This presentation was made at a data science meetup hosted at N3TWORK's offices during GDC 2018. The presentation is a follow-up to a post with the same title published on Mobile Dev Memo in early 2017.
4. Review: How do we build an LTV estimate?
- Get clarity on model;
- Get clarity on early stage signal
5. Review: How do we build an LTV estimate?
- If we’re confident in the model, we
can use early-stage monetization
data to project out to some goal LTV
metric (“Day X LTV”);
- Eg. if we want to recoup our money
in 90 days, we should project a Day
90 LTV; if we are confident in our
model, perhaps we can estimate a
Day 90 LTV with just 7 days of data.
6. Review: How do we build an LTV estimate?
- 7 days isn’t a long time. But how much data do we need to even get a comfortable, robust estimate of
7-day LTV?
7. The Approach
- Simulate the collection of monetization data to understand how many users it takes to get a good
estimate on Day 7 LTV;
8. The Funnel
- As we acquire users into our app,
we accumulate data over time;
- The amount of data we have at Day
X is a function of the input new
users + retention + days that new
users have been acquired;
- With 100% user retention, with 100
DNU running for 4 days, we have
400 data points for Day 1, 300 data
points for Day 2, 200 data points for
Day 3, and 100 data points for Day
4 (only one cohort has reached Day
4 in our app);
- Adding in a typical retention curve
makes this degradation in data size
even more pronounced.
9. The Approach
- Simulate the collection of monetization data to understand how many users it takes to get a good
estimate on Day 7 LTV;
- Build some reasonable assumptions of how our app / marketing performs:
15. The Approach
- Calculate confidence intervals for each Day X LTV - 5 Days of cohorts,
500 users per day
- Day 4 LTV CI 95%
is $0.38 -> $0.62
- $0.24 difference,
nearly 50%!
- And that’s just Day
4!
16. The Approach
- What about 90 days of traffic buying? - Day 5 has a tight CI
- But Day 90 spread
is large