[sijie-Z]

P0. Third of the four research-integrity defects.

What was wrong

The engine is handed train_start - 252 .. test_end so its first rebalance has a warm-up — that part is right. But result["daily_returns"] covers that whole range and was appended as the fold's OOS in full:

oos_ret = result["daily_returns"]
oos_returns_list.append(oos_ret)
m = all_metrics(oos_ret, oos_bench)

Measured on 900 dates × 30 assets (train 252, test 63):

	fold 0 oos_days
before	315
after	63

So a 63-day test window was reported over 315 days, 252 of which were in-sample. Every fold's metrics, the concatenated OOS series, the aggregate metrics and the stability analysis were all computed over train + test.

This is a reporting-window defect, not signal leakage

The fold signal is recomputed from train-only ICs, and that part is correct — the audit agent and I verified it independently.

Recording this as a violation of NO_LOOKAHEAD_CONTRACT 第四条 would have been wrong: that clause governs signal construction, and signal construction was never the problem. The distinction matters, because calling it "walk-forward has lookahead" would both overstate it and hide where the real defect is.

Also fixed, one line away

fold_details["oos_days"] was len(fold_signal) — the signal length, i.e. train + test — rather than the length of the OOS returns actually being reported. Same confusion, different field.

Tests — both checked against the old implementation and both failing there

• test_each_fold_reports_only_its_test_window
• test_concatenated_oos_is_not_the_whole_sample

The second one exists because a per-fold check alone would not catch a regression that kept the per-fold slice but concatenated the unsliced series.

1320 passed, 1 skipped.

Impact

Every walk-forward report produced so far — main.py walkforward, the Strategy Gates input, and the front-end WalkForward panel — reported OOS metrics that were mostly in-sample. Those results need re-running and should be marked affected rather than deleted.

🤖 Generated with Claude Code

[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: Advanced

Run ID: ac41c6c2-9d79-457d-87dc-7c6c00447f61

---

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.}