Customer Analytics Dashboard

Interactive Streamlit dashboard for customer behaviour analytics using the UCI Online Retail II Dataset.

The app includes:

• sales and product performance views
• RFM customer segmentation
• churn propensity modelling
• probabilistic CLV prediction
• revenue time-series forecasting

Features

• Overview: KPI cards, monthly trends and top countries/products by revenue
• RFM Segmentation: K-Means baseline with elbow/silhouette previews and R/F/M-derived segment labels, plus a Compare tab benchmarking Gaussian mixture, Agglomerative and HDBSCAN side by side
• Churn Prediction: random forest classification (5-fold stratified CV) with threshold tuning, confusion matrix and precision-recall / ROC curves
• CLV Prediction: BG/NBD + Gamma-Gamma modelling via pymc-marketing
• Revenue Forecasting: SARIMA, Theta, ETS (Holt-Winters) and a seasonal-naive model (statsmodels) with holdout backtest metrics and configurable prediction intervals. Every model is scored against the seasonal-naive baseline ("same period last year"), so the backtest shows whether it actually beats the obvious benchmark. Short backtest windows (< 2 seasonal cycles) reconstruct seasonality from training data only, so the holdout forecast stays seasonal without look-ahead

Setup

Requires Python 3.13+.

1) Install dependencies

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

2) Build the data artifacts

The app reads precomputed Parquet files rather than the raw CSV (see Data pipeline). Generate them once from the source CSV:

python scripts/build_dataset.py

This writes data/online_retail_clean.parquet, data/online_retail_cancels.parquet and data/dataset_meta.json. Re-run it whenever the raw CSV or the cleaning logic in utils/data.py changes.

3) Run the app

streamlit run app.py

Project Structure

.
├── app.py                      # entry point; defines st.navigation
├── app_pages/
│   ├── 1_Overview.py
│   ├── 2_RFM_Segmentation.py
│   ├── 3_Churn_Prediction.py
│   ├── 4_CLV_Prediction.py
│   └── 5_Revenue_Forecasting.py
├── utils/                      # shared package
│   ├── data.py                 # loading, cleaning and Parquet artifacts
│   ├── transforms.py           # RFM, cohort, churn and CLV feature builds
│   ├── clustering.py           # segmentation algorithms and metrics
│   ├── theme.py                # Plotly template and colour palette
│   └── ui.py                   # page chrome, headers, sidebar filters
├── scripts/
│   └── build_dataset.py        # precompute cleaned Parquet artifacts from the CSV
├── .streamlit/
│   ├── config.toml             # theme and server config
│   └── pages.toml
├── data/
│   ├── online_retail_II.csv            # raw source (build input only)
│   ├── online_retail_clean.parquet     # cleaned orders (read at runtime)
│   ├── online_retail_cancels.parquet   # cancel/return invoices
│   └── dataset_meta.json               # raw row count for data-quality stats
├── static/                     # fonts, logos, favicon
├── requirements.txt
└── LICENSE

Dependencies

• streamlit for the app UI
• pandas, numpy, pyarrow for data wrangling and Parquet I/O
• plotly for interactive visualisations
• scikit-learn, scipy for clustering and transformations
• statsmodels for SARIMA and Theta-method forecasting
• pymc-marketing (with its pytensor backend) for CLV modelling

See requirements.txt for pinned version constraints.

Data pipeline

The raw CSV is ~90 MB / 1.07M rows. Parsing and cleaning it on every cold start peaks at ~1 GB RSS, which exceeds the Streamlit Community Cloud memory limit and causes OOM restarts. To avoid this, the cleaning runs once, offline, via scripts/build_dataset.py:

• utils/data.py holds the cleaning logic (_build_clean_orders, _build_cancels).
• The build script writes compact Parquet artifacts (~7 MB total).
• At runtime, load_data() / load_cancels() read those Parquet files (with a CSV fallback if they are missing), so the app never re-parses the CSV.

This drops the cold-start peak memory from ~1 GB to a few hundred MB and the resident dataset from ~125 MB to ~80 MB (the heavily repeated Description column is stored as a category).

Notes

• At runtime the app reads data/online_retail_clean.parquet (in utils/data.py); reads are cached with @st.cache_data. The raw online_retail_II.csv is only needed by the build script.
• Navigation is defined programmatically via st.navigation in app.py.

License

Released under the MIT License.