My Last Financial Model
A couple of months ago, I wrote about Terence Tao’s proclamation that mathematics is evolving from an “era of proof scarcity to an era of proof abundance”. The more I use the frontier models, the more I realize this may be the eventual fate for most knowledge work itself. Let me explain why I had that epiphany over the last few days.
For every company Deep Dive, I build a financial model for the company from scratch. I am well aware there are many investors out there who simply do not see the point of building such models. I too strongly discourage anyone, including myself, from falling in love with their models. However, the primary point of the exercise, at least for me, is to let the model help me visualize the economics of the business and more importantly, make the assumptions for the future much more explicit so that I know what I need to underwrite to make decent return from an investment. The reality is whether you are building the models or not, you are always underwriting a lot of assumptions for the future. A spreadsheet just lays it out in a visual form rather than (often hand wavy) mental math. I do acknowledge that the more adept an investor is, the less he or she requires the spreadsheet to guide such an exercise; such adept investors are usually also quite disciplined enough to make the mental math a lot less hand-wavy. This is also why Warren Buffett didn’t ever need to touch a spreadsheet and yet probably understood the essence of the economics of most businesses and the implicit assumptions for such businesses in much greater depth than most people with their thousand line spreadsheet models. For mere mortals such as myself, I always thought spreadsheet can still be an excellent tool to guide your intuition and understanding around the assumptions of the drivers of different businesses.
Since launching MBI Deep Dives back in 2020, I have personally built such models for 70 different companies so far. As the promise of AGI grew louder over the last four years, I tried using the frontier models to build my models. While they could build a decent model, it was often more frustrating and time-consuming than building one myself, as they usually struggled to build it the way I wanted. Even Opus wasn’t great at this, and I continued to build the models myself. However, after trying Claude Fable 5.1, I can say with confidence that I have already built my last financial model from scratch.
While working on my current Deep Dive on Uber, I asked Claude to to look at my past financial models on companies such as Booking, Airbnb and DoorDash and use them as a guide for building a decent shell for a model on Uber. I was pleasantly surprised by the quality. My confidence in Claude’s ability grew even more a few days later when I started working on building a model on OpenAI. I started working on the OpenAI model with a broad sketch of what I was trying to accomplish. I hope to publish my work on OpenAI tomorrow, so I will skip the granular details here but I thought it would be useful to explain my thought process about using AI in more detail first. The frontier models are essentially so good at building models that it is a very poor use of time for me to build one from scratch. However, the job of an analyst or investor is likely changing.
Once I asked Claude to build OpenAI’s compute capacity in three buckets (training, alignment & safety, and inference) and use those capacity assumptions to drive revenue and margins, Claude followed my broad sketch and built a very decent shell of an OpenAI financial model. I was essentially just reviewing Claude’s models line by line and asked for clarifications and debated the assumptions it made. It provided specific sources and rationales for the assumptions made in the model, but when I probed certain assumptions, it sometimes accepted that it needed to do more work to establish the rationale for those assumptions and sometimes I myself accepted its arguments and moved on. Through this back and forth, Claude ended up building the model half a dozen times already. I’m not done yet, so by the time I publish it tomorrow, Claude may have to generate the model dozen times from the initial shell it built for me.
As you can imagine, Claude is already perhaps top 1 percentile or higher in excel proficiency. But building the model still requires a tight collaboration and feedback loop between the user and the tool. “Hey Claude, build me a financial model for OpenAI. Make no mistakes” will give you at best an okay-ish result, but unless the user has a good idea about what “good” looks like and in what ways Claude/ChatGPT need to be probed to elicit a better way to visualize the drivers and economics of the business, the frontier models won’t be able to take you there on their own. Using the same prompt for every company also doesn’t seem to give the best outcome. Knowing what the end product should look like is still a clear advantage…for now!
Unfortunately, many people have a lot of preconceived notions about AI and its effectiveness. I have no qualms in admitting that these models are much more proficient at Excel than I ever was. Moreover, the speed and ease with which Claude can incorporate different scenarios and new information were simply inconceivable even a year ago. The OpenAI model that I will publish tomorrow will focus on a couple of specific things I am trying to understand, but I certainly do not expect everyone to agree with the assumptions made in the model. So, my suggestion for you would be to download the model, upload it to your favorite AI, and just ask it to change the assumptions to cater to your narrative or point of view. This is, of course, a suggestion for every model going forward. One of the recurring emails that I receive from subscribers is “when will you update the company XYZ model?” You no longer need to wait for me. Just download the model and ask the AI to update it. Of course, you will still need to steer the model by asking good questions. I should caution you that the model building exercise with Fable 5.1 (extra high) is a fairly token intensive process and can often eat up your entire token limit in a particular session. I will have more thoughts in tomorrow’s post related on OpenAI.
Of course, if advanced mathematics is moving from “an era of proof scarcity to era of proof abundance”, financial modeling is understandably child’s play for the frontier models. Frankly speaking, I won’t miss building these models from scratch. While going line by line for the model Claude built for me, I actually found the job to be just as enjoyable. It’s hard to imagine anyone entered this industry because they like using excel. But the core job of understanding, and thinking through the economic drivers to help allocate capital will likely prove to be timeless unless you simply give up and ask the frontier models to manage your money.
Subscribers get the daily journal and five+ years of Deep Dives, i.e. full-length analyses with financial models on 70 companies. The daily is just how I think out loud between the Deep Dives!
Current Portfolio:
Please note that these are NOT my recommendation to buy/sell these securities, but just disclosure from my end so that you can assess potential biases that I may have because of my own personal portfolio holdings. Always consider my write-up my personal investing journal and never forget my objectives, risk tolerance, and constraints may have no resemblance to yours.
My current portfolio is disclosed below: