Research data integrity boundaries

中文页面

This page defines the market-data platform's responsibilities for preventing look-ahead bias, survivorship bias, and research-sample contamination. Model selection, combinatorial purged cross-validation (CPCV), the Deflated Sharpe Ratio (DSR), probability of backtest overfitting (PBO), feature ablation, and candidate promotion belong to downstream research repositories. Execution dry runs, paper/live gates, and order audits belong to the execution repository.

Platform controls

Risk	Platform control	Main evidence
Data-version drift	Current data contract fixes the downstream entry point	metadata/current_assets/a_share_current.json
Ambiguous asset scope	Manifest and registry record paths, date range, row count, and lineage	manifest.yml, metadata/dataset_registry.csv
Current constituents backfilled into history	By-date universe is the research-universe entry point	universe_by_date, universe_meta
Latest financial snapshot used in historical research	Raw → normalized → PIT layers publish by disclosure and availability dates	normalized_fundamentals, pit_fundamentals
Current industry labels backfilled into history	Historical industry assets require effective intervals or dated-snapshot provenance	industry_changes
Daily-asset quality gaps	Baseline and research-profile checks record missing/duplicate rows, price limits, suspensions, and calendar issues	reports/a_share_daily_clean_*_validation_*.json
Insufficient publication evidence	Validation and health checks precede current-contract publication	reports/a_share_current_release_*.json, reports/a_share_current_health_*.json

These controls ensure downstream research reads versioned assets with lineage and quality evidence. The platform does not assess strategy overfitting or read a research run's Sharpe, IC, or backtest curve.

A-share research assets

The authoritative current A-share contract is metadata/current_assets/a_share_current.json. Consumers should verify that:

• daily_clean passes baseline checks; formal research profiles also require research checks.
• universe_by_date comes from the platform-published by-date universe, with research settings research_universe.mode: pit and require_by_date: true.
• daily_basic valuation fields are a daily valuation overlay. PIT financial-statement research uses pit_fundamentals.
• is_st from the latest instruments snapshot is only a non-PIT label.
• PIT fundamentals v2 available_date is generated from disclosure date with a calendar-day delay. Research must use the formal as-of loader and validate revision_covered, freshness_verified, field-level revision provenance, and _source_retrieved_at <= as_of_date. Report-period query ranges do not prove observation availability. Legacy v1 and historical backfills collected in 2026 cannot claim revision-safe PIT semantics.
• Historical industry assets include effective_date / end_date or explicit dated-snapshot provenance.

See A-share research profile for the full asset policy and A-share fundamentals for raw-to-PIT operations.

Pre-publication checks

Before publishing or switching the current A-share contract, record these steps:

1. Download raw data or ingest a licensed local extract.
2. Build normalized or clean assets.
3. Run baseline validation.
4. Run research validation for research use.
5. Check the current data contract's health.
6. Generate publication evidence.
7. Update metadata/current_assets/a_share_current.json and metadata/dataset_registry.csv.

Related references: contracts covers current contracts, asset keys, manifests, and the registry; A-share research profile covers PIT and historical-industry policy; A-share fundamentals covers raw-to-PIT ingestion and publication; TuShare operations covers routine refreshes.

Relationship to downstream overfitting controls

Downstream research controls should check that runs use the current platform contract and a by-date PIT universe; do not treat daily_basic as PIT fundamentals or backfill current industry / ST snapshots; and keep final out-of-sample (OOS) evaluation, CPCV, feature evidence, DSR, and promotion gates in the research layer.

The execution repository consumes research-exported targets.json. Broker logs, paper-trading evidence, and live audits are not part of the platform's current-data contract.