During the Oct. 2022 Mountains, Capital and Commerce conference, Tadpull Co-founder and CEO Jake Cook sat down with Marketing Professor Dr. Peter Fader to both unpack his conference presentation, and to learn more about how his research is impacting private-equity backed exits.
Dr. Fader is the Frances and Pei-Yuan Chia Professor of Marketing at The Wharton School of the University of Pennsylvania. His expertise centers around the analysis of behavioral data to understand and forecast customer shopping/purchasing activities. He works with firms from a wide range of industries, such as telecommunications, financial services, gaming/entertainment, retailing, and pharmaceuticals. Much of his research highlights the consistent (but often surprising) behavioral patterns that exist across these industries and other seemingly different domains.
In addition to his various roles and responsibilities at Wharton, Fader co-founded Zodiac, a predictive analytics firm, which was sold to Nike three years after its founding. He then co-founded, and continues to run, Theta Equity Partners to commercialize his more recent work on “customer-based corporate valuation.”
Dr. Fader’s keynote address at the 2022 Mountains, Capital and Commerce conference centered around how to blend customer data to drive validations using the customer lifetime value metric.
Listen to this article
Cook: Pete, thanks for being here. Appreciate your time. Let’s just get right into it.
You started [your presentation] with customer-based corporate valuation. About a year and half ago you published an article on this topic with the Harvard Business Review. What’s changed since you guys put that out?
Fader:
A couple of things.
For one thing, many more people have noticed. Many more people have realized the kind of impact that [our research] can have.
And, secondly that it’s real, that it’s not just an academic exercise, but it can apply to real companies big and small. And as a result of that, we are working on more projects, with a wider variety of companies, changing the statistical methods, making them more comprehensive, more robust. And so it’s been just a natural evolution since then. But the publication of that article really was a big step forward.
Cook (laughing): Did you get some “nasty-gram” emails based on some of the takeaways you presented?
Fader: I can handle it. I like a good “nasty gram,” it kind of makes life interesting! I almost wish there were more. The worst that we get are from investors or from an “executive Sam” saying, “Yeah, we do that already. Thanks for validating what we already know and do.” But, even in those cases, they’ll still say, “But, you help us do it more boldly and more accurately.” But most people had the reaction of “I never thought about this before. Man, that makes so much more sense. Thanks for helping me open my eyes.” And, to anybody [that feels good], but especially to a professor, that feels really good.
Cook: And so when you hear the “we’re-already-doing-that” comment, what is your counter to that? What are the things that they think they’re doing, but they’re not?
Fader:
So, it’d be a couple of things. We’re always talking about a suite of models here: We’re modeling
- acquisition,
- retention,
- repeat purchase,
- and spend.
So, they might be doing some of this right, typically one component of it. The most common example is, a lot of people would be modeling retention. They would be looking at the [customer] churn rate. They might say, “Oh, our churn rate is 10 percent, and therefore people are staying around 10 years.”
Well, that’s a baby step in the right direction. Sure, and I’ll tip my hat. But there’s more you could do on that one behavior. And of course, you should be looking at the other behaviors as well. So usually, they’re not quite broad or deep enough, but I’m happy to see people trying it all.
Cook: [During your conference] talk you gave us four key pieces of information to orientate around. Can you repeat those for us and unpack for us why each one matters?
Fader:
That is the key and it’s worth unpacking.
Okay, the way we look at it is every dollar of revenue is arising from either acquiring a new customer, or a customer has chosen to stay with you longer, where they’ve made incremental purchases, or they spent a little bit more money than they would have otherwise.
Those four behaviors are: acquisition, retention, repeat purchase, and spend. There’s nothing new about them. Companies have been observing and managing those things forever. But they tend to be lost [or buried] way, way, way, way below [the surface.]
So taking those somewhat mundane behaviors and elevating them, evaluating them with more rigor, and drawing greater visibility across the organization to what they really mean. That’s what it’s about.
