How to create good BI for UA activities? How we did it at Pixel Federation? How we structure our systems and our teams? What we can say based on our experience? Presentation run by Michal Grno at 8th edition of GameCamp (www.gamecamp.io).
11. Data Discovery Workbench / Lab
Raw Data
Staging
Layer
Performance
Layer Custom Apps
Dashboards
Our Data Ecosystem
DWH API
API
API
www.pixelfederation.com
Data behind UA
Our BI Stack
13. Processing
Data Discovery Workbench / Lab
Raw Data
Staging
Layer
Performance
Layer
~4,500
50
automated jobs daily
GBs of compressed data stored
9
analysts digging into data
www.pixelfederation.com
Data behind UA
Our BI Stack
14. Reporting
20,000,000
101
events sent to 3rd parties
reports refreshed
Performance
Layer Custom Apps
Dashboards
DWH API
API
www.pixelfederation.com
Data behind UA
Our BI Stack
16. www.pixelfederation.com
Data behind UA
Our BI Team
Centralized
DATA
team
DA
TS
SP
EA
● Benefits
○ Knowledge sharing
○ Replaceability
● Drawbacks
○ Analytics only as a support
○ Not being part of product team
○ Communication
18. www.pixelfederation.com
Data behind UA
Our BI Team
CDO
+
DWH
DA
SP MKT
TS2AC
*
*
*
*
*
● Benefits
○ Agility
○ Analytics as part of a product
○ Working on things that matter
○ Communication with PMs and GDs
● Drawbacks
○ Some tasks take too long
○ Doing things that have been
already done
22. Predicting the LTV
● Lifetime value of a player
○ “Lifetime” = predefined time period (3m, 6m, 1y, …)
○ Usually period to regain the costs
This is where we are now
● Methods
○ Historical revenue
○ Retention method (ARPDAU*Lifetime)
○ Continuous ARPU curve
○ Normalized ARPU curve
○ More advanced methods
■ Pareto/NBD
■ BG/NBD
■ Gamma-Gamma
www.pixelfederation.com
Data behind UA
Doing UA the smart way
24. Predicting the LTV
● ARPU curves
○ calculated from historical data
○ for project + platform
○ function of days_in_game
○ normalization means that norm_ARPU(364) = 1
● How does the prediction works?
○ pred_ARPU(x) = actual_ARPU(x) / norm_ARPU(x)
www.pixelfederation.com
Data behind UA
Doing UA the smart way
25. Predicting the LTV
AFK Cats Android 21%
Diggy
Android 9%
Canvas 4%
Seaport
Android 19%
Canvas 12%
TrainStation Android 17%
D7 ROAS Benchmark
www.pixelfederation.com
Data behind UA
Doing UA the smart way
26. Predicting the LTV
AFK Cats Android 31%
Diggy
Android 23%
Canvas 13%
Seaport
Android 31%
Canvas 25%
TrainStation Android 40%
D30 ROAS Benchmark
www.pixelfederation.com
Data behind UA
Doing UA the smart way
31. Predicting the LTV
● Easy to calculate
○ only payment data needed
● Easy to implement
○ not so big data
● Easy to interpret
○ everyone understands $$$
● It works!
● Unstable over time
○ game design changes
○ optimization/network
● Inaccurate with small samples
○ only payment data used
● Predictions made on cohort level
○ not on player level
● Easily “fooled” by whales
● Unreliable with early predictions
Pros Cons
www.pixelfederation.com
Data behind UA
Doing UA the smart way
44. To summarize
● LTV > CAC is not everything!
● Give UA managers the tools they need
to make the right decisions
● Measuring LTV is kind of difficult
● Garbage in = garbage out
www.pixelfederation.com
Data behind UA
45. Time for Q&A
Michal P. Grno
mgrno@pixelfederation.com
https://www.linkedin.com/in/michalprokopgrno/