A comprehensive summary about the lifetime value predictions that Martin, Head of Marketing at AppAgent, has learned by building analytics platforms for clients and consulting with the best of the best in the mobile industry.
LTV Predictions: How do real-life companies use them & what can you learn from it?
1. LTV Predictions
How do Real-Life Companies use Them &
What can you Learn From It?
Martin Jelinek
Head of Marketing
AppAgent
2. Introduction
Martin Jelinek
• University of Economics
• 2007 – 2016: Independent Game Developer
• 2017 – 2018: AppAgent
• 2019 – Head of Marketing at AppAgent
• Published / launched / consulted 40+ mobile games and apps
• Designed and built AppAgent‘s in-house Marketing Analytics
• Speaker at App Promotion Summit, GIC and other events
4. LTV Calculator
I could calculate LTV
and ROI myself.
Mastermind
My company uses a
custom LTV predictions
model.
LTV Predictor
My company uses some
form of LTV Predictions.
ROI Primate
I understand the concept of
LTV and ROI and could
explain it to this audience.
Audience Level Check
What is the level of YOUR understanding?
5. 20 minutes / 5 mins questions
3 Sections:
- Predictions 101
- 3 most common modes
- Insights & Learnings
Today‘s talk:
7. Basics 1/4 - What is a prediction?
• What‘s a prediction?
• Using past experience to estimate what‘s
going to happen.
• Using historical data to understand what
will happen with % certainty.
•LTV vs. ROI predictions
• ROI = LTV / CPA
8. Basics 2/4 – What are predictions for?
• User Acquisition:
• How much can we spend (and scale)?
• When and how to adjust bids, creatives,
targeting etc.
• Confidence when scaling
• Evaluation of new channels
• (Re)-Engagement:
• Re-engagement efforts focus
• Identification of at-risk users
• Special offers
• Ads vs. in-apps
Many benefits, two key areas:
9. Basics 3/4 – How do they work?
• Using past to estimate
future
• Explore past data to find „rules“ that can be
leveraged in calculations
• Build a MODEL – set of rules and calculations
on how to transfer input into output
• Collect inputs (retention and ARPDAU) to
• Point vs. Interval
• The LTV for (segment) is $2,5.
• Prediction: „There‘s a 95% chance that
the LTV will be between $1.5 and $3.“
10. Basics 4/4 – No data, no predictions.
• Characteristics of data
• Variability
• Level of spend, first purchase timing, behavior
differences between segments
• Amount of data
• What is the minimum amount of data
points for reliable predictions?
• Userbase vs. refined segments
• App versions, seasonality..
To predict, we need past observations (data).
12. 3 Most common LTV prediction models:
- Retention x ARPDAU
- „Ratio“ model
- „Behavioral“ model
Concepts of LTV Models
13. „Getting a couple of retention datapoints for a cohort to model the shape of the
retention curve.“
pLTV = pRetention x ARPDAU
Concept 1 – „Retention x ARPDAU“
14. „If a cohort spent 100 bucks during the first 7 days, we know from past experience
they will spend in average 4x more by the time they mature to day180.“
Revenue D180 / Revenue D7
Concept 2 – „Ratio“
15. User
behavior > > LTV
„If user xyz had 10 sessions during the first three days, at least three of them
in the morning, and visited a promotional package page two times, he has a
probability of becoming a payer of 35%. His pLTV is 35 USD.“
Concept 3 – „Behavioral“
Revenue D180 / Revenue D7
16. „Let‘s get a couple of retention datapoints to model to see how sticky is the
game and model their retention curve. The more days they stay, the more they
will pay.“
Retention + ARPDAU >>> pLTV
Subscription App
17. „Let‘s get a couple of retention datapoints to model to see how sticky is the
game and model their retention curve. The more days they stay, the more they
will pay.“
Retention + ARPDAU >>> pLTV
Ad-monetized app
18. „Let‘s get a couple of retention datapoints to model to see how sticky is the
game and model their retention curve. The more days they stay, the more they
will pay.“
Retention + ARPDAU >>> pLTV
Casual Game
19. „Let‘s get a couple of retention datapoints to model to see how sticky is the
game and model their retention curve. The more days they stay, the more they
will pay.“
Retention + ARPDAU >>> pLTV
Hardcore Game
20. „Let‘s get a couple of retention datapoints to model to see how sticky is the
game and model their retention curve. The more days they stay, the more they
will pay.“
Retention + ARPDAU >>> pLTV
Airline Ticket Reseller
21. „Let‘s get a couple of retention datapoints to model to see how sticky is the
game and model their retention curve. The more days they stay, the more they
will pay.“
Retention + ARPDAU >>> pLTV
E-Commerce
24. A model needs to be assigned depending on the app vertical (and type),
monetization type, user behavior, business model and other factors.
Which model is the best?
NONE.
25. In case there‘s a huge variance in the data, predicting LTV with a sensible level of
certainty could be almost impossible.
Mission Impossible
(SOMETIMES)
MISSION
IMPOSSIBLE
26. From what we‘ve seen and
heard, the „ratio“ model is the
most common.
Even for the „big players“ – still, this conclusion could be biased, but still – even
big companies often favour this. The question remains
So what are companies using in real-life?
27. Even if there‘s added value of higher precision, the increased hassle is not worth it
- Lots of engineering time to create / maintain
- Apps keeps changing
... And complex is hard to change
- UA team understands
... And btw, do you REALLY need a prediction model?
Simple vs. Complex
SIMPLE vs. COMPLEX
28. 3 main approaches – retention-based, ratio based, behavior-based
Ratio model most frequent – probably as it‘s the best bang for the buck
No one-size-fits all – each app is different and so is the optimal prediction model
Not always possible – for some apps, predicting can be close to impossible
Simplicity is favored – most companies we talked to prefer simplicity
Marketers are responsible – most often, marketers drive the whole process
Key Takeaways