Building In-House Hierarchical MMM Capability Across 50 US DMAs

L.L. Bean

Retail & E-Commerce
PyMCPyMC-MarketingCausalPyHSGP

The Challenge

L.L. Bean was paying for a vendor-built marketing mix model that covered 50 US markets with dozens of variables across multiple purchase channels. It might have been excellent. They had no way to know — the model was a black box. They couldn't see the methodology, couldn't validate the assumptions, and couldn't adapt it when business questions changed.

They wanted to build their own, with full transparency, and they had a particular attribution challenge that mattered deeply to the brand: understanding how their iconic catalog drives searches and eventually purchases through a multi-step funnel that most marketing models ignore entirely.

Our Approach

We coached the L.L. Bean team through building a geographic hierarchical model from scratch — one that shares patterns across all 50 markets while allowing each region to have its own character. The model incorporated time-varying parameters so marketing effectiveness could shift over time rather than being frozen at a single estimate.

For the catalog question, we built a funnel model that traces the path from catalog distribution through to online searches and eventual purchases, giving the team a principled way to attribute sales that flow through that distinctive channel. We also integrated geographic experimentation methods so they could validate model predictions with real holdout tests in specific markets.

Results

The team progressed from a single-market prototype to a full 50-market hierarchical model. What started as a knowledge transfer engagement extended twice as the team's ambitions grew alongside their capabilities.

They went from being unable to inspect their own marketing model to owning every assumption in it — and having the skills to evolve it as their questions change.

PyMC Labs Team

  • Niall Oulton
  • Ben Vincent
  • Will Dean
  • Joe Wilkinson
  • Carlos Trujillo

More Work in Retail & E-Commerce

HelloFresh

Rebuilt HelloFresh's marketing measurement and testing infrastructure, cutting model runtimes from 20 minutes to 2 and overnight test pipelines from 6 hours to 6 minutes across 15+ global markets.

MMMGaussian ProcessesHSGPA/B Testing

Fabletics

Guided the Fabletics data science team through a major model modernization after their lead modeler departed, upgrading their marketing mix model and adding time-varying parameters.

MMMHSGPTime-Varying ParametersGeo Testing

Swarovski

Improved prediction accuracy by 20% for Swarovski's marketing mix model by letting the baseline sales level shift naturally over time instead of holding it fixed.

MMMHSGPTime-Varying InterceptAdstockSaturation

Let's Chat, We Respond Fast

Tell us about your problem. We typically respond within 24 hours.

Schedule a Consultation