Marketing Mix Modeling with Robyn

Multi-channel marketing budget optimization with Meta’s Robyn framework

Context & Problem

Business question: How should a CMO reallocate marketing budget to maximize sales?

In a context where marketing budgets are scrutinized and attribution becomes increasingly complex (end of third-party cookies, multiplication of channels), Marketing Mix Modeling (MMM) is experiencing renewed interest.

MMM allows you to:

• Measure the incremental impact of each marketing channel on sales
• Account for saturation effects (diminishing returns)
• Model carryover effects (adstock) - the impact of a campaign that persists over time
• Optimize budget allocation across channels

Approach & Methodology

Robyn Framework

I used Robyn, the open-source framework developed by Meta (Facebook), which combines several advanced techniques:

# Robyn model architecture
# 1. Time decomposition with Prophet
#    - Trend
#    - Seasonality (weekly, annual)
#    - Holiday effects

# 2. Marketing variable transformations
#    - Geometric adstock: effect(t) = spend(t) + θ × effect(t-1)
#    - Logistic saturation: response = α / (1 + exp(-β × spend))

# 3. Ridge regression to avoid overfitting

# 4. Multi-objective optimization with Nevergrad
#    - Minimize NRMSE (prediction error)
#    - Minimize DECOMP.RSSD (decomposition consistency)

Channels Analyzed

• Google Ads: Search and Display
• Facebook Ads: Social and remarketing
• LinkedIn: B2B
• Email: Newsletters and promotions
• Organic: SEO and direct traffic

Complete Workflow

Raw data → Feature Engineering → Robyn Training → Model Selection → Budget Optimization

1. Data collection: Daily sales, spend by channel, contextual variables
2. Feature engineering: Creating seasonality variables, special events
3. Training: Calibrating adstock and saturation parameters
4. Model selection: Pareto analysis (NRMSE vs DECOMP.RSSD)
5. Optimization: Budget allocation scenario simulation

Key Results

Sales Decomposition

The model decomposes sales into:

• Baseline: “Natural” sales without marketing (~60%)
• Channel contribution: Incremental impact of each investment
• Seasonality: Calendar-related variations

Response Curves

Each channel has a different response curve, showing diminishing returns:

• Google Ads: High initial ROI, rapid saturation
• Facebook: More linear curve, less saturation
• Email: Excellent ROI but limited capacity

Optimization Recommendations

The dashboard allows simulating different scenarios:

• Current scenario: Historical allocation
• Optimized scenario: Rebalancing toward under-invested channels
• Constrained scenario: Optimization with fixed budget

Demo

View Shiny Dashboard

The interactive dashboard allows you to:

• Visualize sales decomposition
• Explore response curves by channel
• Simulate allocation scenarios
• Compare marginal ROIs

Technologies

Component	Technology
Language	R
MMM Framework	Robyn (Meta)
Time Decomposition	Prophet
Optimization	Nevergrad
Dashboard	Shiny
Environment Management	renv

Source Code

The project is organized into several components:

mmm-robyn-project/
├── R/
│   ├── 01_data_preparation.R    # Data preparation
│   ├── 02_robyn_training.R      # Model training
│   ├── 03_model_selection.R     # Pareto selection
│   └── 04_optimization.R        # Budget optimization
├── app/
│   └── app.R                    # Shiny dashboard
├── data/
│   └── marketing_data.csv       # Input data
└── outputs/
    ├── exploration/             # EDA visualizations
    ├── robyn/                   # Model results
    └── optimization/            # Optimized scenarios

Learnings

This project allowed me to deepen:

1. Marketing modeling: Understanding adstock and saturation effects
2. Multi-objective optimization: Finding the right balance between accuracy and interpretability
3. Results communication: Translating complex models into actionable recommendations

---

← Back to Portfolio Analysis | View Dashboard