|
Abstract:
|
Marketing campaign metrics often include counts of non-purchase customer-engagement actions, such as membership sign-ups and first-time subscriptions. These actions are usually sparse, occurring infrequently during the entire customer lifetime, sometimes months after the campaign that might have prompted the initial interest. Measuring the effects of campaigns on these metrics is thus a challenging problem, requiring both view-through and multi-touch attribution models, over time periods longer than what is commonly used for purchase-based metrics. In this work, we argue that instead of focusing on simple counts of actions in a given time period, a better strategy is to model the time to the action event. Time-to-event models are able to extract information from both the counts and the timing of these events, resulting in more accurate conversion metrics for different groups of customers. In addition, these models can predict which customers’ events have not been observed yet (“censoring”) and which will never happen (“cure fraction”), thus allowing marketing managers to decouple demand-shifting from demand-creating mechanisms of marketing.
|
