In this presentation, AppAgent's mobile marketing experts demonstrate and explain the process behind creating ROAS prediction models for Ad-Monetized Games.
They cover what data is needed, how to turn this data into insights that are useful for predictions, how to practically create calculations and charts in Google Sheets, and how to use the model to understand the game’s monetization potential.
10. WHERE WE’LL START | FUNDAMENTALS What the...
Warm-up and terminology
● What ROAS components can we actually predict?
● Why do we always want to look at curves?
11. CONCEPT 1/3 | CALCULATING LTV
Getting the basic curve shape from Retention and ARPDAU data.
● How to calculate LTV and create curves using “Retention x ARPDAU” Methodology
● Visual example from the ground up
● Calculation using linear interpolation
12. Predicting retention points if data is not available.
● "Predictions" without having the data
● Example: We have all dates until D120 but lack any info on D365
● We'll discuss the approach and limitations
CONCEPT 2/3 | PREDICTING UNKNOWN RETENTION POINTS
D360
???
13. CONCEPT 3/3 | RETENTION PROFILES AND HOW TO USE THEM
Using the full curve to perform LTV / ROAS predictions.
● Retention profiles allow you to practically perform predictions on your new cohorts (so you can quickly
evaluate your paid UA) or "predict" and analyze performance of newly opened channels.
14. Practical example.
● How to do an actual analysis to identify retention profiles
● Using retention profiles for prediction and new channel analysis
FINAL OUTCOME | EXAMPLE DATA ANALYSIS AND PREDICTIONS
16. Martin Jelinek
● 7+ years in game development
● 6+ in mobile marketing
● Head of Marketing
● Data Enthusiast
YOUR HOSTS FOR TODAY
Roberto Sbrolla
● 15+ years in digital marketing
● 7+ in mobile marketing
● Growth Consultant
● Hobby: learning Unity and game development
17. WHO IS APPAGENT?
Founded in 2016, AppAgent is a strategic & creative marketing
partner for top game and app publishers.
The team consists of 30 people from 14 nationalities and
various specializations.
AppAgent was awarded App Marketing Agency of the Year in
2018 and 2020 while it has been shortlisted ever since 2017.
18. WARMUP & TERMINOLOGY
● Concept 1 - Calculating LTV from Retention and ARPDAU
● Concept 2 - Predicting Unknown Retention Points
● Concept 3 - Retention Profiles & How to Use Them
● The Final Boss: Example Analysis & Predictions
20. No predictions and calculations needed!
TRIVIAL EXAMPLE 1: PREMIUM APP
ROAS =
Purchase Price (one-time)
CPA (avg cost per payer)
________________
21. But most games (apps) don’t monetize with one initial purchase.
TRIVIAL EXAMPLE 2: AD MONETIZED GAME
ROAS =
CPI (avg cost per install)
________________
LTV (accumulated over time)
22. It should not look like a snapshot (eg. D90 LTV) but rather like this:
TRIVIAL EXAMPLE 2: AD MONETIZED GAME
ROAS =
LTV
CPI
___
D(x)
D(x)
Thecurveiscool!
Belikethecurve!
23. We are Live!
You still did not
persuade me about the
curves, young men!
24. IMAGINE 3 GAMES WITH THESE D30 LTV’S:
1 2 3
D30 LTV
$1
D30 LTV
$0.6
D30 LTV
$0.8
If you have to pick one to invest into - which one would you pick?
25. 1 2 3
D30 LTV
$1
D30 LTV
$0.6
D30 LTV
$0.8
Example 1’s LTV
is like…this big!
IMAGINE 3 CAMPAIGNS WITH THESE D30 LTV’S:
26. LOOK AT THE CURVES TO SEE THE WHOLE STORY.
1 2 3
Daily snapshots can be deceiving - aim for curves, not snapshots!
28. CONCEPT 1:
CREATING LTV CURVES
● Concept 2 - Predicting Unknown Retention Points
● Concept 3 - Retention Profiles & How to Use Them
● The Final Boss: Example Analysis & Predictions
29. This is the context of whole ROAS calculation.
LTV CURVE THIS again?!
30. We’ll need 2 core ingredients.
LTV CURVE
LTV Retention x ARPDAU
(avgactivedays) (avgads/day*eCPM)
D(x)
Howmanyadswillheseeperday?
