is mature enough to provide a useful analysis of the stock the results are extremely promising.

One potential pitfall will not be a legal problem but rather a regulatory one. What happens if GPT3 returned factually incorrect information leading to a liability exposure.

Clearly, that is only an experiment and shouldn't be used as an investment recommendation nor seen as investment advice in any way.

But if it works, it could be a real game changer.