GulShair/marketing-mix-modeling-app

Marketing Mix Modeling Dashboard

An interactive Streamlit app for exploring how advertising spend across TV, Radio, and Newspaper relates to Sales. It uses a polynomial regression model trained on the classic advertising dataset to predict sales, analyze ROI, simulate scenarios, optimize budget allocation, and surface data-driven insights.

What it does

The app has seven pages, selectable from the sidebar:

• —Dashboard — headline metrics, channel/sales correlations, spend share, a correlation heatmap, sales distribution, and a 3D predicted-sales surface.
• —Sales Predictor — enter a spend mix and get predicted sales, a data-grounded Low/Medium/High category, and an estimated error range based on the model's test RMSE.
• —ROI Analyzer — converts predicted sales to revenue and computes profit and ROI%, with a gauge chart, interpretation bands, and CSV export.
• —What-If Simulator — apply percentage changes to a baseline mix and compare baseline vs scenario sales (and revenue) side by side.
• —Budget Optimizer — finds the channel split that maximizes predicted sales for a given total budget, bounded to the model's trustworthy range.
• —Insights — automated findings computed from the data and model (correlations, Newspaper's weak signal, TV diminishing returns, contribution breakdown, reallocation upside) plus per-channel diminishing-returns charts. Exportable to CSV.
• —Upload Dataset — upload your own CSV, validate it, preview it, and optionally retrain the model with a side-by-side metric comparison before applying.

How to run

Requires Python 3.9+.

pip install -r requirements.txt
streamlit run app.py

The app opens in your browser (default http://localhost:8501).

To retrain the bundled model from the command line:

python model/train_model.py

Folder structure

MarketingMixApp/
├── app.py                  # Streamlit app: all pages, layout, styling
├── requirements.txt        # Python dependencies
├── README.md
├── .streamlit/
│   └── config.toml         # Dark theme configuration
├── data/
│   └── advertising.csv     # Training dataset (TV, Radio, Newspaper, Sales)
├── model/
│   ├── train_model.py      # Training pipeline + reusable retrain functions
│   ├── predictor.py        # Loads the saved model and predicts sales
│   └── saved_model.pkl     # Persisted model, transformer, and metrics
└── utils/
    ├── roi.py              # ROI calculation logic
    ├── optimizer.py        # Budget allocation optimizer (scipy SLSQP)
    └── insights.py         # Data/model-driven insight generation

Model and its limitations

The model is a degree-2 polynomial regression (PolynomialFeatures + LinearRegression) trained with an 80/20 train/test split.

Please read these caveats before trusting any number:

• —Small dataset. The advertising dataset has only ~200 rows. Estimates are directional, not precise, and the model can be sensitive to the specific train/test split.
• —Newspaper has weak signal. Newspaper spend correlates only ~0.16 with Sales in this data — essentially no meaningful predictive relationship. The app flags this explicitly and the optimizer caps Newspaper allocation so it isn't over-weighted on numerical noise.
• —Extrapolation is unreliable. The model is only trustworthy within the range of spend it was trained on. Inputs outside that range trigger a warning, and the optimizer is bounded near each channel's observed maximum.
• —Predictions are estimates, not guarantees. Real sales depend on many factors outside this dataset (seasonality, competition, pricing, etc.). Retraining on new data is not guaranteed to improve the model — always compare metrics before applying a retrained model.