market-data-sanity-checker

Trust layer for market data. Validate, score and explain anomalies in OHLCV, dividends, splits and fundamentals — before they break your screener, dashboard or tracker.

• Quickstart
• Why this exists
• Install
• Usage
• The sanity score
• Rule catalog
• When NOT to use it
• FAQ
• Development & contributing

Data quality / validation for market data, as a library and CLI. Not another generic data validator: this tool understands the semantics of financial data — corporate actions, adjustments, currencies, exchanges — and tells you what is suspicious, why, what impact it has, and whether to block, flag or review the datum.

A real mdsc check on MOB.ST: the phantom −50% "crash" is an unadjusted 2:1 split — detected, scored, and explained with the volume evidence. One self-contained .html file — works offline, interactive chart, no server.

Quickstart (30 seconds)

No install, no API key — audit a live symbol straight from Yahoo Finance:

npx market-data-sanity-checker check --provider yahoo --symbol AAPL

AAPL (yahoo) — sanity_score: 95/100
findings: 0 critical · 0 warning · 1 info

Or point it at your own data (CSV/JSON — headers are auto-mapped) and get the interactive HTML dashboard:

npx market-data-sanity-checker check --file your-data.csv --html report.html

That's the whole loop: score (0–100), findings (what's suspicious, why, with evidence), action (block / flag / review). See docs/examples.md for real caught-in-the-wild cases — the phantom −50% split crash, the $185 Berkshire tick, pence/pounds ×100 — each reproducible from the shipped fixtures.

Why this exists

Market data is wrong more often than you think:

• June 3, 2024 — a SIP software bug showed Berkshire Hathaway (BRK.A) at $185 instead of ~$621,000 (−99.9%). Trades executed on the bad price cost Interactive Brokers $48M.
• July 3, 2017 — Nasdaq test data propagated to Google Finance, Bloomberg and Yahoo, showing AAPL, GOOG and MSFT all at $123.47.
• Yahoo Finance regularly ships unadjusted splits that look like −50% crashes, dividends scaled 100× (the classic LSE pence/pounds confusion), and duplicated dividends.
• Tickers get reused (AB has named 6 different companies), silently splicing two firms into one price series.

Generic data quality tools (Great Expectations, Soda, Pandera) check structure — nulls, ranges, schemas. Data vendors (yfinance, ccxt) fetch data but don't audit it. Nothing open source sits in between. mdsc is that missing layer.

Principles

1. Causal explanation, not binary checks. Every finding carries a hypothesis ("possible unadjusted 2:1 split"), the evidence behind it, and a recommended action.
2. Flag, don't delete. This tool never rewrites your data. It detects, explains and scores — the decision stays with you.
3. Conservative thresholds. Every threshold comes with a citation (academic literature or a documented real-world incident) and is configurable. False negatives are acceptable; false positives destroy trust.
4. Provider-agnostic. The core works on CSV/JSON with a canonical internal schema. Provider connectors are plugins, never core.

Install

npm install market-data-sanity-checker   # or: pnpm add market-data-sanity-checker

Usage

CLI

# check a CSV of OHLCV bars (date,open,high,low,close,volume — headers are auto-mapped)
mdsc check --file aapl.csv --symbol AAPL --source yahoo

# fetch and check straight from a provider (no API key needed for Yahoo)
mdsc check --provider yahoo --symbol AAPL --range 1y

# compare two sources for the same symbol
mdsc compare --symbol AAPL --providers yahoo,alpha-vantage
mdsc compare --files yahoo-aapl.json,stooq-aapl.csv
# (PowerShell: quote values with commas — "yahoo,alpha-vantage" — or the shell swallows the comma)

# machine-readable report
mdsc check --file aapl.json --format json > report.json

# provider status and rule catalog
mdsc providers
mdsc rules

Exit codes: 0 gate passed · 1 gate failed (critical findings, or --fail-on/--min-score) · 2 operational error (file unreadable, invalid flags, provider unavailable, invalid config).

Providers are plugins: yahoo works keyless; alpha-vantage needs the ALPHA_VANTAGE_API_KEY env var (free key at alphavantage.co). New providers plug in via the Connector contract (src/connectors/types.ts).