Cook:
And of those four metrics, where do people get misled by the data? Do you find they go down a rabbit hole that perhaps they shouldn’t?
Fader:
It comes back to the one I just mentioned: retention. It’s maybe the most important one, it is after all, called customer lifetime value. The question is: How long is that lifetime going to be?
In some cases, such as in a contractual setting, where we can measure a churn rate, and then infer lifetimes, but even in that situation we can do better.
But in the non-contractual setting, which is most businesses, you don’t have some kind of subscription relationship with the customer. So the best you can do is to look at:
- How often they bought,
- The last time they bought,
- And, try to project how long they’re going to keep buying.
Well, that’s hard right? Of course, that’s what keeps me gainfully employed. I raise that very question and come up with really good statistical methods to address it.
Cook: And the models for the non-contractual customers, the ones we can’t observe when they “die.” Can you talk about the concept of “buy till you die” and where that probabilistic stuff came from?
Fader:
Too often these days, people are using some kind of machine-learning model. And there’s nothing wrong with machine learning. But every tool has its place.
And when we’re talking about behaviors that are pretty sparse, meaning they don’t happen very often, over a very long horizon. It really is inherently probabilistic, for example who knows exactly when and why you’re actually going to die. In the end there is a big random element to it. So, we need to focus on models that really embrace that randomness.
One of the issues with machine learning is this desire to try to specify and capture everything … we want to just focus on the fundamental behaviors, we don’t want to sweat the small stuff.
And we want to acknowledge: long horizon and random. This is what actuaries have been doing for a long time–coming up with the life tables, and also knowing their limitations. Knowing that it’s going to be really hard to say exactly how old you are going to be when you die, but having an astonishingly good ability to say, if we look at a cohort of people who share the same behavioral characteristics, what percent of them will live to be 90 years old?
Yes, actuaries do that really well. And we’re basically borrowing those methods. And that whole perspective as we build these kinds of models.
So one of the interesting things is, we can drill down and make statements about individual customers. But, we’re just much more comfortable at the micro segment level, looking at customers that share characteristics to talk about that small group collectively, rather than any one individual.
Cook: So let’s talk about eCommerce. How would you advise people to think about cohorts?
Fader:
It depends on the business. So, if [the product] is something that you’d be buying pretty frequently, then you might want monthly cohorts, sometimes quarterly, sometimes yearly. So it really does depend both on the cadence of the purchasing, as well as the kinds of decisions that the company wants to make.
If I had to choose, I’d say quarterly. When you think about seasonality, because so much happens in Q4. We don’t necessarily care whether someone has been acquired in November versus December, that doesn’t matter as much. And when we look at these quarterly cohorts, we see those Q4 cohorts, always standing out and often in a not very good way. But doing that the quarterly analysis often helps you shine light on a lot of those holiday activities, often discovering that those Black Friday doorbuster sales are not a good idea, because we’re bringing in a bunch of bad customers that aren’t going to buy again, until we offer the same sale next year.
Cook: You famously wrote a great blog post on this topic, about Black Friday & Cyber Monday, where you recruit your worst customers and cut them a deal.
Fader: Yeah, treat them like royalty. And the problem is, I’ve been writing that blog post every single year. People say what’s your New Year’s resolution? Mine is to get companies to stop doing that!
There weren’t a lot of good things about COVID. But that was one good thing: Companies took their foot off the gas a little bit. And, I was hoping that they’d see the benefit and some of those [COVID] practices would persist. I’m not optimistic. I think that companies are falling right back into their old habits. Time will tell, but, I think in the long run, we will see a ratcheting down of some of that Black Friday activity. And, one of the big reasons why is that companies will look at those quarterly cohorts and just realize the mistakes they’re making.
Cook:
If you’re an eCommerce manager, and you’re interested in lifetime value, maybe you go into Google Analytics, or Shopify, or something that spits back an LTV or CLV number to you, how would you recommend getting started in scratching beyond that?
