Publication and updates
The Quarto website uses the same publication pattern as player_value and side_order. Documentation and saved results are built on pushes to main and published only after local-equivalent tests and rendering checks pass. Pull requests build and check the site without deploying it. A manual workflow trigger can rebuild a publication.
The Monthly forecast refresh action runs on the 2nd of each month at 08:23 UTC, and can also be started manually from GitHub Actions. It fetches current sources, runs tests, issues new dated 6-, 12- and 24-month forecasts, validates the website, commits the results, and passes that exact commit to the website workflow. Schedules may be delayed by GitHub.
Reproduce the website
Install the package’s development dependencies and Quarto, then run:
make check
make siteCI uses Quarto 1.10.20. Set QUARTO=/path/to/quarto to use another installed executable. make site reads the included data, validates forecast dates and model-selection metadata, generates tables, renders the site and checks local links and download checksums. It does not fetch new data or refit models.
The overview includes README.md; backtest pages render the saved analysis reports. Forecast tables join saved predictions to features reconstructed at exactly the saved cutoff, using one-to-one player IDs. Publication metadata in out/publication.json records input hashes and the forecast horizon and dates. Generated include files and rendered HTML are ignored by Git.
Data cadence versus documentation cadence
make refresh fetches the official IPL, Test, ODI and T20I Cricsheet archives, the side_order calendar, and ipl_auction identity aliases. Archive extraction uses the validated numeric-file pattern in player_value’s downloader. Upstream Git commits, downloaded-file hashes, archive hashes and retrieval times are saved. The auction-affiliation reference is a versioned snapshot of the local sibling data; that source file is not available in its GitHub checkout. It is not guessed or silently replaced during refreshes.
The observation endpoint is the validated side_order calendar’s cutoff, which must be a completed day within the last ten days. The forecast starts the next day. Thus a scheduled run can publish a forecast dated a few days earlier than the run. The workflow stops if a source fails, is stale, or loses previously recorded matches. Failed runs do not publish a new site.
New inputs live in data/current/, predictions in out/current/, and dated snapshots in out/history/. The forecast history includes downloads and source receipts. The original data/build/ and out/ backtests remain the frozen benchmark. The validation-selected model family stays fixed; parameters use only training labels that matured before January of the forecast year. Refreshing observations does not change the prior model-selection decision or turn the current forecasts into a new backtest.
The evaluated forecast grid is monthly. Newly issued forecasts use the available calendar cutoff, so their exact dates need not be month starts. The site separates these unscored forecasts from the frozen backtest. Rebuilding docs alone, via make site, does not fetch or advance data.
The acquisition command uses curl and GitHub CLI (gh), both available on the GitHub-hosted runner. A verified downloaded source bundle can be reused locally with python -m national_selection.refresh --workspace /path/to/sources.