A practical implementation of Marketing Mix Modelling (MMM) using Bayesian methods, based on the concepts from Marketing Mix Modelling: The Complete Guide by Mark Stent.
The accompanying guide was written to incorporate as many MMM concepts as possible into a single resource. However, it's important to note:
- Not exhaustive: The guide doesn't cover every aspect of MMM - it's a comprehensive starting point, not the final word
- Bayesian focus: This implementation uses Bayesian methods exclusively. There are other valid approaches (frequentist regression, machine learning methods, etc.) that are not covered here
- Two implementations: This repository includes both additive and multiplicative model specifications, demonstrating different approaches to MMM
The goal is to provide a solid foundation for understanding and implementing Bayesian MMM, which you can then adapt to your specific needs.
MMM-guide/
├── notebooks/ # Jupyter notebooks with full MMM examples
│ ├── mmm_complete_example.ipynb # Additive model
│ └── mmm_multiplicative_example.ipynb # Multiplicative model
├── frontend/ # Next.js web application
│ └── src/app/ # React pages (upload, explore, config, etc.)
├── backend/ # FastAPI backend server
│ ├── main.py # API endpoints for model training
│ └── core/ # Core MMM modules
│ ├── __init__.py
│ ├── transformations.py # Adstock, saturation functions
│ ├── models.py # PyMC model building
│ ├── attribution.py # Channel contribution analysis
│ └── optimization.py # Budget optimization
├── conjura_mmm_data.csv # Raw dataset
├── mmm_weekly_clean.csv # Preprocessed weekly data
├── pyproject.toml # Python dependencies
└── README.md
All notebooks are located in the notebooks/ folder.
| Notebook | Model Type | Description |
|---|---|---|
notebooks/mmm_complete_example.ipynb |
Additive | Standard MMM: Sales = Baseline + Media_Effects + Controls |
notebooks/mmm_multiplicative_example.ipynb |
Multiplicative | Log-log and lift-factor models with Shapley decomposition |
- Additive Model: Simpler interpretation, straightforward decomposition, works well for most cases
- Multiplicative Model: Better when channels interact strongly, coefficients are elasticities, requires Shapley values for attribution
The project includes a modern web application (MMMpact) that provides a no-code interface for building and deploying Marketing Mix Models. It's designed for marketing analysts and data scientists who want to run MMM analysis without writing Python code.
MMMpact guides you through the complete MMM workflow:
- Upload your data - Drag and drop CSV/Excel files or use the built-in demo dataset
- Explore and validate - Visualize trends, check data quality, and understand your marketing spend patterns
- Configure the model - Set adstock decay rates, saturation curves, and Bayesian priors per channel
- Train with confidence - Run MCMC sampling with real-time progress and convergence diagnostics
- Understand results - Interactive charts showing channel contributions, ROI, response curves, and model fit
- Plan budgets - Optimize allocation across channels, set constraints, compare scenarios, and export recommendations
- Marketing analysts who need to measure channel effectiveness without coding
- Data scientists who want a faster way to iterate on MMM configurations
- Agencies presenting MMM insights to clients with interactive dashboards
- Teams collaborating on budget planning with exportable scenarios
- Start the backend server:
cd backend
uv run uvicorn main:app --reload --port 8000- In a new terminal, start the frontend:
cd frontend
npm install
npm run dev- Open http://localhost:3000 in your browser
- Data Upload: Upload CSV files or use the demo dataset
- Data Exploration: Interactive charts and data quality analysis
- Column Mapping: Map your columns to target, media spend, and controls
- Model Configuration: Configure adstock, saturation, and MCMC parameters
- Model Training: Train Bayesian MMM models with progress tracking
- Results Analysis: View contributions, ROI, response curves, and diagnostics
- Budget Planning: Optimize allocation, set constraints, create scenarios, and export plans
This project uses the Multi-Region Marketing Mix Modeling Dataset from Figshare, which contains e-commerce marketing data across multiple brands and channels.
- Loading and exploring marketing data
- Selecting appropriate data granularity (weekly aggregation)
- Exploratory data analysis
- Handling missing values
- Identifying and handling outliers
- Scaling variables for modelling
- Creating derived variables
- Correlation matrix between channels
- Interpreting correlation levels
- Strategies for handling high correlation
- Adstock (Carryover Effects): Geometric adstock transformation
- Saturation (Diminishing Returns): Hill function implementation
- Complete transformation pipeline
- Prior specification (informed by domain knowledge)
- Complete model structure in PyMC
- MCMC sampling
- R-hat interpretation
- Effective Sample Size (ESS)
- Trace plot analysis
- In-sample fit (R-squared, MAPE)
- Residual analysis
- Posterior predictive checks
- LOO-CV (Leave-One-Out Cross-Validation)
- Response curves with uncertainty
- ROI calculation with credible intervals
- Sales decomposition
- Channel contribution analysis
- Marginal ROI calculation
- Optimal budget allocation
- Constraint handling
Covers the same data preparation as the additive model, plus:
- Log-Log Specification: Elasticity-based model where coefficients represent % change
- Lift-Factor Specification: Multiplicative lifts from baseline with natural interactions
- Shapley Value Decomposition: Fair attribution for multiplicative models
- Model Comparison: Side-by-side comparison of both multiplicative approaches
- Budget Optimization: Allocation optimization using elasticities
Key differences from additive:
Additive: Sales = Baseline + b1*X1 + b2*X2
Log-Log: log(Sales) = a + b1*log(X1) + b2*log(X2)
Lift-Factor: Sales = Baseline * (1 + lift_1) * (1 + lift_2)
- Python 3.11+
- uv package manager
- Node.js 18+ (for the web frontend)
- Clone this repository:
git clone <repository-url>
cd MMM-guide- Install Python dependencies using uv:
uv sync- Install frontend dependencies:
cd frontend
npm install
cd ..- Start Jupyter Lab for notebooks:
uv run jupyter lab- Open
notebooks/mmm_complete_example.ipynband run the cells
Key packages used:
- PyMC: Bayesian modelling framework
- ArviZ: Bayesian diagnostics and visualization
- Pandas/NumPy: Data manipulation
- SciPy: Optimization algorithms
- FastAPI: Backend API server
- Next.js/React: Web frontend
- Recharts: Interactive charts
Sales = Baseline + Trend + Seasonality + Media_Effects + Noise
Sales = Baseline * (1 + TV_lift) * (1 + Digital_lift) * Seasonality * exp(Noise)
Or in log-log form:
log(Sales) = alpha + elasticity_TV * log(TV) + elasticity_Digital * log(Digital) + ...
Models how advertising effects persist over time:
Adstock(t) = Spend(t) + lambda * Adstock(t-1)
Models diminishing returns:
Response = x^S / (K^S + x^S)
Where:
- K: Half-saturation point (spend level at 50% effect)
- S: Shape parameter (controls curve steepness)
The optimizer uses the log-log model formula to maximize total sales:
contribution = avg_sales * (spend / avg_spend)^elasticity
At the optimal allocation, marginal ROI is equalized across all channels.