"Equity Liquidity-Shock Research — Data Audit (Step 2)"

Purpose: find temporary stock-level mispricing created by liquidity/order-flow shocks. This document is the mandatory pre-implementation audit. No strategy exists yet; this only records what data is actually available and reliable.

#Verdict up front

Conclusion: true order flow exists but is a single-name, 22-day sample. Every broad panel is OHLCV-only. Per the brief this means two separately labelled tracks, never conflated:

• Track A — OHLCV liquidity-pressure PROXY (broad: 499 sym × 3y, 30-min; plus 100 sym × 1y, 5-min). Not order flow. Factors may only use price, volume, dollar-volume.
• Track B — TRUE microstructure (AAPL 1-second NBBO/trades: spread, size imbalance, Cont OFI, Lee-Ready signed volume). Single name, 22 days → underpowered by construction; used to validate/reject Track A's proxy, not to claim a cross-sectional edge.

#Dataset inventory

#True order flow — data/parquet_micro/{SYM}_{YYYY-MM-DD}_1s.parquet

• Source: data/equity_ticks.py — Alpaca SIP historical trades + NBBO.
• Cached: AAPL only, 22 files, 2026-08-03 … 2026-09-15 (non-contiguous).
• Grid: 1-second RTH (09:30:00–15:59:59 ET), 23,400 rows/day.
• Columns: mid bid ask bsize asize spread imbalance ofi nq vol sv tcount notional micro vwap.
  • spread = ask−bid (mean/sec), imbalance = (bsize−asize)/(bsize+asize),
  • ofi = Cont et al. event-level OFI summed per second,
  • sv = signed volume (Lee-Ready vs prevailing NBBO, tick-rule fallback),
  • micro = size-weighted microprice.
• Raw density ~670k quotes + ~660k trades per liquid symbol-day; pull ~2.3 min/symbol-day. Fails closed on 401/403.
• No Level-2 depth. No continuous multi-level book. No historical deep book.

#OHLCV bars — data/parquet/daybot/

#Daily adjusted — data/parquet/{SYM}_ADJ.parquet

• 503 symbols, ~6.1y (2020-07 → 2026-09), open high low close adj_close volume dividend split. Delisting-inclusive pool (framework.load_close_volume).

#Auxiliary

• data/cache/sectors_full.json — 497 names, GICS sector+industry. Overlap with the 499-name 30-min universe: 97. Sparse.
• data/cache/pit_sp500_membership.json — monthly PIT membership 2000-2026.
• data/parquet/{SPY,QQQ,IWM,VIX,TLT}_*.parquet, SPY_30Min via daybot.
• data/alpaca_news.py — timestamped news fetch (earnings/analyst/macro filters noted); data/cache/event_study/news/ is empty.
• Fundamentals companyfacts present (market-cap computable); no direct mktcap cache.

#Reliability notes / hazards

• Do not manufacture order flow from OHLCV. Track A factors are named *_proxy and kept separate from Track B.
• 30-min cache contains extended hours; RTH filtering is mandatory.
• All bar timestamps are UTC in the raw repo convention, but the daybot pooled caches are already tz-aware America/New_York.
• Sector confirmation is limited by 97/499 GICS coverage; use correlation peers or residual-vs-market as fallback.
• Survivorship: backtests over the fixed 499/503 lists exclude delistings except the _ADJ pool; use PIT membership for OOS honesty where it matters.

#Plan implied by the audit

1. Track A event study on 30Min_3y (499×3y): return/volume/impact proxy shocks, forward returns, market + sector confirmation, costs. This is the power test of whether the phenomenon exists at all.
2. Track B event study on parquet_micro (AAPL 1s): true OFI/spread/depth channels, validates the proxy and tests reversal-vs-continuation directly.
3. Only if an event-level effect survives → cross-sectional portfolio + OOS + placebo. No production changes regardless of result.