Let's wrap it all up with a case study. At Theta Equity Partners most of the time were working with private companies, private equity firms looking to buy some private whatever. But just for fun, we'll look at IPOs. When a company is about to go public, they'll often put out in their S-1 filing. There's a lot of fanfare about it. The company announces its intention to go public and there's this big document and they say, "Here's a bunch of metrics that describe the company." Usually, I talk about these things as trophy metrics. It's like, "Hey, we can make pretty charts" and that's pretty much it. But if they give us the right kinds of data and enough of it, then it turns out to be just enough for us to reverse-engineer and pull out the Buy Till You Die model. Again, exactly the same models we've been talking about all along, but estimated on different data inputs. Once again, big shout out to my former student and co-founder, Dan McCarthy, for doing a lot of the math that shows us that if we get the right metrics, we really can match up those metrics with different combinations of the model parameters to pull out the insights that basically to put us in a position where it's as if we had the transaction log data in the first place. Let me give you an example right over here. We've done this many, many times, but I want to talk about a company called Farfetch. It's a nice company. They're an online luxury marketplace headquartered in the UK. When they went public back in 2018, actually in their F-1 filing, because they're a foreign company, they put a bunch of metrics out there, and you can see them right here. For those of you who've done anything finance-related, you'd be looking at these metrics and going, "Okay. Yeah. Fine. Whatever. Yeah." There's some descriptions of your customer base. But it turns out that this data is just enough for us to do the reverse engineering and run the model. I want to emphasize what we're about to do here. You're looking at these figures, again, this is exactly the way Farfetch puts it out there, literally cut and pasted from their filings. We are not coming up with a model that's going to fit this data as well as possible. It's kind of a machine learning thing, we've already covered that. We're coming at this in a very principled way. I know exactly what story I'm going to tell about how many customers I'm going to acquire, and then at the heart of it all is a Buy Till You Die model. I just know in my heart of hearts that Buy Till You Die is the right model for them. I'm not looking at the data. I'm not saying, well, given the kind of company, I just know it's the right model to run. It turns out that they're giving me just the right amount of data to run it. They provide all of this data. We run our Buy Till You Die model, of course, the specific parameters that we get will be a function of their data, but the overall model specification is something that we've locked in before we touch any of the data at all. Let me show you on this next slide over here how well things match up. I hope you'll be impressed by it. If you look over on the right, if you take the aggregate metrics that we saw in the upper left of the previous slide, how many active customers were there? How many orders did they place? You can see how amazingly well we're picking up each of those bars. But then when you look at the top left, that's what we call the C3 chart, the customer cohort chart, it's fairly self-explanatory. It's basically saying for each cohort of customers we acquired in each different year, how much total gross merchandise volume? How much revenue did each of these cohorts provide in the years to come? You can see the way those revenues are working their way down, consistent with the Buy Till You Die model. You can see each of these cohorts layered on top of each other. We fit our BTYD model in addition to the spend model and the acquisition model, which are very, very simple. Happy to provide more details about those. If you look at the lower left, you can see how amazingly well the Buy Till You Die model picks up on it. Basically, it captures everything worth capturing and gives you that good feeling in the belly that we could really take this model and start to do things with it, even if it's things that aren't revealed in the raw data itself. For instance, on this next slide, here's the retention curve. Now, you should be very careful about using a word like retention, because as we've discussed before, we tend to save that for a contractual setting. Of all the customers we had last time, how many are still with us? Well, in this case, it's the equivalent of that, but it's non-contractual. It's simply saying: Of a given cohort of customers, what percent of them continue to be active as we move one, two, three, four months, years ahead? Notice the shape of the curve over here. You see that really steep drop and then it levels off? That's interesting. We'll always see a drop, but seeing the drop so steeply and then leveling off tells us that this is a very heterogeneous customer base. It's somebody who said there's a bunch of customers who buy once and go. Not that great. But when it levels off, it's saying there's a large group of customers who are going to stay with this company basically forever. That's the lifetime value, that's the customer asset value that companies want to see, and not all of them have the luxury of seeing it. Farfetch does. Even though there's nothing in the disclosures that tell us specifically about how customer retention varies over time and how it varies across customers, we can pull that out, because we understand that the Buy Till You Die model is working so well that it lets us make statements about behavioral aspects that aren't even included in the dataset but are consistent with the model. We can do that for each and every one of the different behaviors. How many customers will we acquire? How long will they stay? How many purchases will they make? How much will they spend when they do? Again, those things are not directly revealed in the filings, but we specify this holistic model, it pulls them all together, we estimated, it works pretty well, and then we can start to pull them apart. As you can see on this chart over here, we show our ability to capture and then forecast each of those separate behaviors. Once again, if we can capture and forecast each of those behaviors, that gives us overall revenue. We can start to make statements about what the revenue of this company will be, and therefore what its stock price will be, based on these relatively simple, seemingly innocent customer metrics. I'm going to stop right out. If you're interested, you can go further and on the Theta website, we actually post the full description of this thing. Just to tell you real quickly, a month before the company went public, we estimated that they were worth about twenty dollars a share. When they opened on the market, they went much higher and then they went much lower. Basically, we have a number of blog posts that just talk about what the company's worth and how that value changes as the company itself changes some of its ongoing policies. There's nothing special about Farfetch. Again, it's a company I admire. But if you're interested, we've done the same thing with Slack, the enterprise SaaS company. We've done the same thing with Lyft. We've done the same thing with Revolve, the women's clothing company out of the West Coast. That's just beautiful. We can take basically the same modeling framework, the one that you've seen and hopefully enjoyed through all of these different sessions with me, and reverse-engineer what a company is actually worth, and then go back to the marketing people to say, "Hey, you have these loyal, sticky customers, what are the services should you be offering for them?" Takes us back to customer centricity as well. I hope that you enjoy the models. I hope that you've learned a lot about Buy Till You Die. I hope you've learned a thing or two about just modeling in general and how it compares with machine learning, and just a lot of the issues that arise from putting a model together, as well as a lot of the managerial motivations for why you'd want to do so in the first place and the improved decisions that you can make afterwards. Again, it's a conversation that I would love to keep going with you. Please do connect with me. It is your own questions, your own feedback, or at least those who've taken the course before you, that motivated me to put this additional content together. There's plenty more where that comes from. Again, I wish you the best with your different modeling adventures and the decision-making that you're going to make on the basis of them. Then, again, let's keep in touch beyond that.
