Using Chat GPT For Deal Analysis

To say Chat GPT is useful isn't very insightful, but I recently got access to their advanced data analysis plugin and it's amazing. If you haven't poked around with it yet, you should!

For a deal we currently have under LOI, I've made great use of it to ask all kinds of data related questions that would normally take up to 10 ish minutes each, or at least annoying data wrangling in excel.

Flip Fund

We will hopefully have exciting news next week about the Flip Fund target. We're in the manic final days before a deal closes.

I will say it seems across the board that even growth-ey companies have decelerating growth. This seems to be true based on ~10 ish deals we've looked closer at recently. It makes for an interesting negotiation.

One diligence question we came up with after looking at 1 specific company was around growth experiments. Danny came up with the idea:

I think rather than asking exactly what worked or what didn’t… the better question more plainly what did you do and how often/how much did you do it. - Danny Chu

The question tries to get at a more important point for us as the new owners which is just documenting experiments that were run so we can think through our own experiments to run. It's actually very difficult without excellent attribution to say ads "worked" or didn't during a particular period. Yes, turning on ads can easily force growth in signups for example, but it's harder to say definitively "The CAC was $1k and our LTV is $500 so ads are not profitable". Let's say ads were run for 3 months, but randomly during those months, a youtube video picked up and drove a bunch of traffic as well as a blog post that started ranking better. Unless you're very careful, you might accidentally attribute the growth to the ads.

Using Chat GPT Advanced Data Analysis

All you need to do is upload a CSV and make sure the "Advanced Data Analysis" is turned on.

You can use a simple export of Stripe customers as the base data set. Be sure to export the "status" column so you can look at active subscribers, vs delinquent, cancelled, etc.

Once you upload the CSV, you can just ask it what kinds of regressions you could run, you don't even have to know what question you want to ask yet!

While it will output "something", it's not always clear what the heck it did, so be very careful using the output directly.

The next level is to just tell it to output the regression in R for example which you could then run yourself to get the output and be sure it's doing the thing you expect it to.

Prompt engineering in these data scenarios seem to be less like alchemy. The y-axis in a regression had the wrong label and I pointed it out. Chat GPT recognized it and asked if I wanted to run the regression.

The subsequent result still had the error, but this is still very impressive. I ultimately had to modify the data to add the date in a different format to get it to work.

The things that worked well were more data wrangling tasks such as these

or asking it to bucket customers by how long they've been a customer:

Overall i think it's an amazing tool to answer questions yourself instead of asking the seller to do this type of work, or an analyst, or doing it yourself. If you haven't yet tried this out, you should! It will at worst, save you a lot of time and let you ask many many more questions about a dataset than you otherwise would if you had to answer each one manually.