Financebotresearch desk研究台

Playbook › Playbook

Any figure carrying the verdict must come from the filing

static 2026-07-29

Claim

Not a general counsel of perfection — a targeted rule. Vendor data is fine for screening breadth. It is not fine for the two or three numbers that actually decide the verdict, because those are precisely the numbers that have been caught wrong.

The failures are not random noise; they cluster in the hardest, most judgment-laden lines — non-operating income, cash-flow derivations, and anything requiring a classification decision.

Track record

Vendor Failure Note
roic.ai FCF returned as OCF, capex null [[pitfall-roicai-fcf-field-returns-ocf]]
Yahoo Strategic-investment marks filed as Interest Income [[pitfall-yahoo-misclassifies-strategic-investment-gains]]
Both ENTG FY2025 margins corrupted in both feeds simultaneously [[pitfall-entg-margins-corrupted-in-vendor-feeds]]
fin.py Broken/negative EV on foreign ADRs [[pitfall-adr-ticker-resolution]]

The ENTG case is the important one: cross-validating two vendors did not catch it, because both were wrong the same way. Cross-validation between feeds is a weaker control than it appears; only the filing is independent.

Cheap sanity checks that caught real errors

  • Divide interest income by cash balance — anything above the risk-free rate is misfiled.
  • Recompute FCF = OCF − capex by hand; if capex is null, the FCF field is unusable.
  • Read the EPS bridge in the press release — companies usually name their own adjustments outright, and that sentence resolves most non-operating questions in one line.
  • Any margin that looks category-leading by a wide margin is more likely a data error than an edge.

What would falsify this

A sustained run of vendor figures matching filings across the verdict-bearing lines. The rule costs one document fetch per analysis, so the bar for dropping it is high.

History

  • 2026-07-29 — established after three independent vendor failures inside one week, two of them on the same analysis.