Marina

Point-in-time trading rules for financial markets, as data.

Price limits, fees, trading sessions, settlement — every record says when it took effect and where it was published. Any backtest or trading agent can cite the exact rules it ran under, and anyone can re-run it against byte-identical rules.

中文 · 市场交易规则的「tzdata」：涨跌幅、费用、tick 等制度从哪天起、依据什么，每条记录带官方原文、引文和内容哈希。给回测引擎和 AI Agent 用——查不到就报错，不猜；coverage() 预检整段窗口，resolve() 返回统一判定，diff_rules 告诉你升级后哪些回测要重跑。目前覆盖 A 股（含科创/创业/北交所、ST、IPO 窗口、退市整理期、ETF、REITs、可转债、B 股、T+1）。CC0 数据 + MIT 代码。

rules: marina-rules@2026.10.26.4 sha256:ab6d42b9282414d5

Marina is to trading rules what the tz database is to time zones and what exchange_calendars is to trading days: small, boring, free, and meant to be the one copy everyone imports instead of the thousand copies everyone gets slightly wrong.

Status: 2026.10.26.4. China A-share price limits and transaction fees — 69 records, 63 verified + 6 single-source against primary sources (exchange / ministry / depository notices, with quotes and page hashes). Other markets are not yet covered; see Coverage. Treat the exchange's own notice as authoritative, always.

Why

Every backtest engine, every trading agent, and every research notebook re-implements the same handful of rules — "ChiNext limit is 20%" — and a large share of them get the date wrong (it was 10% before 2020-08-24), the scope wrong (ST stocks differ), or the rounding wrong (exchanges round half-up; round() doesn't). These errors used to live in one firm's code. Now that AI agents write backtests by the thousand, the same error gets copied by the thousand.

A shared, cited, versioned table fixes this the way tzdata fixed time zones: not by being authoritative, but by being the same for everyone and traceable to the source.

Use

pip install marina-rules      # self-contained: the dataset ships inside the wheel, byte-pinned
# or take the JSON / SQLite dump attached to each GitHub Release — same bytes, same digest

import marina_rules as mr

mr.rule_at("stamp_duty_sell", "2023-08-28")           # 0.0005
mr.rule_at("price_limit_pct", "2020-08-23", "chinext", "XSHE")   # 0.10
mr.limit_price(10.00, "chinext", "2021-03-01", "XSHE") # (12.0, 8.0)  — exchange rounding
mr.limit_price(10.00, "ipo_first5_main", "2024-01-30", "XSHG+XSHE")  # None — unlimited band
mr.rule_at("stamp_duty_sell", "2005-01-01")            # KeyError — no defaults, ever
mr.citation()                                          # 'rules: marina-rules@2026.10.25 sha256:…'

# Pre-flight a whole backtest window before running it:
mr.coverage("XSHE", "chinext", ["price_limit_pct"], "2019-01-01", "2026-10-19")
# {'status': 'complete', 'gaps': [], 'conflicts': [], 'verification': 'verified',
#  'required_context': ['board', 'st_status', 'listing_day_index'], 'excludes': [...], ...}

python -m marina_rules.diff 2026.10.24b 2026.10.25b      # what changed, and what must be re-run

coverage() answers "can Marina serve every (rule, day) in this window for this scope?" — complete, partial (with the exact gaps), or unsupported. It also tells you what you must already know about each security to select the scope legitimately (required_context): Marina holds rules, never a security's ST or listing history, and it will not guess them for you. Gaps carry known_gap_ids when they match data/cn/known_gaps.yaml — the list of things Marina knows it does not cover (T+1, sessions, REITs, pre-2022 convertibles…), each with a reason and a status, so "unsupported" can be cited rather than worked around. Agents: read SKILL.md.

The data is plain YAML under data/. You do not need the Python package; read the files from any language.

Cite it

Put mr.citation() in every result you publish. A result without it cannot be reproduced; a result with it can be re-run by anyone against the same bytes.

Check a fill log

python -m marina_rules.check fills.csv [--strict]

Verifies each fill against the rules in force that day — no buys above limit-up, no sells below limit-down, fees not below the statutory floor — and emits a rules-verified@<version>:<hash> stamp when clean. The stamp is recomputable by anyone holding the same log and the same rules version.

What the stamp does not mean: it says nothing about look-ahead bias, survivorship, capacity, or statistical validity. It says only that the fills respected the rules in this dataset. Name it accordingly.

