Marketing Mix Model: Budget Optimization

Business Question

Given a fixed marketing budget, are we spending it in the right places, and how should we reallocate it?

---

Approach

Traditional attribution (last-click, or "naive ROAS") systematically misleads budget decisions: it ignores the delayed effect of advertising (you see an ad Monday, you buy Friday) and the diminishing returns from overinvesting in a single channel (doubling Facebook spend doesn't double Facebook sales).

Marketing Mix Modeling (MMM) solves both problems. By fitting a Bayesian model with adstock (time-lagged carry-over effects) and saturation (S-curve diminishing returns) transformations, we can estimate the true, causally-correct contribution of each channel to sales and use that to mathematically optimize how the budget should be split.

---

Key Findings

• Email drives ~47% of total incremental sales, despite representing only 25% of total spend: the highest absolute contributor of any paid channel.
• Facebook and Google Search are over-invested relative to their saturation curves: both channels are operating in the diminishing-returns zone, where additional spend yields minimal incremental lift. The optimizer eliminates both from the recommended allocation.
• YouTube Paid and YouTube Organic show significant headroom: current spend sits on the steep part of both saturation curves, indicating strong marginal returns from additional investment.

---

Recommendation

Reallocating ~23% of budget away from Facebook and Google Search toward YouTube Paid and YouTube Organic is projected to increase incremental sales by approximately 1.2% with no change to total spend. The key reallocation is shifting dollars away from over-saturated channels (Facebook: -9pp, Google Search: −14pp) and into channels with remaining capacity on their response curves (YouTube Paid: +11pp, YouTube Organic: +10pp).

---

Repository Structure

mmm-budget-optimization/
├── data/
│   └── media_spends.csv          # Kaggle dataset
├── notebooks/
│   └── mmm_analysis.ipynb        # End-to-end analysis: EDA → MMM fit → optimization
├── requirements.txt
└── README.md

---

How to Run

1. Install dependencies

# Create a virtual environment (recommended)
python -m venv .venv
source .venv/bin/activate   # macOS / Linux
# .venv\Scripts\activate    # Windows

# Install packages
pip install -r requirements.txt

2. Launch the notebook

jupyter notebook notebooks/mmm_analysis.ipynb

Run all cells top-to-bottom (Kernel → Restart & Run All). Model fitting (Phase 2) takes approximately 5–15 minutes.

---

Caveats

• Single market: The model is fit on Division A (one of 26). Results are directionally representative but should be validated across divisions before acting on them.
• Impressions as spend proxy: The dataset contains impression/view counts rather than dollar spend, so budget allocation is expressed in normalized activity units. A production version would use actual media spend in dollars.
• No external controls: Seasonality, competitor activity, and macro-economic factors are not modeled. In a production MMM, these would be included as control variables to isolate true media contribution.