Configuration (mdsc.config.json)

Every threshold is project-tunable without touching code — drop a mdsc.config.json in your working directory (or pass --config <path>). Unknown rule ids are a hard error with the valid catalog:

{
  "rules": {
    "RETURN_SPIKE": { "params": { "zscoreThreshold": 5 } },
    "STALE_PRICE": { "enabled": false },
    "SPLIT_NOT_ADJUSTED": { "severity": "warning" }
  },
  "compareRules": {
    "VOLUME_DIVERGENCE": { "enabled": false }
  }
}

CI gating, history and dashboards

# gate your pipeline: exit 1 on warnings too, or when the score drops
mdsc check --provider yahoo --symbol AAPL --fail-on warning --min-score 80

# keep an audit trail and see the trend
mdsc check --provider yahoo --symbol AAPL --save
mdsc history --symbol AAPL

# self-contained HTML dashboard (offline, interactive — great CI artifact)
mdsc check --file aapl.csv --html report.html

SDK

import { ingestFile, marketDataSetSchema, checkMarketData } from 'market-data-sanity-checker'

const raw = await ingestFile('aapl.csv', { symbol: 'AAPL', source: 'yahoo' })
const dataset = marketDataSetSchema.parse(raw)
const report = checkMarketData(dataset)

console.log(report.sanity_score) // 0–100
for (const finding of report.findings) {
  console.log(finding.severity, finding.rule, finding.explanation)
}

Fetch and compare sources with the connector plugins:

import { yahoo, alphaVantage, compareDatasets } from 'market-data-sanity-checker'

const a = await yahoo.fetchDaily('AAPL', { range: '1y' })
const b = await alphaVantage.fetchDaily('AAPL') // needs ALPHA_VANTAGE_API_KEY
const comparison = compareDatasets(a, b)

console.log(comparison.consistency_score, comparison.compared_dates)

Every threshold is a parameter — rules can be tuned, re-severitized or disabled:

const report = checkMarketData(dataset, {
  rules: {
    RETURN_SPIKE: { params: { zscoreThreshold: 5 } },
    STALE_PRICE: { enabled: false },
    SPLIT_NOT_ADJUSTED: { severity: 'warning' },
  },
})

Report shape

{
  "symbol": "MOB.ST",
  "source": "yahoo",
  "sanity_score": 60,
  "findings": [
    {
      "rule": "SPLIT_NOT_ADJUSTED",
      "severity": "critical",
      "action": "block",
      "dimension": "accuracy",
      "where": { "date": "2023-04-14" },
      "explanation": "Price moved -50.0% in one session, matching an unadjusted 2:1 split, but no split is registered near this date. Hypothesis: the vendor recorded the split without adjusting historical prices. Supporting evidence: volume moved 2.0× in the opposite direction, as unadjusted feeds typically do.",
      "evidence": { "return": -0.5, "hypothesized_ratio": "2:1", "previous_close": 99.4, "close": 49.7, "volume_ratio": 2 },
      "references": ["https://ranaroussi.github.io/yfinance/advanced/price_repair.html"],
      "occurrences": 1
    }
  ],
  "summary": { "critical": 1, "warning": 0, "info": 0 },
  "generated_at": "2026-07-31T17:00:00.000Z"
}

The sanity score

sanity_score = max(0, 100 − Σ penalties). Each finding penalizes by severity: critical −40, warning −15, info −5. A rule penalizes once per dataset, at its highest severity — 847 bars with the same defect must not drive the score below zero; the finding carries occurrences instead.

Score	Band
90–100	reliable
70–89	usable with caution
40–69	suspicious
<40	not fit for production

Rule catalog

The corpus ships 36 rules: 29 single-source across four blocks, plus 7 multi-source compare rules. Full details with defaults and references: mdsc rules (or --format json). Real-case fixtures anchor the headline rules: tests/fixtures/, documented in docs/fixtures.md.

