Multi-touch attribution in the mobile gaming industry
Loading...
URL
Journal Title
Journal ISSN
Volume Title
School of Business |
Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.
Authors
Date
Department
Major/Subject
Mcode
Degree programme
Language
en
Pages
58
Series
Abstract
User acquisition spend is a big investment for mobile gaming companies. Because of the large scale, even small improvements in how this spend is allocated can provide big returns. To allocate advertising spend well; it is important that the credit of a conversion be attributed as accurately as possible. The current attribution model - standard to the industry - is a last-touch attribution model, which attributes 100% of the credit to the last touch-point. However, before a user installs a game they might see ads from multiple channels that might all affect the user’s propensity to install. With the last-touch attribution model, the uplift of these ads is not observed which skews the returns on advertising spent for different channels. This study looks at how install probability develops as impressions per user increase, how long the effect of an ad lasts and attempts to find better attribution models that attribute credit better than the last-touch model. Three multi-touch attribution models are proposed; two based on the Shapley value and one based on the ad effect time decay of different channels. The data for this study comes from a mobile gaming company and consists of impressions seen by both installed and non-installed users as well as impression channels, impression time and install time. The data was collected during a 38-day period and has data from 44,719,217 users who were divided into a training set and a test set with a 70%/30% split. The test set is used to validate the proposed models against the last-touch attribution model by using the models trained on the training set to generate predictions on install probability for user paths in the test data set. The study finds that the ad effect of all channels declines very quickly after the first day and is almost zero at seven days after the impression. The study also attempts to find the correlation between install probability and the amount of impressions a user has seen. Regarding this objective, the study is inconclusive. This correlation behaves very differently between different channels and because the amount of impressions per users could not be controlled for, it is difficult to deduce causation. Out of the three proposed attribution models, only one is able to outperform the last-touch model when it comes to predicting install probabilities from the training set’s paths. The model that outperformed is a Shapley value based model that considers the times of impressions for each path when calculating credit attribution. Finally, the study finds that only 9.5% of observed installs had impressions from more than one channel during a seven-day attribution window. This combined with the difficulty of validating attribution models based on return on advertising spend means that developing a multi-touch attribution model probably is not a very low hanging fruit for performance marketers. What would be worth looking into would be to test optimizing the frequency of ads shown to users.Description
Thesis advisor
Malo, PekkaKuosmanen, Timo