Fader:
It’s sometimes dangerous. I don’t want to disparage any one of these companies, but they’ll put a lifetime value number out there. And very often it will be a singular number: here’s your LTV to CAC ratio: your lifetime value relative to the cost of acquisition, as if every customer’s worth that amount.
So, the first scratching of the surface to be done is to break away from that singular number and understand the nature of the spread.
Understanding how many high-value customers versus how many so-so ones do we have? And even if we don’t get the full, beautiful, continuous picture that I like to draw, even if it’s just three segments, to show me the relative size and nature of those three segments of customers, that will still be a good step in the right direction.
- Let’s look at, let’s celebrate, let’s understand what that spread means.
- Let’s not take these models for granted. Let’s validate them.
So again, if we look at a quarterly cohort of customers and we come up with these predictions, not just what they’re going to do over their entire lifetime, but what they’re going to do over the next year or two. Then, let’s go back a year or two later and ask, “How were those predictions?”
The problem is companies don’t do that. And even if they do that, they’ll see predictions that were often say, “Well, we couldn’t have foreseen this change.” And, they don’t learn from it. They dismiss it. You gotta learn from it. It’s okay that forecasts aren’t perfect. They never are. But, let’s learn from those mistakes or those inaccuracies. Whether it is a mistake on our part that we can fix or whether it’s some random thing that happened.
Cook: But this comes back to a cultural piece. Right? So are there things you’ve noticed from some of the companies you’ve worked with, from a cultural perspective?
Fader:
So, we have to talk about culture at two levels:
So first, most importantly, there’s just the overall corporate culture.
We have some folks who are doing these kinds of analyses and making [data-backed] decisions. Are they “locked up in a backroom and only let out for air” when people have questions for them? Or, are people actively and frequently engaging with them and caring what they have to say and holding themselves accountable for what these models and what the modelers are saying?
So first, we have to have this culture that’s going to be more data driven, more inquisitive, more scientific, so don’t just tell us what the numbers are. But let us know what’s going on below the surface.
So that’s the biggest cultural problem of all, and I really pride myself on trying to get that big, broad, cultural conversation happening. One of the reasons I’m writing books on customer centricity, I’m not aiming them at the people crunching the numbers, they don’t need these kinds of books, they’re already reading my papers. I’m aiming [to get them in front of] the other people in the organization who tend to be downplaying the importance of [data analysts].That’s the big culture conversation.
But then there’s the micro one, which would be among the analytics people, this idea that there’s a lot of folks who are just just locked into machine learning. And they have this idea that it will fix everything, and that we’ll come up with the best forecasts for everything. And getting them to be a little bit open minded, and getting them to recognize that they actually need to expand their toolkit a little bit, and understand where these probabilistic models fit in with machine learning (ML), and in some cases actually improve upon it. I’m not saying, “Put your ML skills aside,” I’m suggesting a broadening of skills that enable you to use the ML findings more effectively.
Cook: Let’s go deeper into machine learning for a second. It’s a hot topic for software companies and many believe that putting .ai at the end of their domain is a surefire way to increase their valuations. But if you had to stack-rank data, and how much predictive power it has, what would that look like to you?
Fader:
It’s such an important question. Having that discipline to know what data is essential and what data is just nice to know. So, takes us back to the fundamental behaviors: acquisition, retention, repeat purchase and spend.
If you can give me a metric, maybe two, if you’re going to be really generous, that’s going to let me capture each of those processes and the heterogeneity, the differences among the customers with regard to each of those processes. Now, I’m like a kid in a candy store.
I can get away with a surprisingly small set of metrics. And it’s a surprisingly consistent set of metrics for seemingly unrelated kinds of businesses. Very often when we’re working with a firm, and they “let us into the data room” we find–if you couldn’t tell by my metaphor choice–that there is just too much data.