Price / OHLCV (13)

Rule	Severity	What it catches
SPLIT_NOT_ADJUSTED ⭐	critical	Price jump matching a split ratio with no split registered (phantom −50% "crashes")
PRICE_SPIKE_INTRADAY	critical	Isolated bad ticks: far outside the neighbourhood AND reverted next session (Berkshire 2024)
PRICE_NONPOSITIVE	critical	Zero/negative/NaN prices
OHLC_INCONSISTENT	critical	high < low, open/close outside [low, high]
VOLUME_NEGATIVE	critical	Negative or non-finite volume
TS_DUPLICATED	critical	Two bars for the same timestamp
RETURN_SPIKE	warning	Robust statistical outlier returns (modified z-score), deduplicated against splits
CURRENCY_SCALE_SUSPECT	warning	Persistent ~100× level shifts — pence/pounds (GBX/GBP) block confusion
STALE_PRICE	warning	Identical close for 3+ sessions with volume — frozen feed
BAR_MISSING	warning	Gaps of a full trading week or more
TS_UNORDERED	warning	Bars out of chronological order
ZERO_VOLUME_MOVED	warning	Zero volume but price moved intraday
INSUFFICIENT_DATA	warning	Too few bars to evaluate plausibility — the score reflects absence of evidence

Corporate actions (9)

Rule	Severity	What it catches
EXDATE_AFTER_PAYDATE	warning	Payment date preceding ex-date (field swap, unless a large special dividend mandates it)
DIV_DUPLICATED	warning	Same dividend recorded twice within days (ALC.SW 2023)
EXDATE_MISPLACED	warning	No price drop on the registered ex-date, matching drop days later (TETY.ST)
DIV_NOT_ADJUSTED	warning	Dividend registered but pre-ex-date prices unadjusted (8TRA.DE, 1398.HK)
DIV_SCALE_100X	warning	Single dividend implausibly large — ×100 pence/pounds error (HLCL.L, LTI.L)
DIV_YIELD_IMPOSSIBLE	warning	TTM dividend yield above the 20% plausibility bound
DIV_SPECIAL_MISCLASSIFIED	info	"Regular" dividend that is an extreme outlier vs the issuer's own history
SPLIT_RATIO_IMPROBABLE	warning / info	Split ratio extreme or 1:1; near-1 ratios flagged as probable spin-offs (info)
CORPORATE_ACTION_MISSING_FROM_FACTOR	warning	Prices adjusted (factor < 1) with no dividend/split registered to explain it

Fundamentals (4)

Rule	Severity	What it catches
FUNDAMENTALS_SIGN_VALIDITY	critical	marketCap or sharesOutstanding non-positive or non-finite
MARKETCAP_MISMATCH	warning	marketCap ≠ sharesOutstanding × price (Alphabet case)
PE_EPS_INCOMPATIBLE	warning	P/E incompatible with EPS and price
PAYOUT_IMPOSSIBLE	warning	Payout ratio negative or far above 300% of EPS (REIT-aware)

Metadata (3)

Rule	Severity	What it catches
CURRENCY_SUSPECT	warning / info	GBP label on pence-looking prices; currency absent (info)
SYMBOL_MAPPING_SUSPECT	warning	ISIN checksum invalid; ISIN country inconsistent with exchange (ticker reuse)
DIVIDEND_FX_MISMATCH	warning	Dividend currency differs from the series currency (Shell pays USD, trades GBX)

Compare (multi-source, 7)

Rule	Severity	What it catches
SYMBOL_MISMATCH	critical	The two datasets are for different symbols — comparison is meaningless
INSUFFICIENT_OVERLAP	warning	Too few shared dates between sources — agreement cannot be evaluated
CLOSE_DIVERGENCE	warning	Systematic close disagreement between two sources
DIVIDEND_MISMATCH	warning	Dividend missing or different in the other source
SPLIT_MISMATCH	warning	Split missing or with a different ratio in the other source (phantom-crash factory)
PRICE_DATE_MISMATCH	info	Specific dates where the sources disagree beyond tolerance
VOLUME_DIVERGENCE	info	Systematic volume disagreement (volumes are not cross-comparable)

Findings carry rule, severity (critical/warning/info), action (block/flag/review), dimension (DAMA data quality dimension), a human-readable explanation, evidence, and references to the incident or literature justifying the rule.

