Marketing Mix Modeling (MMM) - Weekly Dataset Analysis

📌 Overview

This repository contains a Marketing Mix Modeling (MMM) analysis using a 2-year weekly dataset.
The goal is to explain revenue as a function of marketing spend, promotions, pricing, and other drivers while considering a causal mediation effect of Google spend (as a mediator between social/display channels and revenue).

The analysis is implemented in both:

• mmm_model.ipynb — interactive notebook with plots and insights
• mmm_model.py — reproducible Python script for automated runs

---

⚡ Dependencies

• pandas
• numpy
• matplotlib
• seaborn
• scikit-learn
• statsmodels
• jupyter

⚙️ Environment Setup

1. Clone the repository

git clone  https://github.com/Shruti1suman/MMM_modeling.git

2. Create virtual environment (if not already):

python -m venv venv
venv\Scripts\activate 

3. Install dependencies

pip install -r requirements.txt

#🚀 How to Run Jupyter Notebook

jupyter notebook mmm_model.ipynb

VS Code

python mmm_model.py

🧩 Dataset Columns

• week — Week of observation (datetime)
• facebook_spend, google_spend, tiktok_spend, instagram_spend, snapchat_spend — Paid media spend
• social_followers — Total social media followers
• average_price — Average product price
• promotions — Promotion spend/events
• emails_send, sms_send — Direct marketing channels
• revenue — Revenue (target variable)

🛠️ Modeling Approach

1. Data Preparation

• Converted week to datetime
• Log-transformed paid media spend (log1p) to handle zeros and scale distributions
• Standardized all features using StandardScaler
• Handled weekly seasonality implicitly using time-series CV

2. Causal Framing

• Assumed Google spend is a mediator between social/display channels and revenue
• Included log-transformed social media spend and Google spend as separate features
• Considered back-door paths to avoid leakage

3. Model

• Ridge Regression (with cross-validation)
• TimeSeriesSplit for validation (respecting temporal order)
• Hyperparameters (alpha) tuned using logarithmic grid search

4. Diagnostics

• Residual analysis to check model fit
• CV R² scores for stability
• Feature importance plots for interpretability
• Sensitivity to average_price and promotions explored

📊 Outputs

• All outputs (plots and tables) are saved in the outputs/ folder:

• residuals.png — Residual analysis over time

• feature_importance.png — Coefficients of each feature (interpretable importance)

• Feature importance
  

• Residuals
  

🔍 Key Insights

• Social media spend drives Google search spend, which mediates revenue.
• Promotions and pricing have significant elasticity on revenue.
• Email/SMS campaigns show measurable but smaller effects.
• Ridge regression reduces multicollinearity and produces stable, interpretable coefficients.

📝 Author

-Shruti Suman