Panel Time-Series Forecasting — Sales (Notebooks)

A portfolio-ready set of Jupyter notebooks for panel time-series forecasting (daily sales across multiple stores × items).
The project emphasizes clean validation (holdout + rolling-origin backtest), interpretable statistical baselines (SARIMAX, TBATS, ARIMA), and practical automation (AutoTS), with optional Prophet / Darts / NeuralProphet experiments.

Note on data: the original dataset is not redistributed here due to licensing/ownership.
To reproduce, place your CSV/Parquet under data/ and adjust the path in the notebooks. A minimal schema is described below.

What’s inside

• Notebooks (cleaned for sharing):
  • practice.ipynb — main workflow (panel setup, feature engineering: calendar & Fourier terms, backtesting, leaderboard).
  • tbats_prophet_arima.ipynb — focused comparison of TBATS / Prophet / ARIMA with Optuna-tuned ARIMA.
• Validation:
  • Holdout: last 90 days.
  • Rolling-origin backtest (expand) with forecast horizon H=90.
  • Primary metric: SMAPE; secondary diagnostics via MAE/MAPE where relevant.

How to run

1. Python 3.9+ recommended. Create a fresh environment and install deps:

   pip install -r requirements.txt

2. Put your dataset under data/ (see schema below). In the notebooks, set the file path accordingly (e.g., data/HW_train.csv or data/train.parquet).
3. Open the notebooks and run cells top-to-bottom:

   jupyter lab

Data schema (example)

Minimum columns:

• date — daily timestamp (YYYY-MM-DD).
• store_id — store identifier (int/str).
• item_id — item identifier (int/str).
• sales — target numeric value per (store_id, item_id, date).

Optional exogenous features (if available):

• holiday, is_weekend, dow, month, quarter, custom Fourier seasonalities, etc.

If you used a file named HW_train.csv in your course, place that file into data/ and update the paths. If that file originates from a public source (e.g., Kaggle), you can reference its official page in your local notes; this repo intentionally avoids re-hosting third‑party data.

Results (short)

• Leader (stable): AutoTS on panel segments — SMAPE ≈ 8–9% on holdout and backtest.
• Strong classical baseline: SARIMAX (weekly + yearly Fourier) — SMAPE ≈ 9–11%.
• Prophet / TBATS / ARIMA: competitive on select segments; Prophet tends to be stable but slightly weaker; ARIMA benefits from Optuna tuning.

Roadmap (optional)

• Add holiday calendars as exogenous factors for SARIMAX/Prophet.
• Add prediction intervals (Quantiles / Pinball loss) visualizations.
• Try a light ensemble (e.g., average of AutoTS + SARIMAX on overlapping segments).
• Per‑segment error analysis (top outliers, volume vs error).

Environment

See requirements.txt. Heavy libraries (Prophet, PyCaret, Darts, Neural* packages) may take time to build; feel free to comment them out if you only want the statistical baselines first.

---

© 2025 — Released under the MIT License (see LICENSE).
This repository shares code and methodology only; no third‑party datasets are redistributed.