Mathematical Approach to Advertising on Facebook Gleb Mashchenko Business Development Director, RoasUp, Inc. [email protected] 

ROASUP is a performance marketing agency for mobile games focused on user acquisition from Facebook [email protected] 2 Our Clients 

Facebook is a Powerful UA Platform for Mobile Games [email protected] 3 * The AppsFlyer Perfomance Index - Edition VII, H1 2018 https://www.appsflyer.com/2018indexpage/ Power Ranking* 

Facebook Ad Account Structure [email protected] 4 Facebook Account Ad Campaign Ad Set Ad Set Ad Set Ad Ad Ad Ad Campaign 

Ad Set Parameters [email protected] 5 to show WHAT to show WHERE to optimize HOW Creative Headline Text CTA Audience Placement Device Optimization Conv. window CBO BUDGET Now you have to set 25+ parameters for an ad set! 

Conversion Funnel and ROAS [email protected] 6 Impressions Clicks Installs Payers 10000 100 50 3 $$$ Spend $$$ Revenue ROAS = Revenue Spend 

ROASUP Data-driven Mathematical Approach [email protected] 7 • Collect data from Facebook • Automated data analysis 

How Do Ad Set Parameters Influence ROAS? [email protected] 8 # Parameter H* 1 Creative 847 2 Optimization 199 3 Optimization conversion window 54 4 Location 40 5 Gender 28 * Kruskal W. H., Wallis W. A. Use of ranks in one-criterion variance analysis. // Journal of the American Statistical Association. — 1952, 47 № 260. — pp. 583–621. Asymptotic significance (p << 0.001) 

Creative Performance [email protected] 9 Most of creatives do not receive enough spent for valid A/B tests. Facebook automatically optimizes creatives in an ad set 

Creative Performance [email protected] 10 Performance of creatives dynamically changes 

Creative Performance [email protected] 11 ROAS is not distributed normally among launches of any creative Standard methods of classic statistics are not applicable! K-S d=0.27405, p<0.01, Shapiro-Wilk W=0.59, p=0.0000 

Creative Performance [email protected] 12 • ROAS distribution estimation • Bayesian Statistics • Automated analysis of creatives performance dynamics ROASUP Approach 

Optimization and Conv. window Performance [email protected] 13 LTV of users from different optimizations varies LTV Days 

Optimization and Conv. window Performance [email protected] 14 LTV is high, but CPI is high too! Cohort ROAS Days 

Optimization and Conv. window Performance [email protected] 15 CPI ROAS CPI and ROAS depend on optimization and targeting High CPI is not an issue if ROAS is high! 

Optimization and Conv. window Performance [email protected] 16 • Analysis of cohort ROAS for each optimization and targeting • Maximization of ROAS through selection of optimization and conversion window for each targeting • Performance dynamics tracking ROASUP Approach 

Ad Set Control Problem [email protected] 17 Ad set dynamics is described by non-stationary stochastic process 

Ad Set Control Problem [email protected] 18 Lifetime of the most of ad sets is less than 7 days 

Ad Set Control Problem [email protected] 19 N of Ad Sets N of Ad Sets Opt. day Opt. day Data imbalance: +90% of ad sets do not meet KPI The first day is optimal for turning off for the most of ad sets 

Ad Set Control Problem [email protected] 20 Simple methods of machine learning lead to the issues: • Unstable classification of ad sets (classes imbalance, sparse data) • Complex dynamic of an ad set depends on the fast changing market and changing performance of creatives • Low amount of data 

Ad Set Control Problem [email protected] 21 • Deep Reinforcement Learning • Generative models (GAN, VAE) • Autoencoders ROASUP Approach 

Conclusion [email protected] 22 Efficient user-acquisition on Facebook requires automated mathematical analysis 

[email protected] 23 Thank you! Gleb Mashchenko Business Development Director, RoasUp, Inc. Email: [email protected] Facebook: gleb.mashchenko Skype: gleb.mashenko Telegram: @mgleb