Optimising user acquisition through LTV prediction
3 likes2,398 views
The document discusses optimizing user acquisition (UA) through lifetime value (LTV) predictions, highlighting strategies to improve UA early in campaign lifecycles. It emphasizes the importance of accurate, data-driven predictions based on game monetization and cohort analysis to enhance decision-making and campaign performance. Additionally, it outlines the use of machine learning and cloud technologies to streamline data processing and improve the efficiency of UA efforts.
Optimising user acquisition through LTV prediction
- 1. Optimising user acquisition through LTV prediction Róbert Magyar Warsaw 14 -15 October 2019
- 2. Warsaw | 14-15 October 2019 Róbert Magyar Data Science Lead robert.magyar@superscale.com
- 3. “We’re forming growth partnerships with world’s top developers to scale their games to maximum potential.”
- 4. Warsaw | 14-15 October 2019 ● 70+ World Class Experts In-House ○ (Self-)Publishing Infrastructure i. Business Intelligence ii. Analytics & Data Science ○ Monetization i. LiveOps Optimization ii. Game & Monetization Design ○ User Acquisition i. Creatives ii. UA Campaign Management iii. ASO ● Founded in 2016 ● Bratislava, London, Berlin, Helsinki & Prague offices First Light Games Our partners Games
- 5. Warsaw | 14-15 October 2019 UA Optimization Motivation
- 6. Warsaw | 14-15 October 2019 How can we improve User Acquisition? Huge topic - lots of angles ● Lookalike redesign ● Different UA channels ● New creatives ● Different spending strategy ● Passively through game related efforts e.g. Economy Balancing ... ● Optimization of UA investments through LTV prediction - focus of this talk
- 7. Warsaw | 14-15 October 2019 LTV predictions as an actionable insight for UA team Comes early in the campaign lifecycle Helps to identify opportunities Accuracy is stable over time Supports decision making Defines strategy of UA investment optimization
- 8. Warsaw | 14-15 October 2019 Waiting for data is not fun Usual approach for building LTV prediction model: ● wait several weeks or months to gather data about the campaigns ● create LTV model ● use the model to estimate UA performance .. but market changes over time, new competitors arise etc
- 9. Warsaw | 14-15 October 2019 The Question How can we improve UA efforts with LTV predictions ● early in campaigns lifecycle? ● without need of waiting several months to understand performance trend? => by understanding monetization of the game and building predictions from bottom-up How to actually do it? ● utilizing cloud machine learning
- 10. Warsaw | 14-15 October 2019 Some games have steep early monetization which creates an advantage in modeling. Game monetization defines data needed for predictions = Game monetization => defines how many days of data we need and helps to understand the payback of UA 50% of revenue in first couple of days = enough data to work with
- 11. Warsaw | 14-15 October 2019 Variation in campaign cohorts performance = accuracy loss => Accuracy of models is impacted by daily market changes, this can be seen on significant movements in campaign’s cohorts performances over time
- 12. Warsaw | 14-15 October 2019 Building accurate predictions Focusing on daily cohorts => cohort’s distribution of payers does not change over time which improves accuracy while building up the predictions: 1. Predictions based on cohort level not on campaign level 2. Aggregate predictions for each campaign Idea is to go hierarchically to the lowest level possible (cohort/ad/adset etc) while: - Keeping enough players in cohorts = Daily cohorts => logarithmic growth can be leveraged => no need to wait several months for data
- 13. Warsaw | 14-15 October 2019 Keeping enough players in cohorts Number of players needed is based on: - Conversion and number of payers Unusual growth of revenue in the cohort can be the clue of not having enough players to work with. Looks like step function - reason could be: 1. Game relying mainly on Liveops offers (not so good design) 2.Conversion for the cohort is not high enough
- 14. Warsaw | 14-15 October 2019 UA Optimization LTV prediction in Action
- 15. Warsaw | 14-15 October 2019 Information we are able to use: - Monetary : Conversion, ARPU, ARPPU, # of purchases, revenue, payers distribution, probability of 2nd purchase etc - Absolute value - Relative change/growth over first couple of days - per cohort or campaign and aggregated - Behavioral (Engagement metrics) : playtime, retention, first day drop-off, percentage of one-shots etc Leveraging all the data
- 16. Warsaw | 14-15 October 2019 Business optimization goal
- 17. SuperScale Analytics + UA Stack Data Sources (downloaded daily) Game Data (Appsflyer, Google Firebase, ...) Attribution Platform (Appsflyer, ...) UA Channel Data (Facebook, Google UAC, Ad-n...) Store Revenue Data (G Play, iTunes)
- 18. SuperScale Analytics + UA Stack Data Sources (downloaded daily) Data Storage Materialized Views Google BigQuery UA Campaign Evaluation MVs Game Data (Appsflyer, Google Firebase, ...) Attribution Platform (Appsflyer, ...) UA Channel Data (Facebook, Google UAC, Ad-n...) Store Revenue Data (G Play, iTunes) Google Cloud Platform Machine Learning (stochastic algorithms, random forest, k-means clustering, …) SBDW ETL Processing (User States, Materialized Views)
- 19. SuperScale Analytics + UA Stack Data Sources (downloaded daily) Data Storage Materialized Views Application Google BigQuery Data Validator (store revenue vs app revenue, AF revenue vs store) ROI/ROAS Spreadsheets UA Campaign Evaluation MVs Game Data (Appsflyer, Google Firebase, ...) Attribution Platform (Appsflyer, ...) UA Channel Data (Facebook, Google UAC, Ad-n...) Store Revenue Data (G Play, iTunes) FB Campaign Optimization Data Visualisation Campaign LTV/ROI prediction Google Cloud Platform Machine Learning (stochastic algorithms, random forest, k-means clustering, …) SBDW ETL Processing (User States, Materialized Views)
- 20. Warsaw | 14-15 October 2019 Building predictions from a bottom-up Daily cohort 1 Daily cohort 2 Daily cohort 3 .. UA DATA PREDICTIONS for each cohort AGGREGATION of cohorts predictions on a higher level .. Weekly Monthly Quarterly Campaign Overall CPI LTV Conversion, ARPU, ARPPU, # purchases, revenue, payers distribution (absolute values, relative values) Campaign success prediction Weighting of data points based on recency
- 21. Warsaw | 14-15 October 2019 Cloud pipeline takes care of automatic scaling of projects. Cloud prediction pipeline Data warehouse Batch preprocessing Machine learning BigQuery Creating models for: Daily, weekly, monthly cohorts and campaigns Additional analysis: Geo analysis Break-even analysis Business Intelligence Data Studio Scheduler Cloud Functions Models storage Google Cloud Storage Google Dataflow ML Engine on AI platform Raw GAME Data + Campaign data + LKLs data Periscope
- 22. Warsaw | 14-15 October 2019 Supporting decision making - examples Weekly cohort estimate d LTV IAP D7 ROAS performance per country Weekly cohort LTV performance estimation Campaigns breakeven analysis Top country to target for LKLs Which strategy was the best When can we expect payback
- 23. Warsaw | 14-15 October 2019 - LTV/ROAS predictions can help to identify opportunities but timing is the factor that makes or breaks any modelling - Aggregating weekly and monthly predictions from daily ones can improve accuracy significantly - Weighting data points based on recency can improve accuracy of estimations Takeaways
- 24. Warsaw | 14-15 October 2019 Thank you for attention. Questions?