So, instead of [taking all of the data], I’m selective: “I’ll take this retention metric over there, I’ll take this active-user metric over here.” We know what we want, doesn’t mean we can always get it. But, the key is having the discipline to use as little data as possible to run these models effectively. We could look at some of those other metrics and, very often, we’ll bring those things in after we run our models, then we’ll see if it can help us slice and dice and find some greater insights. So that one-two punch, run our probabilistic models, and then throw some machine learning on top of it. Very, very powerful. But you just have to prioritize.
Cook: So say, for example, you bring in a net promoter score or some of these other ancillary datasets. Have you found, in your experience, broadly speaking, it’s made a big difference in the predictive power?
Fader:
Not at all. But it is a wonderful way to do the post-hoc slicing and dicing. So instead of just looking at the NPS, for the company as a whole, let’s look at the NPS differently for customers based on their projected financial value.
You’ll find some interesting insights about it. You would think that all the high-value customers are promoters and all of the low-value customers are detractors. It doesn’t work that way, there’s still going to be a spread within those groups. And there could be a lot of insights about it. [You might find] that even some of your high-value customers are pretty critical about you. And even some of your low-value customers, it’s not that they hate you, they love you, but only once a year.
And so looking at something like an NPS or, or other engagement metrics and on top of these lifetime value segments, very powerful, very practical.
Cook: Let’s shift gears a bit and talk about due diligence for private equity investors. Based on your presentation, I understood that a powerful tool is good old-fashioned, boring customer-transaction logs, right?
Fader:
Yep, it is. And we really want to get operators, not just investors, to be looking at these things and understand that a lot of the transaction-log data isn’t just “nice to know” but it’s mission critical.
And even if they’re a small company, they need to get comfortable with it sooner rather than later. Because, what happens with too many companies is they grow, grow, grow and as a result they expand their footprint and expand their product line. But then [the growth] starts to plateau. And at that point, they start reading my books and say, “Okay, now we got to lean into better customers, how do we start?” And my point is, you should have started from day one.
Build the infrastructure. Build the mindset of the [data-centric] culture, to at least look at that data, play around with it, start taking baby steps with it, so that when it’s time to flip that switch, you’re in a better position to do so.
Cook: Talk a little about why the data model that you guys developed at Theta with Dan McCarthy is so great for private equity.
Fader:
It’s really amazing. These models were originally developed for mundane marketing tasks–to answer questions such as “How many customers do we have? How long are they going to stay?” And I always wondered, I always dreamt about: “Wouldn’t it be great to get investors to pay attention to these things?”
But, I expected some serious limits, that if we don’t have access to the full transaction-log data, or the whole CRM system around it, that the models aren’t going to work as well.
Now, in a private equity setting, often we do have the full data. So I’m not entirely surprised that they work as well as they do. But not always. And there are some situations where either the company doesn’t have its act together and doesn’t have all the rich data, or at the stage where they bring us in, they say, “We’re gonna give you these rolled-up metrics.”
I figured there’d be a big loss of fidelity and accuracy as we go from granular data to aggregated data. And this is where Dan [McCarthy] changed the world. When we first connected, he was my Ph.D. student, and while I’m oversimplifying here, I basically said: “Dan, your task is to figure out how to run our Buy-’Til-You-Die models on this very, very limited data. Show us which metrics we’d want to use and how many of them we need.” [I wanted him to show me] how bad models would be by sacrificing granular data.
And he did that superbly well, and he found out that boy, was I wrong.That’s the loss in fidelity and predictive value as you go from transaction-log data to the right kinds of metrics and the right math is very, very little–to the point where we’re sometimes saying, “You know what? Who needs that transaction-log data, just give me these metrics.”
We’d rather have the transaction logs because you want to slice and dice. But there’s very little loss and infidelity. And it’s great that we can sometimes run the models with an investor on the aggregated data. And if they say, “We like this,” then we can share all the data. And it’s just nice to know that we’re not going to get answers that conflict. It’s nice to know that we’ll just get slightly more accurate, slightly more precise versions of what we already got.
