protostar-climate / code /README.md
AlecAlameddine's picture
Upload 105 files
d79a78d verified
|
Raw History Blame Contribute Delete
2.93 kB
# Code
Everything that produces `submission.csv`, as plain scripts (no notebook, no agent needed).
From the repository root: `make all`, or step by step with `PYTHONPATH=code`:
| Step | Command | Writes |
|---|---|---|
| verify / fetch pinned data | `python code/fetch_data.py` | `data/raw/` (sha256-checked) |
| data audit | `python -m protostar.audit` | `reports/audit/` |
| backtests + model selection | `python -m protostar.backtest` | `reports/backtest/` (incl. `selected.json`) |
| tried and not shipped | `python -m protostar.experiments` | `reports/experiments/` |
| refit ≤ 2018 + forecast | `python -m protostar.forecast` | `submission.csv`, `reports/forecast/` |
| file check (template) | `python code/check_submission.py` | — |
| tests | `python -m protostar.tests` | — |
| README tables | `python -m protostar.docs` | number tables in `README.md` |
| NOAA cross-check | `python -m protostar.ghcnd --offline` | `reports/ghcnd/` (from `data/external/ghcnd/`; without `--offline` missing stations are downloaded, capped at 2018-12-31) |
Modules in `protostar/`:
- `common.py`: constants: pinned dataset revision and sha256, metric definitions, the organizers' rules (10 missing days, 10 valid years, 30-year window), the backtest folds, the seed.
- `data.py`: loaders with the **leak guard** (`assert_no_leak`: nothing dated after 2018), the valid-year rule (`long_table`), city × year panels.
- `scoring.py`: the challenge's Score, as in the Space's `scoring.py`.
- `models.py`: the baselines and the candidate models, each a pure function of training panels that end at the origin; a model is a small dict, so backtest and final refit run the same code.
- `backtest.py`: folds, bootstrap, the selection ladder and its acceptance rule, report tables.
- `forecast.py`: refit on ≤ 2018, `submission.csv`, the Space's file checks, plausibility flags.
- `audit.py`: coverage, units, outliers, recomputation from daily data, station changes.
- `experiments.py`: headroom estimate, station-shift correction, coefficient shrinkage: what was tried and not shipped, judged by the backtest's rule.
- `ghcnd.py`: cross-check and gap-fill experiment against NOAA GHCN-Daily (downloads capped at 2018-12-31), and a check of the fallback for city–metrics without a baseline.
- `docs.py`: regenerates the number tables in the README from `reports/`.
- `tests.py`: leak guard, no peeking (raw data after each origin corrupted), neighbors from coordinates only, the selection reproduces, scoring equals the Space's, published practice-run numbers reproduced, model invariants, `submission.csv` equals the code's output, the fallback serves unscored rows only, the NOAA extract respects the date cap.
Other files: `requirements.txt` (exact versions), `check_submission.py` (from the template; reads the pinned local city list first), `AGENTS.md` (instructions for AI agents, including the data rule), `LICENSE` (MIT).