Estimating Marketing Component Effects: Double Machine Learning from Targeted Digital Promotions

For your marketing

Possible founder application

A detailed example of measuring message components against both engagement and purchase outcomes. Test words, offer type, and audience interactions; do not optimize opens as a proxy for sales.

Reported result

What the study found

Content and framing had heterogeneous effects along the funnel. Clearance framing outperformed product-specific discount framing in the reported setting; top-of-funnel engagement did not always predict purchases. Off-policy evaluation estimated about an 11% revenue gain from shifting better clearance offers toward highly engaged customers.

Research question

What the authors investigated

Which individual email-promotion components affect opens and purchases, and how do responses vary across customers?

Design

Double machine-learning analysis of promotion components and customer heterogeneity in a large retailer's email campaigns, with off-policy evaluation of alternative targeting.

Sample and setting

34 email promotions sent to more than 1.3 million customers of an online apparel retailer over two months in 2015.

Scope and transfer

Limitations to keep in view

The promotions are observational rather than randomized treatments, and the estimated counterfactual policy is model-based. Confidentiality prevents reporting full subject lines, and limited margin data constrain profit conclusions.

Research sources

INFORMS publisher abstract and page retrieved with Firecrawl; sample, approach and stated limitations were included in the extracted page text.