reproducibility.md

Reproducibility — how to verify the published numbers

Every study's headline numbers are produced from cached market-data snapshots and stamped with data fingerprints (short hashes of the exact series used), quoted in each study's docs/results*.md. The policy:

• Caches are gitignored. The _cache/*.parquet snapshots (~hundreds of MB, and some upstream licences forbid in-tree redistribution) are never committed — only their fingerprints are, inside the results docs.
• A release bundle carries the snapshots. python tools/make_repro_bundle.py collects every existing cache (root _cache/ plus each study's _cache/) into repro_bundle_<date>.zip, with a manifest.json listing path, size and sha256 per file. The bundle is attached to the matching GitHub Release, so a third party can drop the caches in place and reproduce the published fingerprints byte-for-byte. (--dry-run lists what would be bundled without writing anything.)
• Or rebuild from the vendor. Each study's examples/verify*.py --fetch re-downloads the data and re-runs the audit. Beware vendor drift: Yahoo occasionally restates history (splits, dividends, late corrections), so a fresh fetch can yield slightly different fingerprints and slightly different numbers. That is expected — the release bundle is the byte-exact reference; a fresh fetch checks that the conclusions survive today's data.

In short: same caches → same fingerprints → same numbers, to the decimal. Fresh data → possibly new fingerprints → the verdicts should still hold.