Paris air quality forecasting (X-HEC Time Series)

Forecast hourly CO, NO₂, O₃, PM10 and PM2.5 in Paris for the 504 hours (3 weeks) after the training data. The project compares baselines, zero-shot Chronos-2 and an LSTM.

• Deliverable notebook: notebooks/main_colab.ipynb, the run of record, executed end to end in Google Colab on a Tesla T4. notebooks/main.ipynb is the same notebook as run locally; it is generated by scripts/build_notebook.py.
• Report: report/report.md

Run locally

pip install -r requirements.txt
bash scripts/download_data.sh          # needs ~/.kaggle/kaggle.json or ~/.kaggle/access_token
python3 -m pytest tests/ -q            # 91 tests: metrics, no-leakage checks, submission format
jupyter nbconvert --to notebook --execute --inplace notebooks/main.ipynb --ExecutePreprocessor.timeout=-1

Run in Google Colab (the assignment asks for the foundation-model run there)

1. Copy this whole folder, including data/raw/, to Google Drive, e.g. MyDrive/class_time_serie.
2. Open notebooks/main.ipynb from Drive with Colab. Choose Runtime → Change runtime type → GPU.
3. If your folder is elsewhere, edit DRIVE_PROJECT_DIR in the first code cell.
4. Runtime → Run all. The first cell installs chronos-forecasting==2.3.2 and mounts Drive. The Chronos-2 cell then records the model, version and hardware for the report.

Layout

src/          data.py, metrics.py, features.py, validation.py, evaluation.py,
              models.py (baselines), foundation.py (Chronos-2), lstm.py, adjust.py, plots.py
tests/        unit tests (metrics at y = 0, leakage, windows, submission format)
scripts/      download, notebook builder, standalone experiment scripts
results/      leaderboard_local.csv, Kaggle scores, run info, predictions
figures/      EDA and results figures
submissions/  submission CSVs