Whatistherevenueperad?
Howmanytimes(days)will
theuserbeactivefor?
31. WINDOW OF FUNDAMENTAL LOGIC
Following the logic:
“If the USER plays the game on a given day,
he will spend some avg time and see some avg number of ads; Each will generate REVENUE (cents).
If we understand how many ads per day the user consumes,
and HOW MANY DAYS will he be active (in a given timeframe),
we'll get the LTV.”
32. We’ll need 2 core ingredients.
LTV CURVE
LTV Retention x ARPDAU
(avgactivedays) (avgads/day*eCPM)
ARPDAU-fromreports
(canbebrokendownbutnotsomethingwe'll
dotoday-cannotbepredicted)
D(x)
Lifetimeactivedays-calculatefrom
retention(cumulativeoverdays)
34. Cumulative Lifetime days
How many days was an average user active in X days after
install?
Retention chart
What % of users from our cohort will come back on D(x)?
LTV FROM RETENTION & ARPDAU - LET’S OBSERVE THE CURVE FORMATION
35. Cumulative Lifetime days on D0 = 1.
On average, each newly acquired user has been active for
exactly 1 day.
D0 = 100%
Everyone who launched the app is considered “active”.
DAY 0 - WE STARTED THE CAMPAIGNS!
36. Cumulative Lifetime days on D1 = 1.6.
We add D0 (100%) and D1 (60%). On average, each person
has been active for 1,6 days so far! But this will increase
tomorrow..
D1 = 60%
Not all users came back on D1. This was expected!
DAY 1
37. Cumulative Lifetime days on D2 = 2.
We add D0 (100%), D1 (60%), and D2 (40%).
On average, each person has been active for 2 days so far!
But this will increase tomorrow..
D2 = 40%
We get another retention datapoint..
DAY 2
38. Cumulative Lifetime days on D7 = 3.13.
We summed all the retention numbers for each day so far.
This gives us an average of 3.13 active days on D7!
D7 = 19%
But more importantly, we now have all the data points!
DAY 7 (LET’S SKIP AHEAD!)
44. We need retention for each D(x) - but we often get just the key retention points:
LET’S DO THIS BEYOND D7 NOW!
45. LET’S DO THIS BEYOND D7 NOW!
It looks like this in the chart. To calculate the lifetime days, we need each and every D(x) value!
46. LET’S DO THIS BEYOND D7 NOW!
It looks like this in the chart. To calculate the lifetime days, we need each and every D(x) value!
Interpolate!
You have to
In-Ter-Po-Late!!!
47. THERE ARE MULTIPLE WAYS TO DO THIS. OPTION 1
Option 1:
Use the Calculator that we provide :)
49. SO LET’S DEMONSTRATE HOW TO SIMPLY INTERPOLATE
Let’s first zoom into the first 30 days or so!
50. INTERPOLATION BETWEEN D1 & D7
We see 3 data points: D1, D7, and D30.
Let’s create the rest of the data points between D1 and D7 first.
51. Retention decreased by (48-26)=22% from D1 to D7 (in 6 days).
On average, the retention decreased by (22%/6)=3,67% for each of these days!
INTERPOLATION BETWEEN D1 & D7
53. Retention decreased by (26-16)=10% from D7 to D30 (in 23 days).
On average, the retention decreased by (10%/23) = 0.43% for each of these days!
INTERPOLATION BETWEEN D7 & D30 - USE THE SAME LOGIC
54. SAME LOGIC CAN BE APPLIED TO ANY LINE SEGMENT
Let’s demonstrate how to do the interpolation between any 2 points!
55. NOW WE HAVE A GOOD APPROXIMATION OF THE FIRST 120D RETENTION!
(we could now simply flip this to active days, LTV, and ROAS if we wanted to!)
57. NOW WE HAVE A GOOD APPROXIMATION OF THE 120D!
(we could now simply flip this to active days, LTV, and ROAS if we wanted to!)
But you know what they
say - life begins at 120!
58. CONCEPT 2:
PREDICTING UNKNOWN RETENTION POINTS
● Concept 3 - Retention Profiles & How to Use them
● The final Boss: Example analysis & Predictions
59. WHAT IF WE RUN OUT OF DATA?
What if we don’t have any historical data beyond D120? Can we somehow predict the additional data points?
