Marketing Foundation Models Bayesian MMM Without Per-Dataset Fitting

Building a trustworthy Bayesian marketing mix model (MMM) takes a lot of skilled, repetitive work. Someone has to choose a structure, fit it, diagnose it and revisit the assumptions, then do it all again for the next brand, market or quarter. And because the reported uncertainty depends on the one structure the modeler picked, it is usually too narrow. In this session, Carlos Trujillo and Teemu Säilynoja present Marketing Foundation Models (MFM): a model trained on simulated marketing worlds that estimates baseline, channel and control contributions, with uncertainty, in a single pass, without fitting a model to each new dataset. They will walk through how it works, what the evaluation on synthetic data shows, and where this early research prototype still falls short. Luca Fiaschi hosts, with time for your questions. What you will learn: How MFM combines a simulator of marketing worlds, a transformer encoder and flow matching to return contribution estimates without per-dataset fitting. What the synthetic evaluation found: channel contribution error about a quarter of a standard PyMC-Marketing MMM's, and better calibrated uncertainty than standard regression. Why sampling over causal structures gives more honest uncertainty than a single hand-built model. The model's known limits today, and how it differs from foundation models such as TabPFN and TimesFM. Who should attend: Marketing science and analytics teams that already run MMMs, or have been put off by how long they take to build. Data scientists working on Bayesian modeling, causal inference or amortized inference. Analytics leaders who want measurement answers inside the planning cycle. 📜 Outline of Talk / Agenda: 5 min: Introduction to PyMC Labs and speakers 40 min: Panel discussion 15 min: Q&A 💼 About the speakers: Dr. Luca Fiaschi (PyMC Labs Partner, Gen AI Vertical) Luca helps organizations unlock the value of data and AI. With 15+ years of experience, he’s led and scaled teams at Mistpl