Overview
Detect common data quality issues in market analysis documents before publication. The checker validates five categories: price scale consistency, instrument notation, date/weekday accuracy, allocation totals, and unit usage. All findings are advisory -- they flag potential issues for human review rather than blocking publication.
When to Use
- Before publishing a weekly strategy blog or market analysis report
- After generating automated market summaries
- When reviewing translated documents (English/Japanese) for data accuracy
- When combining data from multiple sources (FRED, FMP, FINVIZ) into one report
- As a pre-flight check for any document containing financial data
Prerequisites
- Python 3.9+
- No external API keys required
- No third-party Python packages required (uses only standard library)
Workflow
Step 1: Receive Input Document
Accept the target markdown file path and optional parameters:
--file: Path to the markdown document to validate (required)--checks: Comma-separated list of checks to run (optional; default: all)--as-of: Reference date for year inference in YYYY-MM-DD format (optional)--output-dir: Directory for report output (optional; default:reports/)
Step 2: Execute Validation Script
Run the data quality checker script:
python3 skills/data-quality-checker/scripts/check_data_quality.py \
--file path/to/document.md \
--output-dir reports/To run specific checks only:
python3 skills/data-quality-checker/scripts/check_data_quality.py \
--file path/to/document.md \
--checks price_scale,dates,allocationsTo provide a reference date for year inference (useful for documents without explicit year in dates):
python3 skills/data-quality-checker/scripts/check_data_quality.py \
--file path/to/document.md \
--as-of 2026-02-28Step 3: Load Reference Standards
Read the relevant reference documents to contextualize findings:
references/instrument_notation_standard.md-- Standard ticker notation, digit-count hints, and naming conventions for each instrument classreferences/common_data_errors.md-- Catalog of frequently observed errors including FRED data delays, ETF/futures scale confusion, holiday oversights, allocation total pitfalls, and unit confusion patterns
Use these references to explain findings and suggest corrections.
Step 4: Review Findings
Examine each finding in the output:
- ERROR -- High confidence issues (e.g., date-weekday mismatches verified by calendar computation). Strongly recommend correction.
- WARNING -- Likely issues that need human judgment (e.g., 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).
Step 5: Generate Quality Report
The script produces two output files:
- JSON report (
data_quality_YYYY-MM-DD_HHMMSS.json): Machine-readable list of findings with severity, category, message, line number, and context. - Markdown report (
data_quality_YYYY-MM-DD_HHMMSS.md): Human-readable report grouped by severity level.
Present the findings to the user with explanations referencing the knowledge base. Suggest specific corrections for each issue.
Output Format
JSON Finding Structure
{
"severity": "WARNING",
"category": "price_scale",
"message": "GLD: $2,800 has 4 digits (expected 2-3 digits)",
"line_number": 5,
"context": "GLD: $2,800"
}Markdown Report Structure
# Data Quality Report
**Source:** path/to/document.md
**Generated:** 2026-02-28 14:30:00
**Total findings:** 3
## ERROR (1)
- **[dates]** (line 12): Date-weekday mismatch: January 1, 2026 (Monday) -- actual weekday is Thursday
## WARNING (2)
- **[price_scale]** (line 5): GLD: $2,800 has 4 digits (expected 2-3 digits)
> `GLD: $2,800`
- **[allocations]**: Allocation total: 110.0% (expected ~100%)Resources
scripts/check_data_quality.py-- Main validation scriptreferences/instrument_notation_standard.md-- Notation and price scale referencereferences/common_data_errors.md-- Common error patterns and prevention
Key Principles
Advisory mode: All findings are warnings for human review. The script always exits with code 0 on successful execution, even when findings are present. Exit code 1 is reserved for script failures (file not found, parse errors).
Section-aware allocation checking: Only percentages within allocation sections (identified by headings like "配分", "Allocation", or table columns like "ウェイト", "目安比率") are checked. Random percentages in body text (probability, RSI, YoY growth) are ignored.
Bilingual support: Handles both English and Japanese date formats, weekday names, and section headings. Full-width characters (%, 〜, en-dash) are normalized before processing.
Year inference: For dates without an explicit year, the checker infers the year using (in priority order): the
--as-ofoption, a YYYY pattern found in the document title/metadata, or the current year with a 6-month cross-year heuristic.Digit-count heuristic: Price scale validation uses digit counts (number of digits before the decimal point) rather than absolute price ranges. This approach is resilient to price changes over time while still catching ETF/futures confusion errors.
Supported weekday checks
The checker accepts YYYY-MM-DD, YYYY/MM/DD, M/D and Japanese
M月D日 dates followed by a weekday. Numeric dates accept English three-letter
or full names (case-insensitive), or Japanese 火 / 火曜日 style labels.
Use matching ASCII (Tue) or full-width (火) parentheses, optionally
preceded by spaces or tabs. Japanese weekday labels also accept mixed-width
parentheses for compatibility. Dates accept Unicode decimal digits, including
full-width digits. Tight ranges such as 11/10(火)-11/12(水) and
12/9-12/10(金) check every dated endpoint. A year prefix written as
2025-11/10(水) is ambiguous, so the weekday is reported only when it is
wrong for both that year and the inferred year. Invalid calendar dates and unrecognized labels
are outside this weekday check and are skipped, not reported as valid dates.
An explicit numeric year always wins. Yearless forms reuse --as-of year
inference (including the existing six-month rollover), then document year,
filename year, or current year. Supply --as-of for reproducible checks.
Every computed mismatch is an advisory ERROR, including Japanese forms
that previously produced WARNING. Findings still leave the CLI exit status 0.