Marketing Mix Modelling

✍ Project Overview

The World Gold Council needed to understand whether their marketing spend was actually working, and if so, then how much, on which channels, and with what lag. Gold is a considered purchase driven by macroeconomic sentiment as much as brand preference, which makes it a uniquely complex environment for marketing effectiveness modelling. I was tasked with building a Marketing Mix Model from the ground up to give WGC the ability to answer these questions with statistical confidence.

📐 Building the Model

The model was built in Python on Google Colab, using a multivariate regression framework with adstock transformations to account for the carry-over effect of media exposure over time, and saturation curves to capture diminishing returns at higher spend levels. Each channel was modelled individually before being combined into a unified model, allowing us to isolate the contribution of paid media from baseline sales drivers like macroeconomic conditions, gold price fluctuations, and seasonality.

📊 The Google Sheets Effectiveness Layer

Alongside the Python model, I developed a complementary effectiveness multiplier model in Google Sheets, designed to make the outputs accessible and actionable for non-technical stakeholders. This layer translated the regression outputs into channel-level ROI scores, budget scenario planning tools, and spend optimisation recommendations that the client could interrogate directly without needing to touch the underlying code.

🔍 Benchmarking Against Meridian

The model's methodology and outputs were benchmarked against Google Meridian, Google's open-source MMM framework. This validation exercise gave WGC confidence that the approach was methodologically sound and comparable to best-in-class industry standards, while also surfacing where the bespoke model offered advantages specific to their business context.

📚 We Learned

1. How to build a full MMM pipeline in Python, from raw data ingestion to model output.
2. How to apply adstock and saturation transformations to media variables in a regression framework.
3. How to isolate paid media contribution from macroeconomic and seasonal baseline drivers.
4. How to translate complex statistical outputs into accessible planning tools for non-technical stakeholders.
5. How to benchmark a bespoke model against an industry-standard framework (Google Meridian).
6. How modelling for a considered, macro-sensitive category like gold differs from FMCG or direct-response contexts.