Forecasting Sandbox

Live demo (Hugging Face Space)

Forecast Sandbox Lite — SKU Explorer (Streamlit)

• Open demo (new tab): https://pranavsharma-forecast-sandbox-lite.hf.space
• Hugging Face Space page: https://huggingface.co/spaces/PranavSharma/Forecast-Sandbox-Lite
• GitHub repo: https://github.com/Sharma-Pranav/forecast-sandbox-lite

Embedded demo

30-second walkthrough

1. Pick a SKU.
2. See the recommended model and Score = MAE + |Bias|.
3. Check demand regime (ADI/CV² → Smooth / Erratic / Intermittent / Lumpy).
4. Compare all models (ranked table).
5. View Actual vs Forecast and download aligned CSV.

What this demonstrates

• No universal winner: model performance depends on demand structure.
• Win-rate is a trap: portfolio loss is driven by tail errors.
• A generalist fallback reduces downside when structure is uncertain.

Data powering the Space (precomputed artifacts)

• data/processed/test.csv — ground truth
• metrics/best_models.csv — per-SKU recommendation
• metrics/combined_metrics.csv — model comparison table
• metrics/demand_profile.csv — ADI/CV² regime tags
• metrics/*_predictions.csv — model forecasts

Downloads

• ⬇️ Slides (PPTX)
• View slides on GitHub

The Space is intentionally a viewer for decision evidence (not a training environment).