In the display advertising market, one of the most popular advertisers' goals is acquiring conversions such as app installs and purchases, and an important technology that enables the advertising platform to support this campaign goal is to predict conversion rate (CVR).
There are two major difficulties in predicting CVR: one is that conversions often don't happen immediately after a click, and the other is that some advertising products have to accept multiple conversions.
In this paper, we introduce a new model - jointly trained Negative Binomial and Order Statistics - to tackle the multiple conversions and a series of conversion delays, simultaneously.
Our proposed model shows the significant improvement in the real traffic data.