research: 최근 경기의 종목·리그별 확률 보정 격리 실험

[choigod1023]

Summary

• Add a standalone standard-library calibration experiment, fixed protocol, minimized sports-prediction fixture and reproducible report.
• Train on 76 earlier official-result-linked picks; evaluate on 103 later picks from Sept 6-7. Four fixed candidate offsets, fixed selections and odds.
• League/market calibration reduces Brier by about 1.37%; recency adds almost nothing. Day-cluster diagnostics and Holm correction do not establish significance. NPB worsens despite overall improvement.

Limits and safety

• Exploratory retrospective dates, previously inspected outcomes and unverified historical label availability. NOT a pristine future holdout or deployable backtest.
• No xG/starter effect, new selection, hit-rate or ROI improvement claimed. No runtime DB access, collector calls, operating weights, model allowlist, restarts or deployments changed.

Validation

• 781 Python tests and 58 subtests passed, including 23 new experiment regressions.
• 229 frontend tests passed; isolated production build passed.
• Saved report exactly reproduces from the committed dataset and protocol; staged diff check passed.

See experiments/recent-calibration-20260907/README.md for Korean findings and reproduction commands.

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Team

Run ID: b46f98ae-92c8-4f0f-9eb7-ce2120403b4e

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}