Cook: Was Dan’s big breakthrough in the mathematics behind that? His ability to transpose “the gravel to the boulders”?
Yes, the main part, at least in my eyes, would be the math. Investors are going to take the math for granted, it would be the choice of the metrics.
Let’s be fair, the math is fun. But the original driving question to Dan was, how many metrics do we need? And which ones? Because obviously, if we have weak ones, we’re going to need more of them.
I always figured we would need at least two metrics for each of the behaviors to fully calibrate everything. And turns out, we can get away with a lot less than that. It’s kind of obvious in hindsight, but a lot of these metrics will reflect multiple behaviors. For instance, if we look at the number of active customers, that’s indirectly reflecting both repeat-purchase propensities, as well as retention propensities, we can get some double dipping going on. So, we can get away with a surprisingly small set if they’re chosen carefully. We do the math to demonstrate that. And then, we’re off to the races. Now, it’s just a matter to get more companies to disclose these things. That’s a big challenge.
Cook:
Let’s walk through a quick example: A shoe company with repeat-purchase behavior, where customers buy new shoes once a year. How should the company look at a Buy-’Til-You-Die framework? What would be some of the basic things you would need to track?
Fader:
It’s actually shockingly simple. We did this research and found we need these metrics: How many active customers did we have in a given period of time, how many customers made a purchase–did something of economic value, and what’s the average number of orders they placed?
Over the right time frame, if you can give me those two metrics, with enough history on them, that’s perfect to replicate our “Buy-’Til-You-Die” model.
That gives us models No. 2 and 3: retention and repeat purchase. We still need a metric to tell us how many customers we acquired, or in other words how many new customers we have. And we still need a metric to find the average spend among them. Basically, you can give me one metric for acquisition, two for repeat buying and retention, one for spend. Ideally, it would be nice if you can give me more, but if you give me those four [metrics] with a long enough history, measured over the right time frame–we’re good. It’s amazing to see how well it works.
Cook: For the private-equity investor, this data isn’t super sophisticated in one sense, but when you “mix the batter together, you have quite the cookie.
Fader:
You’re right. And that’s why on one hand, it is joyous to see how well these models can work and the implications that they have. And [on the other hand] frustrating that companies that actually do this are still pretty rare.
I’m an eternal optimist. And I really believe that as the word spreads, thanks to some of the work that we’ve done together, and other firms that are saying the right things, that there’ll be a generational change.
And when today’s eCommerce entrepreneurs become the CEOs of big, giant retail firms in 15 to 20 years from now, they will carry these practices with them. And, we will look back in 2040 and we’ll take for granted these kinds of analysis and we’ll say “I can’t believe they weren’t doing this even in 2020.” So, maybe we will feel foolish in hindsight. But let’s get there first.
Cook: You advise doctoral students, research, teach … what are you having the most fun with right now? Also, real quick for the record, how long have you been working around customer data?
Fader:
I’ve been a professor at Wharton doing this kind of thing for 36 years and a bunch of years before that as a Ph.D. student at MIT. And so, going back to the early 80s with my fairy godmother, Lee McAllister, who talked me into this whole thing, she painted this picture back around 1982 of rebuilding the “electron microscope” of the customer. “Pretty soon we’ll be able to see everything that each person is doing,” she said.
And back then, I was like, “Oh, yeah, right. Sure, whatever you say.” But today, we have it.
And so yeah, it took a generation to build an “electron microscope” [of the customer]. And now it’s, it’s gonna be a generation to realize how to use it most effectively.
So to answer your question, first of all, I love it all. And this sounds really corny. But every day I wake up and say, “What do we get to do today?”
- Do we get to teach this stuff to really smart students?
- Do we get to push the frontiers through research with people like Dan McCarthy?
- Do we get to go out there and change the way companies do things?
On a day-to-day basis, if I had to pick one, teaching brings great joy.
