venvoo/a-share-l2

China A-Share Full-Market Level-2 Tick Archive (2017–2026)

Ten years of full-market Chinese A-share Level-2 data: every listed instrument, every trading day, three synchronized streams — 10-level order-book snapshots (行情), order-by-order submissions (逐笔委托), and tick-by-tick trades (逐笔成交).

To our knowledge this is the largest equity Level-2 corpus in the public domain: existing public A-share L2 datasets cover on the order of weeks; this one covers 2,346 consecutive trading days and keeps growing.

Coverage	2017-01-03 → 2026-08-31, 2,346 trading days; every day carries all three streams (verified file-complete)
Universe	full market as distributed by the feed — 7,880 instruments/day in 2026 (all listed A-shares incl. later-delisted ones, plus ETFs/indices), ~3,000 in 2017
Rows	556 billion messages (orders 287.8B, trades 225.1B, quotes 42.9B — summed from per-day manifests, as of 2026-08-31)
Size	6.18 TB (5.62 TiB) zstd parquet, 7,038 files
Point-in-time	each day is the cross-section as the exchange feed published it — no survivorship filtering
Provenance	Wind Level-2 historical archive; per-day source-archive SHA-256, conversion stats and QC verdicts in manifests.parquet
Daily volume	year average ~0.9 GB/day (2017) → ~6.5 GB/day (2026), ≈7×

Layout

One directory per trading day, three parquet files, no exceptions:

{YYYYMMDD}/
  行情.parquet        # quotes: 10-level book snapshots (~3s cadence per symbol)
  逐笔委托.parquet    # orders: order-by-order submissions/cancels
  逐笔成交.parquet    # trades: tick-by-tick executions

File names are Chinese; column names are English. Per-day QC and provenance records live in a single manifests.parquet at the repo root — one row per trading day (gate verdicts, per-stream row/byte counts, repair provenance, source-archive SHA-256, plus the full original record as a JSON string column).

Schemas

行情 / quotes (67 cols, 29.1M rows on 2026-08-31): wind_code, ex_code, date, time, price, volume, amount, num_trades, iopv, trade_flag, bs_flag, cum_volume, cum_amount, high, low, open, prev_close, ask_px1..10, ask_vol1..10, bid_px1..10, bid_vol1..10, wavg_ask_px, wavg_bid_px, tot_ask_vol, tot_bid_vol, plus index-breadth fields (idx_unweighted, n_sym, n_up, n_down, n_flat).

逐笔委托 / orders (10 cols, 344.9M rows on 2026-08-31): wind_code, ex_code, date, time, order_id, ex_order_id, order_type, order_code (B/S), price, volume.

逐笔成交 / trades (12 cols, 220.2M rows on 2026-08-31): wind_code, ex_code, date, time, trade_id, trade_code, order_code, bs_flag, price, volume, ask_order_id, bid_order_id.

Conventions:

• date is uint32 YYYYMMDD; time is uint32 HHMMSSmmm (e.g. 103807810 = 10:38:07.810), exchange local time (Asia/Shanghai), no timezone field.
• Order/trade IDs are uint64; on days where a stream carries leading-zero IDs the converter falls back to strings for fidelity (flagged as id_varchar_fallback in manifests.parquet).
• Numeric columns are physical DOUBLE. Empty source fields are NULL. Wind's negative-sentinel convention stores negatives as two's-complement uint values — they surface as astronomically large numbers; decode only if you need them.
• ex_code is the exchange's numeric security code, not an exchange identifier. Derive the exchange from the wind_code suffix — and uppercase it first (at least one full day, 2017-12-01, is lowercase .sz/.sh): upper(right(wind_code, 2)).

How to read

Directory-per-day is the only partitioning, and the parquet files currently carry no column min/max statistics (predicate pushdown will not save you). Always restrict by path, never glob the whole corpus:

-- DuckDB (with hf:// support) or local clone
SELECT * FROM read_parquet('20250825/逐笔成交.parquet');
SELECT time, price, volume
FROM read_parquet('2025*/行情.parquet')          -- one year of quotes, still ~305 GB
WHERE upper(wind_code[-2:]) = 'SZ' AND wind_code LIKE '000001%';

from huggingface_hub import hf_hub_download
import pyarrow.parquet as pq
f = hf_hub_download("venvoo/a-share-l2", "20250825/逐笔委托.parquet", repo_type="dataset")
t = pq.read_table(f, columns=["wind_code", "time", "order_type", "price", "volume"])

For serious work, download the date ranges you need (hf download venvoo/a-share-l2 --repo-type dataset --include "2021*/*") and build your own symbol-partitioned feature layer; the day-partitioned raw layout is optimized for archival integrity, not for per-symbol queries.

