Data — Providers & Quality Gates

The single most expensive failure mode in quant research is silent bad data, so every ETL pass runs a quality report and the findings are persisted next to the artefacts.

Providers — alphaforge.data.providers

Provider	Status
sample	ships; fully synthetic but point-in-time (carries a survivorship disclaimer)
local	reads a local long price table
yahoo	reserved adapter (historical OHLCV)
akshare	reserved adapter (A-share market data)
tushare	reserved — wiring point exists, no live adapter ships (supply token + fetch)

DataPipeline persists the canonical long table; build_panel pivots it into the aligned wide MarketPanel (close / returns / volume / market_cap / industry / universe) that every downstream layer consumes.

Quality — alphaforge.data.quality

from alphaforge.data.quality import build_quality_report, clean_prices

report = build_quality_report(prices, provenance="sample")
assert report.is_acceptable()        # worst missingness + no duplicates
clean  = clean_prices(prices, drop_duplicates=True, drop_non_positive=True)

DataQualityReport records coverage, missingness, outliers (robust MAD on the log scale), OHLC violations, stale-price runs, extreme returns and delisting candidates — and emits a survivorship note for names that stop trading before the panel end (they are retained, never silently dropped). Cleaning primitives (clean_prices, align_calendar, detect_missing_blocks) are deterministic and flag, never overwrite, implausible moves.

See tests/unit/test_data_quality.py for the contract.