howard-lynn-ye China trading stack
A backtest engine written for US equities is wrong on A-shares by default — it will let you sell what you bought this morning, fill you at limit-up, and trade through a suspension. §3 is the list of rules your engine must encode.
Data sourcing is a separate skill: china-ashare-data.
1. Pick a framework
| Task | Use | Note |
|---|---|---|
| Live trading, broadest gateway coverage | vnpy 4.4.0 (MIT, 45,089★, pushed 2026-09-01) | The de-facto standard. Ships no data of its own |
| Futures research + live, incl. 夜盘 | tqsdk 3.10.2 (Apache-2.0, 4,991★, only 4 open issues) | Best-maintained here. Free tier covers all futures + options |
| Quant research pipeline / factor ML | Qlib pyqlib 0.9.7 (MIT, 48,253★) |
🚨 official China dataset is disabled — see §2 |
| Event-driven A-share backtest | RQAlpha 6.3.0 | 🚨 custom licence: non-commercial only; sdist-only, no Windows wheel |
| C++-core multi-asset | wondertrader / wtpy (MIT) | wtpy slowing (last push 2025-08) |
| Retail broker automation | easytrader 0.23.7 (MIT, 10,124★) | Slowing; GUI-automation fragility |
| HK/US via a Chinese broker | futu-api, tigeropen |
Vendor-maintained. ⚠️ futu-api is sdist-only but pure Python, so it installs fine on Windows; it does require the separate FutuOpenD daemon |
2. 🚨 Qlib's China dataset is disabled
Qlib's README states verbatim: "Due to more restrict data security policy. The official dataset is
disabled temporarily." The documented qlib_data --region cn command still appears in the README
but the backing dataset is down.
✅ The live replacement is community-run: chenditc/investment_data — 1,440★, Apache-2.0, latest
release tag 2026-09-03 published the same day (it updates daily), asset qlib_bin.tar.gz
at 563.8 MB. It merges Tushare + akshare + baostock + Yahoo + historical Wind/Caihui dumps,
cross-validates them, rescales adjustment factors to a common basis, and explicitly backfills
"delist company's data" → delisted names are included, which the official bundle never handled well.
wget https://github.com/chenditc/investment_data/releases/latest/download/qlib_bin.tar.gz
mkdir -p ~/.qlib/qlib_data/cn_data
tar -zxvf qlib_bin.tar.gz -C ~/.qlib/qlib_data/cn_data --strip-components=1
⚠️ Qlib's cn_data stores a $factor field and adjusted prices — know which convention a dump
uses before mixing it with any other source. Qlib's own check_data_health script exists precisely because
these dumps have gaps. Qlib itself provides no PIT fundamentals; instruments/csi300.txt encodes
membership date ranges whose quality is entirely inherited from the dump you installed.
3. 🚨 Rules a Western engine will get wrong
Encode all of these, or your A-share backtest is fiction.
- T+1 settlement. Shares bought today cannot be sold until the next session (cash is T+0). Any intraday-reversal strategy on A-shares is unimplementable. Futures are T+0 — a mixed equity/futures backtest needs two settlement models.
- Price limits, which vary by board and by ST status. 主板 ±10% (ST/*ST ±5%), 创业板
300xxx/301xxx±20%, 科创板688xxx±20%, 北交所 ±30%; new listings differ again on day 1. A bar closing at limit-up generally could not have been bought — filling there is the second largest source of fake alpha in A-share research. - Suspensions run for months. Sources variously omit the bar, emit
volume=0with a forward-filled price, or emit NaN. Forward-filled bars create spurious zero returns that deflate measured volatility. - The 90-minute lunch break. Sessions are 09:30–11:30 and 13:00–15:00 — exactly 240 minute
bars.
resample('1min')or any wall-clock rolling window silently spans 90 minutes of non-existent time. Roll over bar index, not wall-clock. - Futures 夜盘 runs ~21:00 to 23:00/01:00/02:30, so a futures trading day starts the previous calendar evening. tqsdk and rqdatac handle the mapping; scrapers do not.
- Forward-adjusted (前复权) prices are look-ahead contaminated — see
china-ashare-data§2. - Fees and 印花税: stamp duty is charged on the sell side only; add 过户费 and broker commission with a per-order minimum. A cost model ported from US equities understates round-trip cost asymmetrically.
4. vnpy specifics
vnpy is a framework, not a data source — every datafeed adapter wraps a third-party, mostly paid service. ✅ Verified on PyPI:
| Adapter | Provider | Cost | Latest |
|---|---|---|---|
vnpy_xt |
迅投研 | Paid | 1.4.6 (2025-10-18) |
vnpy_rqdata |
米筐 | Paid | 3.4.7.8 (2026-04-16) |
vnpy_tushare |
Tushare | 积分 | 1.4.21.0 (2025-06-11) |
vnpy_tqsdk |
天勤 | Paid | 3.8.6.0 (2025-10-02) |
vnpy_wind / vnpy_ifind |
万得 / 同花顺 | Paid | 1.1.0 (2025-06-11) |
vnpy_udata / vnpy_tinysoft |
恒生 / 天软 | Paid | ⚠️ 2023 — stale |
✅ vnpy_datayes (通联) does NOT exist on PyPI — if a doc lists it, that is wrong.
Config is uniform: SETTINGS["datafeed.name"/"username"/"password"]. ⚠️ For RQData these are not
your ricequant.com website login.
🔑 vnpy_datarecorder records live tick/bar from a connected trading gateway into a local DB —
so with a broker CTP account you can build your own tick history for free, bypassing vendor fees.
The most under-appreciated capability in this ecosystem.
5. Order safety
Everything in broker-execution-apis §3 applies, plus:
- CTP simulation (SimNow) vs production differ by broker ID and front address — assert on a server-returned account field, not a config flag.
- easytrader drives a GUI; a UI change or a popup silently breaks it mid-session. Never leave it unattended without an external position reconciliation loop.
- QMT / miniQMT ships from the broker, not PyPI, and its API surface varies by broker build. Pin the build you tested against.
6. Reference files
references/<framework>.md for architecture, licence, gateway coverage and setup;
references/_ashare-rules.md for the machine-checkable version of §3.
Per-library deep dives
The optional fin-libraries plugin carries a dedicated skill for each library below. Load one
only after this skill has told you which library you want:
lib-qlib— qlib