Six data truths — read before computing anything

These are the traps we hit while using this corpus for live research. All of them are invisible to schema inspection.

1. SSE order stream starts 2021-07-23, and its semantics differ from SZSE. The Shanghai exchange only began publishing order-by-order data on 2021-05-10 (this feed carries it from 2021-07-23); before that, SSE rows in 逐笔委托 are simply absent. Worse, the two exchanges publish different things: SZSE sends original orders; SSE sends the remainder after partial fills. Computed literally, the 2021-12-15 "cancel rate" is 28% on SSE vs 2.5% on SZSE — a 10x gap that is pure encoding, not behavior:

Consequences: compute order-flow features per exchange; never compare across 2021-07-23 in a time series; SZSE is the only exchange with consistent order semantics over the full 2017–2026 span.

2. SZSE order_type has two dialects, switching on some days — once intraday. Normal days use '0'=limit, '1'=market, 'U'=best-own. Some days use the exchange-native encoding '2'=limit (e.g. 2022-05-23 and 2024-06-17 all day; 2022-08-19 switches from '0' to '2' at exactly 11:00:00.000). Filter limit orders with order_type IN ('0','2') — filtering on '0' alone silently drops entire days. SZSE cancellations are unaffected: they live in the trades stream as trade_code = 'C', stable across the full history.

3. No parquet column statistics. See "How to read": path-restrict by date, don't rely on row-group pruning. (A v2 re-encode with statistics and higher compression is planned.)

4. Repairs are recorded, not hidden. The raw vendor export had systematic defects. Every repair leaves a per-day audit trail in manifests.parquet:

field	meaning	scale
date0_sentinel_junked	date=0 heartbeat/filler rows dropped	405 days / 17.3M rows
date0_repaired	date=0 rows with real payloads, date restored from the archive date	555 days / 15.8M rows
nul_bytes_stripped	embedded \0 bytes removed (→ NULL)	hundreds of millions of bytes
stray_date_rows_dropped	rows from another date mixed into an archive (verified duplicates)	1 day / 0.37M rows
extra_tail_cols_dropped	late-era extra trailing columns (limit-up/down px, arrival ts) dropped for schema stability	5 days

Strict studies can exclude affected days by scanning these fields.

5. Suffix case instability. At least 2017-12-01 uses lowercase .sz/.sh in wind_code all day. Uppercase before any exchange split, or you match zero rows and never notice.

6. IDs are exchange sequence numbers, not accounts. There is no investor- or account-level identifier anywhere in the corpus (verified column-by-column); order_id/trade_id/ask_order_id/bid_order_id support order-lifecycle reconstruction (submission → partial fills → cancel/full fill) but cannot be aggregated to persons.

QC and provenance

Each day passed a four-gate pipeline before admission — G0 listing fingerprint, G1 file inventory, G2 partition integrity, G3 row-count reconciliation vs raw, G4 semantic checks — and the verdicts are stored per day in manifests.parquet (gates_all_green = true for all 2,346 days). It also records the source 7z archive's SHA-256, per-stream raw/parquet byte counts, row counts, symbol counts, and time ranges, so the entire chain from vendor archive to parquet is auditable per day.

Defect classes found and healed during conversion (in addition to the repair table above): variable archive nesting, mixed \r\n/\n line endings within a single day, mid-history schema upgrades (66→69/70 trailing columns), and NUL-byte contamination that silently truncates naive line-based tooling.

Updates

The underlying pipeline converts each new trading day at T+1. The Hugging Face mirror is synced in batches; expect the tail to lag the exchange by days, not months.

License and terms

Research use only (license: other). The underlying market data originates from exchange Level-2 feeds distributed via Wind's historical archive; redistribution here is for academic and non-commercial research. No warranty of fitness; verify against primary sources before trading decisions. If you represent a rights holder and have concerns, open a discussion on this repo.

Citation

@misc{ashare_l2_2026,
  title  = {China A-Share Full-Market Level-2 Tick Archive (2017--2026)},
  author = {venvoo},
  year   = {2026},
  url    = {https://huggingface.co/datasets/venvoo/a-share-l2}
}