Bayesian MMM with Time-Varying Intercept for Luxury Retail
Swarovski
The Challenge
Swarovski's marketing mix model had a stubborn accuracy problem. Predictions weren't precise enough to trust for budget allocation, and the root cause was structural: luxury fashion has dramatic seasonal swings — holiday gifting, Valentine's Day, summer lulls — plus an organic baseline that drifts year over year. The model assumed a flat baseline. When the true baseline rises and falls but the model can't see it, media channels get blamed for (or credited with) variation that has nothing to do with advertising.
Our Approach
The fix was structural rather than incremental. We replaced the fixed baseline with a smooth, time-varying one that could rise and fall to capture seasonal patterns and organic trends. Media contributions were modeled as additive on top of this shifting baseline, so the model could cleanly separate what advertising drove from what would have happened anyway. Each channel retained its own carryover and diminishing-returns behavior, and we calibrated the model's expectations to match Swarovski's actual revenue scale. The real test was simple: does the new model predict better than what they already had?
Results
It did — prediction error dropped by 20%. More importantly, the attribution picture became far more plausible. Seasonal peaks were absorbed by the baseline where they belonged, rather than being falsely attributed to whatever campaign happened to be running at the time.
For a luxury brand where timing is everything, that distinction matters enormously for planning.
PyMC Labs Team
- Maxim
- Niall
- Thomas Wiecki
Let's Chat, We Respond Fast
Tell us about your problem. We typically respond within 24 hours.
Schedule a Consultation