6
$\begingroup$

I am trying to fit the popular Pareto/NBD or Beta/Geometric models for non-contractual, continuous customer data. On top of that I then fit the Gamma/Gamma model for monetary value (using the very helpful lifetimes package). There are a few questions on CV related to this topic, but none that I can find that answer any of my questions below.

Questions:

  1. As asked here (but as of yet unanswered) for customers who made only 1 purchase, what should we project for their future LTV? Currently it would be just the average LTV of all customers, but this seems very naive. Does anyone know of a more scientific approach than simply using some heuristic (e.g. weighted average of entire customer base and individual's first purchase)?

  2. What about completely new customers? Any approach other than just using the population average?

  3. I notice that for newer customers with less historical data, either model provides very small average future LTVs. My sense is that it's because new customers don't have much data to fit a nice distribution to. Does anyone know if the best practice is to (1) ignore new customers until they've made enough purchases to fit a model to, or (2) combine old customer data with new customer data as a heuristic? or (3) some other method? My CLTV estimates for newer cohorts are unreasonably low.

Improve this question
$\endgroup$
3
  • $\begingroup$ did you ever solve this problem. Especially, in regards to new customers? $\endgroup$
    – RDizzl3
    Commented Oct 7, 2019 at 17:58
  • $\begingroup$ Nope. I’ve since moved on but curious... $\endgroup$
    – ilanman
    Commented Oct 7, 2019 at 22:56
  • $\begingroup$ @ilanman I'm facing similar situation. Did you find the solution for this? Also, Is there any relationship between the number of periods you need to have the customers' data and the number of states he has been transited to (in those periods)? $\endgroup$
    – Artiga
    Commented May 15, 2020 at 9:50

1 Answer 1

Reset to default
0
$\begingroup$

For a new cohort without history, you need to make an assumption that the new cohort behaves similarly with an existing cohort and use that distribution as a starting point (the prior distributions). Then as you receive more history for the new cohort, you can update the prior distributions.

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.