Insights on Creating a Growth Machine Using Attribution Modelling

Sammanfattning: Given access to detailed tracking data, the problem of attribution modelling has recently gained attention in both academia and the industry. Being able to determine the influence of each marketing channel in driving conversions can help advertisers to allocate their marketing budgets accordingly and ultimately increase their customer base and achieve a higher Return On Investment (ROI). However, Last-Touch Attribution (LTA), the current industry standard to approach the problem, has been criticized for oversimplification.  In this degree project, two data-driven attribution models are therefore compared to the LTA model on real data from an insurance company, with the objective to optimize for customer base growth and ROI. Raw attributions for each channel are obtained after training the models to predict conversion or non-conversion. By using a linear function to obtain a Customer Lifetime Value (CLV) estimate, the attributions are then adjusted to the ROI of each channel and finally validated through an attribution based budget allocation and historical marketing data replay. The experimental results demonstrate that all models reach approximately 82% accuracy on balanced data, just below the calculated theoretical maximum. While current research consistently argues for more complex data-driven Multi-Touch Attribution (MTA) models, this project provides a nuance to this field of research in showing that the LTA model may, in fact, be suitable in some cases. A new approach to develop specialized models based on correlations between conversion and contextual variables, then shows that attribution models for mobile users specifically yield higher accuracy. The sum of such unnormalized attributions function as indicators for the conversion strength of contextual variables and can further assist decision making.