How Predictive Modeling Changed Marketing Investment Decisions at Enterprise Scale

At AB InBev, marketing investment decisions historically relied on a mix of historical allocation patterns and managerial judgment. The question was straightforward: could predictive modeling improve how the organization allocates marketing budgets?

The answer turned out to be more nuanced than expected. Building a model that estimates expected ROI by channel is technically tractable. The hard part is building a system that stakeholders trust enough to change their behavior.

The first version of our investment optimization system reduced analytical processing time by approximately 20%. But the more important outcome was qualitative: marketing leaders started asking 'what does the model say?' before making allocation decisions. That behavioral shift matters more than any accuracy metric.

What I learned about enterprise ML from this project:

Interpretability is non-negotiable. In a boardroom, no one cares about your RMSE. They care about why the model recommends shifting budget from Channel A to Channel B. SHAP values and partial dependence plots became standard deliverables alongside predictions.

Temporal validation is the only honest approach. We used expanding-window cross-validation that respects the time structure of the data. Random splits would have inflated our accuracy estimates and eroded trust when the model underperformed in production.

The pipeline is the product. The model itself is a small part of the system. Data ingestion from multiple Azure sources, feature engineering in Databricks, experiment tracking in MLflow, and result delivery through executive dashboards — each component needed the same engineering rigor as the model.

Start with a clear counterfactual. Before building anything, we defined what 'better' meant in business terms: faster analysis, more scenarios evaluated, and explicit uncertainty quantification. Without this, there is no way to measure success.

This project reinforced my belief that the highest-leverage skill for a data scientist is not model tuning — it's the ability to connect a statistical output to a business decision.