Chapter 5 of my upcoming book, Freemium Economics, is dedicated entirely to lifetime customer value. In researching this chapter, the work of two academics – Peter Fader of the Wharton School of Business at the University of Pennsylvania and Bruce Hardie of the London Business School – surfaced again and again. Fader and Hardie have contributed an enormous body of work dedicated to lifetime customer value to the marketing discipline.
Dr. Peter Fader
And their contributions have not been purely theoretical; in fact, many of their papers focus on the practical aspects of calculating lifetime customer value, and some are even accompanied by Excel spreadsheets.
I spoke with Fader about the purpose and importance of implementing LTV into a firm’s acquisition process and general trends he’s witnessed within marketing as increasingly larger volumes of data about user behavior become available as model inputs.
Eric Seufert: User acquisition on mobile is conducted through advertising channels which are, generally speaking, fairly opaque: on a per-user basis, acquisitions are difficult to attribute to the channel from which they were generated reasonably accurately, in a short amount of time. This adds a layer of complexity to the calculation process that requires infrastructure beyond the scope of many small firms to overcome — tracking infrastructure, matching infrastructure, and reporting infrastructure. How do you think small firms should approach this? How does a small firm weigh the costs of LTV accuracy against its benefits?
Peter Fader: Too many firms (large and small) focus on “cost per acquisition” as the way to gauge and guide their acquisition efforts – what a mistake! So you’re right that LTV take some effort (and cost) to calculate on a regular basis, but the upside of managing the acquisition process on the basis of value instead of cost can be huge in many cases.
ES: At a high level, do you think a “theoretical” LTV provides value? That is, a value that is derived from rough approximations of user behavior
PF: Yes…it is useful to determine a ceiling to guide the value for acquisition spending (and retention spending for existing customers). So even if it’s just a theoretical norm, the impact on a manager’s mindset can be quite substantial.
ES: For small companies building digital freemium products, what’s an appropriate discount rate to use when discounting LTV, given a very long potential timeline for product use but a high level of risk in developing such products (successful apps can make millions per day, but app revenue is highly stratified, with a huge percentage of total app store revenue being generated by a small percentage of the total ecosystem)?
PF: Excellent question – I have no idea. That’s one of the nasty, unsatisfying aspects of CLV-related work to date: we have very little good ideas about how to choose the discount rate, and this decision can have a sizeable impact on the resulting calculations (and allocation decisions).
ES: Given a now almost incomprehensible volume of data, how has the tenor of marketing changed over the last decade, especially as it concerns LTV estimation?
PF: Ironically, the basic elements of the LTV calculation haven’t changed at all, but many firms now have an appetite (or a demand from investors/senior management) to do so. Companies have raised their “quantitative literacy” overall, and greater use of LTV has been an excellent outcome of this change.
ES: Your work with Bruce Hardie has often focused on practical implementations of LTV calculation. Can this be done in a “big data” environment? How valuable is Excel as a tool for calculating LTV with some minimum level of precision?
PF: Excel is a nice starting point, and sometimes we’ll continue to use it as a way to check our work (and/or convey it to low-tech folks). But it’s rarely acceptable to stick with Excel for commercial-scale applications. This is why we have created/shared a lot of MATLAB code for our models, as well as a new open-source R Library (http://cran.r-project.org/web/packages/BTYD/index.html) for the key models.
ES: What changes have you observed in the backgrounds of MBA and undergrad students at Wharton pursuing marketing concentrations? Has marketing become fundamentally more focused on computational quantitative methods?
PF: Several changes:
1. Many marketing majors are now seeing that marketing can be a science (as well as an art), so they are willing to take more quantitative courses in this field.
2. Other students who are quantitatively oriented are becoming more interested in marketing than they would have even imagined.
3. We’re seeing an interesting split among students who want to actually do the analytics, versus those who want to build strategies around analytics. But it’s important for both groups to learn this stuff very well, so my course enrollments are higher than ever.
ES: How do you think companies should think about user acquisition expenses with respect to sunk costs (like pre-revenue product development, etc.)?
PF: I’m no accountant, but I don’t want to count any expenses that aren’t directly attributable to the customer(s) being acquired. Anything that’s overhead (including general expenses associated with an acquisition campaign) shouldn’t be counted towards LTV. But I recognize that it’s messy/difficult to allocate expenses on a very granular basis, so I’m willing to be somewhat flexible here…
ES: What’s one tip you would give to a new product manager in approaching LTV from a conceptual standpoint?
PF: It’s important that the manager does NOT treat all sales equally. It is essential to distinguish sales to new customers versus “first repeats,” “second repeats,” and so on. This kind of “depth of repeat” decomposition isn’t difficult but it’s critical if one wants to accurately forecast new product sales and draw meaningful diagnostic insights about the underlying customer behavior.
Peter Fader and Bruce Hardie are holding a two-day workshop in San Francisco in mid-February dedicated solely to lifetime customer value as part of Wharton’s executive education program. More information can be found here.
Recent