When NOT to use it

• It does not repair data. mdsc flags, never rewrites — flag, don't delete. If you need automatic price repair (yfinance-style), do it downstream, on your own call.
• It is not for realtime or tick data. The unit of analysis is the daily bar plus corporate actions; intraday streams and tick-level feeds are out of scope.
• It does not cover options or futures corporate actions. The semantics are modeled for equities/ETFs; derivative chains, rolls and expiries are not checked.
• It is not a substitute for paid point-in-time data. It catches vendor defects, but survivorship-bias-free, restatement-aware datasets remain a paid-vendor feature.

FAQ

Why does Yahoo Finance show −50% crashes that never happened — and does mdsc catch them? Those are unadjusted splits: the vendor records the split but doesn't adjust historical prices, so a 2:1 split looks like a −50% crash. SPLIT_NOT_ADJUSTED (critical) matches the jump against plausible split ratios and cross-checks the volume signature; the compare rule SPLIT_MISMATCH catches sources that disagree on the split itself.

What is the sanity_score? A 0–100 quality score per dataset: max(0, 100 − Σ penalties) with −40/−15/−5 per critical/warning/info finding, each rule penalizing once per dataset. 90+ reliable, 70–89 usable with caution, 40–69 suspicious, below 40 not fit for production. See The sanity score.

Does mdsc fix my data? No — flag, don't delete. It detects, explains and scores anomalies, and recommends an action (block/flag/review); the decision and any repair stay with you.

Which data sources are supported? Any CSV/JSON you can map to the canonical schema (mdsc check --file), Yahoo Finance keyless (--provider yahoo), and Alpha Vantage with a free API key. New providers plug in via the Connector contract.

Can I gate CI on data quality? Yes: mdsc check --provider yahoo --symbol AAPL --fail-on warning --min-score 80 exits 1 when the gate fails (0 pass, 2 operational error) — drop it into any pipeline.

How is this different from Great Expectations or Pandera? Those are generic, structure-first validators (nulls, ranges, schemas) you configure yourself. mdsc is domain-first: it ships 36 rules that understand market-data semantics — splits, dividends, pence/pounds, ticker reuse — each with a causal explanation, evidence, a literature or incident reference, and a tunable severity, plus multi-source comparison.

Roadmap

• Phase 0 — canonical schema, CSV/JSON ingestion, scoring model, CLI, CI
• Phase 1 — OHLCV rules engine (12 rules: unadjusted splits, bad ticks, stale prices, currency scale suspects…), per-rule config, mdsc rules → v0.1.0
• Phase 2 — corporate actions, fundamentals and metadata (13 rules), identifiers schema (ISIN/CUSIP/FIGI), golden report-contract tests → v0.2.0
• Phase 3 — provider connectors (Yahoo keyless, Alpha Vantage), comparison engine (mdsc compare, 5 compare rules), provider scoreboard, real-data calibration → v0.3.0
• Phase 4 — mdsc.config.json custom rules, CI gating (--fail-on/--min-score), audit history (mdsc history), self-contained HTML dashboards, OKF datapackage → v1.0.0
• Phase 5+ (if traction) — user rule plugins, hosted API, more connectors, web app

Provider scoreboard

The project audits itself against reality, publicly: scoreboard/ holds the latest weekly audit of ~30 liquid symbols per provider (mean sanity score, worst symbols, findings), regenerated every Monday by a GitHub Action and reproducible by anyone with pnpm scoreboard. calibration/ holds the latest 50-symbol calibration run — the corpus's honesty loop against real Yahoo data.

Development

pnpm install
pnpm test         # vitest
pnpm typecheck    # tsc --noEmit
pnpm lint         # eslint
pnpm build        # tsup → dist/
pnpm dev check --file tests/fixtures/ohlcv-valid.csv

Contributing

See CONTRIBUTING.md. The short version: every rule needs a stable ID, a severity, a documented real-world case or literature reference, and a fixture.

License

MIT