Maximizing Marketing Effectiveness with Media Mix Modeling for a Home Services Leader

Solution Architecture

• Data Ingestion & Preprocessing: Cleaned and validated disparate media, lead, and weather data from multiple vendors
• Modeling Framework: Meta's Robyn MMM library in R
• Insights Activation: Generated budget allocation scenarios and ROI insights shared with leadership and media buyers

Machine Learning Implementation

The core of the solution relied on time-series regression modeling with regularization and business-informed feature engineering:

• Response Variable: Daily lead volume
• Media Inputs: Channel-level impressions and spends for TV, search, display, and social
  • Applied adstock transformations to model carryover effects
  • Used diminishing returns curves to capture saturation effects
• Control Variables: Weather trends, holidays, and seasonality
  • Incorporated cooling/heating degree days to model HVAC demand
  • Captured seasonal patterns relevant to home services
• Optimizer: Robyn's budget allocator for scenario planning
  • Simulated spend reallocations to maximize expected leads
  • Generated multiple allocation scenarios for business stakeholders
• Model Validation: Comprehensive validation approach
  • Evaluated NRMSE, bias metrics, and decomposition reasonability
  • Ensured business interpretability of coefficients

Technical Challenges

Implementing this media mix modeling solution required overcoming several complex technical hurdles:

• Disjointed & Evolving Data Structures: Required building flexible and robust data preprocessing pipelines that could adapt to new formats and sources as the project evolved.
• No Centralized Data Warehouse: Data engineering had to be modular and transparent to support reproducibility, with careful documentation of source transformations.
• Data Latency & Sparsity: Marketing data arrived at different cadences with varying levels of detail. We implemented techniques to normalize across time periods and smooth weekly data to increase signal clarity.
• Adstock Parameter Tuning: Finding optimal adstock parameters for each channel was computationally intensive. We developed an efficient grid search approach to identify the best hyperparameter combinations.
• Model Interpretability: Ensuring stakeholders could understand and trust the model outputs required careful balance between technical sophistication and business relevance in our visualizations and explanations.

Design Tradeoffs & Decisions

Several key architectural and modeling decisions were made to ensure the solution was robust, interpretable, and actionable:

• Robyn Chosen Over Manual MMM: Automated hyperparameter tuning and optimizer features allowed for faster iteration and transparent tradeoff analysis than building a custom MMM solution.
• Impression Data Over Spend: Modeled exposure metrics to better reflect media consumption patterns, with spend used for ROI calculations, providing more accurate attribution.
• Environmental Controls Included: Weather variables significantly improved model accuracy, capturing HVAC seasonality that was critical for this home services business.
• Stakeholder Communication: Delivered multiple models and visual summaries to gradually build trust in the data and shift attitudes toward digital media, overcoming initial resistance.