Return to page


Epsilon Increases its Customers’ Marketing ROI with

Acordocerto Logo Acordocerto Logo

+15K Highly relevant customers into marketing campaigns



+$9M Incremental revenue for campaign

For the past 30 years, Epsilon’s data Coop, Abacus has helped over 3,000 brands across the markets of Catalog, Consumer Services, Pureplay online, Not-for-Profit, Publishing and Business-to-Business. Abacus helps brands reach the right consumers with the right product, and offers, in any channel at the right time.  Epsilon’s customers are able to directly tie advertising costs to revenue, and they’ve continued to lead the market in delivering results.  Epsilon’s Abacus move to machine learning with has furthered their lead, driving a 3-5% improvement in a clients direct mail response rate,.  That 3% translates into finding an average of 15,000 more highly relevant customers into every marketing campaign, driving more gross demand for the brands Epsilon helps.   



When Epsilon’s customers want to run a marketing campaign, Abacus builds them a custom modeled list of consumers who are likely to be interested in their catalog, product, or offer.  This amounts to more than 100,000 models a year.  In 1990, the Epsilon team built out an inventive automated system using SAS statistical software.  The system produced multiple models for each customer campaign, stack ranked the models, and then used the combination of all the models to create the best lists for their customers.  As machine learning adoption increased, Epsilon’s Abacus was challenged to see if they could improve their list accuracy and continue to lead the market in outcome-based marketing.  They faced scale and throughput challenges as the system needed to be able process terabyte sized data sets and build over 100,000 ML models a year.



Epsilon partnered with to develop their ML system, the first product was named Accelerate.  Similar to the SAS workflow, Epsilon builds multiple models for each campaign and combines them to build highly relevant custom lists for their customers.  The solution starts with a proprietary feature engineering process before performing feature reduction in pySpark, a Python API for Spark. Jobs are then queued for the Machine Learning infrastructure.  Based on the job, the system automatically provisions the right-sized machine learning cluster.  The model is then built and the cluster is shut off.  When the model is complete, Epsilon interprets the data, scores the model, and adds it to their audience structure for the campaign.  This workflow allows their data science team to continually monitor the models and audience structure to ensure they are meeting their customers goals and quickly adjusting when goals change.


Epsilon selected based on their ability to directly support the algorithms (written by, process large data sets, and deeply explain models.  Model interpretability was of paramount importance as Epsilon needs to prove why people were added to certain lists and ensure there is no bias when the lists are created.

quotation mark

In the environment of increasing privacy regulations, the tools Epsilon employs need to provide transparent list selection. help us meet this challenging and critical need.

Andrea Thornton, VP of Analytics


Moving from SAS to improved list response rate by 3-5%, providing significant results for the brands Epsilon supports.  By removing inaccurate targets and adding relevant consumers, Abacus customers are now seeing higher gross demand revenue from their acquisition audiences. For a large brand selling gifts for consumers and B2B, Abacus is able to increase their direct mail response rate by 1.10% which drove an incremental $9.0M for just one campaign.  Epsilon is helping to support 8,000 campaigns a year, so the total business impact of improvement is significant and enabling Epsilon Abacus to continue to lead the market.