data-quality-checker

---
name: data-quality-checker
description: Use when validating market analysis documents or blog articles before publication for price-scale, notation, date/weekday, allocation-total, and unit errors. Advisory only; English and Japanese.
category: Risk
version: 1.0.0
tools: []
---

# Data Quality Checker — Market Document Validator

Before publishing a market analysis document (weekly strategy blog, market summary, translated report), scan its text directly for five categories of data-quality issues and flag them as advisory findings for human review — never as a hard block on publication.

## Categories checked

1. **Price scale consistency** — ETF vs. futures/index confusion. Use a digit-count heuristic (digits before the decimal point) rather than absolute price ranges, since it stays valid as prices drift over time while still catching scale confusion (e.g. "GLD: $2,800" has 4 digits when 2–3 is expected for that instrument class).
2. **Instrument notation** — ticker naming conventions per instrument class (equities, ETFs, futures, indices); flag inconsistent or ambiguous notation.
3. **Date / day-of-week accuracy** — compute the actual weekday for every stated date and flag mismatches (e.g. "January 1, 2026 (Monday)" when the actual weekday is Thursday). High-confidence — flag as ERROR.
4. **Allocation totals** — only check percentages inside allocation-labeled sections (headings like "Allocation", "配分", or columns like "Weight", "ウェイト", "目安比率"); ignore percentages in body text (probability, RSI, YoY growth). Flag totals more than 0.5% off from 100%.
5. **Unit mismatches** — flag inconsistent basis-point vs. percentage usage; note as informational when it may be intentional.

## Bilingual support

Handle both English and Japanese date formats, weekday names, and section headings. Normalize full-width characters (％, 〜, en-dash) before checking.

## Year inference

When a date has no explicit year, infer it in priority order: an as-of reference date the user has given, a YYYY pattern found in the document's title/metadata, or the current year with a 6-month cross-year heuristic (a stated month more than 6 months in the past relative to today implies next year; otherwise this year).

## Severity levels

- **ERROR** — high-confidence issues verified by direct computation (date-weekday mismatches). Recommend correction.
- **WARNING** — issues needing human judgment (price-scale anomalies, notation inconsistencies, allocation sums off by more than 0.5%).
- **INFO** — informational notes (e.g. mixed bp/% usage that may be intentional).

## Workflow

1. Read the full document text provided in chat.
2. Walk it section by section, applying the five checks above.
3. For each finding, note severity, category, the specific issue, and the passage it came from.
4. Present findings grouped by severity, plus a one-line total count.
5. Suggest a specific correction for each finding — propose the fix, don't just flag it.

## Key principles

1. **Advisory only** — every finding is for human review, never a block on publication.
2. **Section-aware allocation checking** — only allocation-labeled sections are checked for total percentage; random percentages in prose are ignored.
3. **Digit-count heuristic for price scale** — resilient to price changes over time, unlike fixed price-range checks.
4. **Especially valuable when combining data from multiple sources** (e.g. FRED, market data feeds, screening tools) into one report — cross-source merges are exactly where scale and unit mismatches creep in.