[sijie-Z]

P0. First of the four research-integrity defects. NO_LOOKAHEAD_CONTRACT 第三条 requires exactly this behaviour, and the point-in-time accessors already existed — neither was wired to the panel the factors actually read.

What was wrong

The daily financial panel was forward-filled from the fiscal period end, then backward-filled:

df_full = df.reindex(full_idx)              # df still keyed on qdate
df_full = df_full.groupby("asset").ffill()
df_full = df_full.groupby("asset").bfill()  # "fill initial NaN"

A-share reports land 40–50 days after the period closes, so this published figures weeks before anyone could have seen them. Measured on the provider's own output (seed 7, 000001.SH):

Date	vs publish_date	Value shown
2021-04-10	40 days before	visible	❌ should be unknown
2021-05-10	10 days before	visible	❌ should be unknown
2021-05-20	publication day	visible	✅

(fiscal_period_end 2021-03-31 → publish_date 2021-05-20, lag 50 days.)

The bfill() had a second effect: it pushed a company's first report back to the first date of the sample, so fundamentals were "known" months before anything was published.

Fix

Re-index each report to its publish_date before forward-filling, so a value first appears on the day it became public. Drop the bfill — before a company's first report there is genuinely nothing to know, and filling it backwards is itself a look-ahead.

The fiscal period moves to a new fiscal_period_end column rather than overloading the index, which now carries the visibility date.

Verified

before publish_date      -> nan
on / after publish_date  -> the published value
before the first report  -> nan

Two regression tests added, and both were checked against the old implementation — they fail there:

• test_no_value_is_visible_before_its_publish_date
• test_nothing_is_backfilled_before_the_first_report

The pre-existing test_publish_date_after_report_date asserted the old model (that index dates which happen to be quarter-ends are the original report rows). That assumption no longer holds, so it now compares publish_date against the new fiscal_period_end column — the same property, checked against the new shape.

1315 passed, 1 skipped.

Impact — this invalidates prior fundamental results

Everything computed through the daily panel is affected, so prior fundamental-factor ICs need re-running. That includes the M1 roe run and the M3 factor-zoo entries for roe / pb_ratio / pe_ratio / log_market_cap / asset_growth.

Per the contract's own remediation procedure, those results should be marked affected rather than deleted. The validity metadata to do that does not exist yet — tracked separately.

🤖 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: b64f7af8-d5b6-4406-adc9-19fc5c30bb8c

---

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.}