But, when we work with a private equity firm, and we can either talk them out of doing a deal where where the unit economics of the customers aren’t nearly as favorable as they might have thought, or if we can say, “You know what, you should be spending another $50 million on that company. They’re worth much more than you think.” And to see them write that bigger check and understand, that’s all because of us. You see the impact. And you just can’t get blasé about that. To see that a lowly marketing professor can have such an impact on big-stakes investment decisions, it feels really good.
Cook: And in those rooms, and having been a professor for so long, you’re used to having cantankerous students, maybe cantankerous colleagues. What are some of the cantankerous conversations in the private equity space?
Fader:
It’s going to be these old-school people, basically saying, “I know more about a company by watching a video with its CEO, than anything you can get out of customer data.”
There’ll be a bunch of just outright disbelief about even doing “this metric stuff” at all. There’ll be some disbelief, even if they’re willing to use [the models], to make it a major part of the investment decision. At best, it’s gonna be something they use for operating the company after buying it, or maybe at the margin, to change the last decimal place on the check they write.
To say that this kind of data, and these kinds of models can drive those decisions, yeah, that will generate some skepticism. And I understand that. But then again, I go back to them, and just lay out this idea: Every dollar you get is coming from customers. So, don’t you think that all the multiple stuff that you’re doing is implicitly just a reflection of what you think about the customer base? How sticky it is, and how many more you’re going to acquire. So, why don’t we take that intuition and just quantify it [with data.]
Cook: The proof is in the pudding because Theta has existing relationships with PE firms that have come back to you not once, not twice, we’re talking hundreds of times through due diligence and also on the value creation side. So, once you create value for them, they tend to come back to the well?
Fader:
It’s so nice to see that. For that right tail of PE firms that are using this all the time, what’s wonderful about it isn’t only the knowledge that they’re using these models to make big [business] decisions, but the process of working with them, it’s just so nice.
They know exactly what kind of data that we want. They know how to take the outputs that we give them and then massage them into their investment decisions. And it just shows that this can be done on a day-to-day basis, on every single deal. It doesn’t have to be that oddball, one-shot thing. Now it’s just a matter of getting everybody else to catch up to them.
Cook: If I’m a leader in a private-equity firm. How do I get started with Theta?
Fader:
It’s very common that someone will see me present or read some of my work and reach out. Often, they will give us one deal. Sometimes it’s a deal from the past, they’ll share the data to see if we can validate what actually happened. It’s a proof-of-concept [deal]. Often from there we will be given a second deal. Where things really kick in, is when we can look over the entire portfolio. I just love to do that because we look at every one of the businesses–whether it’s a SaaS enterprise company or a digitally-native women’s cosmetic company–and begin to make apples-to-apples comparisons among them because we are evaluating [the entire portfolio] on the four same measures: acquisition, retention, repeat purchase and spend. They can really start to understand, putting the products or services aside, what makes these companies different and then [be able to] make more informed decisions about what other types of companies to bring into the portfolio based on customer mix. It’s when we can do that type of meta analysis across a bunch of different companies at one time, and get multi-company learning. That’s really great.
Cook: I just wanted to say, I’ve been an adjunct professor since 2007 and I’ve been very fortunate to create and teach classes around eCommerce. The work that you and Dan have done in open sourcing, not just the mathematical models, but the code base and some of the APIs that have been built on top of that, has been huge. If you don’t hear this enough … thank you. To be able to put this in front of a student and give them a real-world data set, you can just see the lights go on. It’s really fun.
Fader:
I have great fun with this myself. Why should I be the only one to have all the fun? As a professor, I like to profess. And, when I see folks finding use from this stuff and then sometimes improving upon it or getting others to appreciate it who I normally wouldn’t have contact with, I know I’m doing the right thing. The fact that I get to do it, makes me feel so lucky.
We are all in this together, and the world of eCommerce is becoming a better place as a result.
Interested in attending the 2023 conference? Request an invite here.