November 27, 2025
By Thomas Wiecki
This event is a discussion between Luca Fiaschi and Thomas Wiecki about Bayesian modeling in online marketing with a special focus on Marketing Mix Models and Customer Lifetime Value models.
In the panel discussion, Luca Fiaschi shares his thoughts on the current state and the future of Bayesian data science in online marketing, based on many years of real-world experience leading teams at HelloFresh and Stitchfix. A special focus will be placed on Marketing Mix Models (MMM) and the work PyMC Labs has done with Luca during his time at HelloFresh. In addition, Customer Lifetime Value models play an ever more important role in marketing but remain stuck in old modeling approaches. We will discuss a potential way forward.
00:00 Introduction by Thomas
01:39 Luca Fiaschi introduces himself
02:37 Alex Andorra introduces himself
03:57 Attribution methods have multiple use cases across businesses
04:39 Analyze media models
05:23 Find Optimal Budget Allocation(What-if scenarios and optimization)
05:55 Forecasting
06:27 Most common attribution methods
11:02 Structure of a Media Mix Model
12:04 Saturation and Adstock functions
14:04 Advantages of Bayesian Media Mix Models
17:14 Bayesian MMM can be calibrated to ensure consistency with incrementality measurements
19:50 Challenges encountered when developing Hello Fresh’s model
24:36 PyMC Labs work with Hello Fresh to build the model
33:12 Comparing the Hello Fresh model with the other different frameworks
36:05 Business insights that can be derived from the Hierarchical Gaussian Processes model
40:10 Question: How can business people and Data science efforts be aligned in an organization?
44:23 Question: Have you considered introducing relationships between regressor variables
48:19 Question: To what level of confidence are you able to say that you are at a certain saturation level and can you use Bayesian methods to do that?
54:28 Question: Have you ever settled for Gaussian Random walk parameters over latent Gaussian Processes(GPs) for the sake of simplification or explainability?
59:19 Question: Media Mix Models suffer from multicollinearity, any advice?
1:02:14 Thank you and closing remarks