D360
???
62. DIFFERENT APPROACHES.. BUT WE JUST DON’T HAVE THE DATA!
We can try fitting to a curve - but that’s just relying on maths and natural curves..
ExponentialTrendline
LogarithmicTrendline
63. MY RECOMMENDATION - APPLY LOGIC AND MODELLING
Let’s check a couple scenarios!
64. MY RECOMMENDATION - APPLY LOGIC AND MODELLING
Let’s take a look at couple
scenarios!
I hope
they’ve got at
least 3!
65. WHAT CAN HAPPEN BETWEEN D120 AND D360?
Scenario 1
Retention does not decrease.
Almost impossible.
66. WHAT CAN HAPPEN BETWEEN D120 AND D360?
Scenario 1 Scenario 2
Retention does not decrease.
Almost impossible.
Drop from D90-D120 will continue.
Very unlikely.
67. Loose the same % of users every day.
Quite unlikely.
WHAT CAN HAPPEN BETWEEN D120 AND D360?
Scenario 1 Scenario 3
Scenario 2
Retention does not decrease.
Almost impossible.
Drop from D90-D120 will continue.
Very unlikely.
68. LTV?
Loose the same % of users every day.
Quite unlikely.
HOW WOULD THESE RETENTIONS TRANSLATE TO THE ACTUAL LTV?
Scenario 1 Scenario 3
Scenario 2
Retention does not decrease.
Almost impossible.
Drop from D90-D120 will continue.
Very unlikely.
69. HOW WOULD THESE TRANSLATE TO THE LTV?
We already get some indication of what the difference in retentions can mean for the resulting LTV!
$5,1
$5.6
$7.2
$8.6
70. Which of these is more likely? Try thinking about your game:
- How much content do you have? Is it enough for long term?
- Are users socially engagement? (this can mean they’ll stay longer)
- How are users currently "engaged"?
- Industry benchmarks (Appsflyer have some?)
IT’S ACTUALLY QUITE HELPFUL IF YOU USE IT WELL!
76. RETENTION MULTIPLICATION PROFILES
Now that we have a good idea of the full curve, we can use it for predictions of new cohorts!*
Retention
what? Like a
repeated tension?
Note: We are simplifying here and expecting the retention profiles to be stable across all dimensions.
77. RETENTION PROFILE - CONCEPT
D7:D1 = 54%
Day 1 retention: 48%
Day 4 retention: 26%
If we do the math (26/48=54%), we see that 54% of
D1 active were retained to D7.
78. RETENTION PROFILE - CONCEPT
D30:D1 = 33%
Day 1 retention: 48%
Day 4 retention: 26%
Only 1 third of D1 users survive until day 30.
79. RETENTION PROFILE - CONCEPT
D90:D1 = 19%
Day 1 retention: 48%
Day 4 retention: 26%
Only 1 third of D1 users survive until day 30.
80. RETENTION MULTIPLICATION PROFILE CREATION
D120/D1=13%
Day 1 retention: 48%
Day 120 retention: 6.5%
If we do the math (6.5/48=54%), we see that 13% of
D1 active were retained to D7.
81. RETENTION MULTIPLICATION PROFILE CREATION
So if you have a table of these fractions
D1=(?)
D7=0.54*(D1)
D30=0.33*(D1)
D90=0.2*(D1)
…
Each time you measure D1, you can apply the
retention profile, “predict” all the retention
points, and take it from there!
83. THE DATA TO USE
All retention values.
Organic / Paid
division
Organic + 3 paid
channels
3 GEOS
5 Months
84. ✅Creating the D1 retention profile
✅ Analysis of the profiles (“how many do we need?”)
✅ Selecting the profiles
✅ Analysis of a new cohort, what-if analysis
✅ New channel analysis based on initial retention
NOW LET’S DEMONSTRATE
85. Call us if you want
to watch TV or
calculate LTV!
86. DO YOU NEED HELP WITH LTV PREDICTION
AND SCALING PAID UA? PING ME!
Email: nenad@appagent.com
Linkedin: /in/nenadstevanovic
Who is this
handsome fella
now?