Reproduce the paper

Everything the paper reports is regenerated by notebooks under publication/2_journal_article/notebooks/, which call the code in publication/2_journal_article/paper/. The reference run is committed under publication/2_journal_article/results/ with a manifest per stage (repository-relative paths and SHA-256 hashes), so a re-run can be checked file by file.

Set up

python -m pip install -e .[dev,paper]

Checking a re-run

tests/test_paper_results.py verifies that every file named in results/manifests/ is on disk with the recorded hash and that the M9 headline table equals the committed one:

pytest tests/test_paper_results.py -q

Show details

The figures and tables (minutes)

Two notebooks read results/ and write the manuscript's figures and tables to results/paper/figures/ (PNG at 600 dpi and PDF) and results/paper/tables/ (CSV and Markdown), and the supplementary ones to results/paper/supplementary/:

notebook	writes
paper_figures.ipynb	one figure per cell, each show("fig05_two_hypotheses") from the registry in paper/figures.py
paper_tables.ipynb	one table per cell, each show("table02_headline") from paper/tables.py

The manuscript entries (fig01 to fig12, table01 and table02) live in paper/manuscript.py, drawn at the journal's column widths; adding one later is one function in the registry and one cell.

The reference run (hours)

Five notebooks regenerate results/ in order; each has CONFIG in its first code cell and states its inputs, outputs and runtime in its header.

notebook	stage	writes	runtime
01_data_and_folds	population and the 18 station-held-out folds	results/01_data_folds/	seconds
02_baselines_m7_m8	M7 rule, M8 trained per fold (nine fits) and predicted	results/02_baselines/	about 20 minutes
03_m9	M9 scored and calibrated per fold	results/03_m9/	about two minutes
04_outcomes_and_metrics	reference terms, metrics, bootstrap, gate, operating points	results/04_metrics/	minutes
05_gamma_forecasting	the forecasting case study	results/05_gamma/	minutes

The configuration is publication/2_journal_article/config/evaluation.yaml; it records the dataset hashes, the fold rule, the settings of every method, the control and the gate. The evaluation refuses to run on a dataset whose hash differs.

The M8 bundles (results/02_baselines/bundles/) are not committed; they are regenerated by notebook 02 and their hashes recorded in the fold manifests.