ATA is the acronym of Action To Action. In online advertising, users go through a marketing funnel defined by several actions that advertisers can track and analyse: impression view, ad click, app install, service registration, purchase, etc. ATA models estimate the probability of a user engaging in a certain action of the funnel provided that it has engaged in the previous one. 


As industry matures, some advertisers try to align the cost of users acquired with the lifetime value of these users. Sometimes this means asking DSPs to optimize campaigns towards latter steps in the user funnel. However, the further the optimization action is in the funnel, the less conversions there will be for DSPs to feed their machine learning algorithms with.

As a matter of fact, DSPs are required to provide better predicting models for smaller conversion rates with less training data. Otherwise, campaigns may not reach performance goals or volume targets. 


To make the most of our data, we break the user journey through the funnel in the actions that are most relevant for the campaign. An e-commerce campaign optimizing towards purchase might be broken down to impression-to-install and install-to-purchase problems.

Independent models are fitted and optimized for their specific problem. This allows us to learn from intermediate actions that would otherwise be ignored. Intermediate conversion rates are one or two orders of magnitude higher and easier to predict. First steps in the funnel have lots of data and models can have lots of features and low regularization. Latter steps in the funnel have less data and models must have fewer but more relevant features and high regularization to prevent overfitting.

The final estimation of the user conversion rate for the campaign is a combination of all estimated intermediate conversion rates. 

We have observed that this methodology not only provides better performance and volumes than other solutions in a wide variety of situations but also automates away a considerable effort in setting up campaign parameters and manual optimizations.


All in all, to provide advertisers with accurate prediction models for their post-install-action optimized campaigns, we break the funnel into relevant actions so that we can squeeze as much value as possible from all the data we have as well as keep manual setup and optimization efforts in check.