How to start working with LTV measurement in mobile gaming? How to move to advanced strategies?

Proprietary + Confidential
How to start working with
LTV measurement in
mobile gaming. How to go
to advanced strategies?
Mariusz Gąsiewski,
gasiewski@google.com
Warsaw, 17.9.2018

Source: Lorem ipsum dolor sit amet, consectetur adipiscing elit. Duis non erat sem
Not all users are the same!

Payer fraction growth in mobile F2P games
Source, DeltaDNA, September 2018

It is not enough

Example approaches in the field
Historical Average or Cohort Simple Predictive Models Advanced Predictive Models
Example:
Probability Models
combined with machine learning
Pro:
Very high accuracy, can become significant
competitive advantage for a company in
monetisation once done right.
Con:
Requires data science team and significant
amount of data.
High computational power needs
Historical Average or Cohort Simple Predictive Models
Example:
Feature selection using Random Forest
Pro:
Strong outcome, possible to do with pretty
limited ressources and for middle size
games. High actionability.
Con:
Less accurate than advanced statistical
methods. Data scientist experience strongly
recommended.
Con:
Doesn’t account for differences between
customers. Combined with strong risk once
not done carefully.
Pro:
Easy to compute with average revenue per
user. Good as the start even for small
studio. Actionable once done at the country
level.
Example:
Simple models based on spreadsheets and
formulas

It really matters...
Version 1 Version 2

K-factor analysis example

Example LTV calculator
Example LTV calculator
https://goo.gl/A4fmKc

Other simple approaches
Source: http://www.gamedonia.com/blog/5-ways-to-calculate-lifetime-value-for-free-to-play-games

Is first payment good prediction of LTV?

Feature selection

Can be used for effective UA

What makes feature selection?
Source, http://blog.citizennet.com/blog/2012/11/10/random-forests-ensembles-and-performance-metrics

Red dots represent users that made in-app purchase
Blue dots represent users that didn’t make in-app purchase
Imaginary Traditional View

Red dots represent users that made in-app purchase
Blue dots represent users that didn’t make in-app purchase
Imaginary Traditional View Actual Case

...to more advanced approaches.ExtraRandomForest
A supervised learning algorithm that builds multiple decision trees and merges them together to get a
more accurate and stable prediction

Source: https://www.machinelearningplus.com/machine-learning/feature-selection/

Source: https://hub.packtpub.com/4-ways-implement-feature-selection-python-machine-learning/

Example

MoonTon increases high
quality user volume by
leveraging Feature
Selection Methodology
Background
MoonTon was founded in 2014 in
Shanghai, focusing on developing
high quality mobile games. It has
become well-known in SEA
markets for its hero game’s
achievements both in install aspect
and revenue aspect.
Goal
MoonTon values the power of
machine learning and tries to
explore a way to balance the
installers’ quantity and quality from
game ecosystem perspective.
Approach
• Utilize Feature Selection
algorithm model to find out the
common in-app actions that high
value players act.
• Start UAC Actions campaigns
which optimize towards feature
selection events.
“We are impressed by the effect of
machine learning technology and we will
surely invest more and dig deeper to
unleash its potential next.”
--Cai Xuwei
Marketing Head
MoonTon
+18%
retention rate
-52%
cost/retention
+423%
ROI

Which customer is more valuable to your business?
CPI
Value generated in Month 1
A B
$9 $19
$10 $10
Value generated in Month 2 $0 $10
ROI end of Month 1 11% -47%
Value generated in Month 3 $0 $10
ROI end of Month 3 11% 58%

BTYD: predictive model based on 3 inputs (RFM)
Unique
identifier
Transaction
value
Date of
transaction

BTYD: example of data output (prediction)
Customer
ID
P(Alive)
Predicted
Transactions
Predicted
AOV
Predicted
Value
1 39.0% 0.2 $68.34 $16.36
2 99.7% 31.3 $45.53 $1,424.94
3 99.1% 7.8 $122.52 $958.64
4 97.9% 7.6 $62.29 $475.35
5 81.3% 2.2 $85.51 $183.55
6 99.9% 9.0 $126.15 $1,138.03
7 97.5% 13.9 $95.65 $1,328.45
8 67.0% 2.3 $167.61 $379.78
9 28.9% 4.8 $230.58 $1,017.02
10 52.8% 1.6 $126.83 $202.36

BTYD: Example of an RFM model at work

BTYD: Example of an RFM model at work
Looking ahead, academic research shows that at Week 78 Customer A is actually
more valuable than Customer B.
Predictive Models help us understand and value this future behavior

Scalable Machine Learning System Successfully Onboard
Cloud & Big Query
Data Architecture / Processing
Data Analytics
TensorFlow Marchine Learning
Predictive Payer + pLTV
Business Intelligence
Reporting via Data Studio

Clash of Kings Boosts
CVR up by 7X with
TensorFlow & UAC
Approach
• Build a TensorFlow Deep Learning Model on
Google Cloud
• Predict gamers’ likelihood to pay based on
hundreds of behavioral signals with an automated
Machine Learning System
• Leverage Google UAC to acquire “will pay gamers“
based on predictive signals
“TensorFlow & Machine Learning are awesome tools, It
helped us explore a smarter way of acquiring high value
customers.”
-- Bin Gao
Senior UA Lead
Elex Clash of Kings
2.6X
ROAS
Over 7X
Paid User/Installer
63% decrease
of cost per paid
user

Worth to look at
Source: https://www.youtube.com/watch?v=-ZEXx6PNRIc

Worth reading
Source: https://medium.com/googleplaydev/predicting-your-games-monetization-future-ce176169b056

Questions?

How to start working with LTV measurement in mobile gaming? How to move to advanced strategies?

More Related Content

What's hot

What's hot (20)

Similar to How to start working with LTV measurement in mobile gaming? How to move to advanced strategies?

Similar to How to start working with LTV measurement in mobile gaming? How to move to advanced strategies? (20)

More from GameCamp

More from GameCamp (20)

How to start working with LTV measurement in mobile gaming? How to move to advanced strategies?