Audit the rules against your own data

python scripts/audit_bars.py /path/to/daily_bars --complete-from-listing

A price-limit rule is falsifiable: after it takes effect, no close may move beyond the band. This script counts violations per rule. A violation means the rule is wrong or your data is — either way you want to know. Marina was first audited this way against 9.1 million A-share bars; see docs/audit-2026-10-24.md. The primary-source audit that followed, and the four corrections it produced, is in docs/sources-audit-2026-10-24.md.

Coverage

Blank cells are blank on purpose. A missing rule raises KeyError; it does not return a guess.

Scope — what Marina will and will not contain

In: rules that change whether an order fills, at what price, or at what cost — price limits, trading hours and auction windows, tick size and lot size, price cages, T+N settlement, statutory fees, exchange margin requirements, trading calendars (delegated to exchange_calendars).

Out, permanently: constituent lists, ST/suspension lists, borrow lists, prices, adjustment factors, anything with an owner or a license. Those are data; Marina holds rules. This boundary is what lets tzdata survive forty years without a legal department, and we are not going to be cleverer than tzdata.

What the dates mean (and don't)

Each record carries three dates: effective_from (when the rule applied), published_at (when the official notice was issued — null when only later restatements were located), and recorded_at (when the record entered Marina). recorded_at is not "known at". Marina was assembled after the fact; it does not by itself let you prove that a backtest avoided look-ahead on rule changes. If that matters, use published_at and treat records with published_at: null as unknown-at-the-time.

One answer shape for every language: the decision contract

python -m marina_rules.contract price_limit_pct 2024-06-14 st_main XSHE --context board=main_board,st_status=true
python -m marina_rules.contract --schema          # JSON Schema of the answer
python -m marina_rules.contract --capabilities    # what can be asked, generated from the data

resolve() returns one dict with status ∈ {DETERMINED, MISSING_CONTEXT, OUT_OF_COVERAGE, CONFLICT, UNVERIFIED}, a result only when determined, the record and evidence refs, known_gap_ids, the ruleset version and hash, and a reason in plain words. It is the boundary other languages, agents and logs should see; the Python exceptions remain for in-process use.

Upgrading: diff_rules

python -m marina_rules.diff <from> <to> lists every change between two tags (or directories) and classifies each as correction (Marina was wrong about the past — results computed under the old version on that interval were computed with a wrong rule), regime_change (the rule itself changed on that date), coverage_added, rename, bookkeeping or metadata. It ends with the (market, scope, rule, interval) list you must re-run. Both versions are pinned by dataset digest. Marina does not estimate P&L impact; that is what the interval is for.

Conformance vectors — check your implementation, in any language

conformance/vectors/ holds 88 input→expected pairs whose expected values come from the exchange texts and hand arithmetic quoted in each vector's oracle, not from running Marina. Regime switches one day before and on the day, ST vs non-ST, IPO windows, both exchanges' one-tick floors, funds, convertibles, fees, missing and contradictory context, known gaps. Marina itself is tested against them in CI; the first run caught a real defect. python conformance/run.py.

Errata

When data contradicts a rule, so far the data has been wrong every time. errata/ records each case with a sample and a one-command detector you can run on your own bars (stale pct-change columns; sub-¥1 adjusted-price rounding). Contributions welcome under the same bar: sample + detector + the fact that settles it.

Verification levels

Every record carries one of three levels, in the data, in the open:

strict=True in the API refuses unverified records. Release tags include all levels; the level is part of the record, not hidden in a changelog.

Contributing

Yes, please — this only works as a commons. One record per pull request, with the primary source linked. CI checks the schema, the sources, interval overlap, and runs the audit on the sample bars. See CONTRIBUTING.md and GOVERNANCE.md.

Acknowledgements

The China A-share timeline was cross-checked against the rule histories compiled in the MIT-licensed PyPI package data-backtest-live-common (2026-09), whose anonymous author did careful work on the fee and limit chronology, and against exchange_calendars for calendar conventions. If you are the dbl_common maintainer: please open an issue; this table should have your name on it.

License

Data (data/, docs/): CC0 1.0 — no rights reserved. Code (src/, scripts/, tests/): MIT.

About the name

Marina is named after my wife, Marina. This project was created and first published on 24 October 2026, her birthday.

A rules table is an unglamorous thing: a few hundred lines of dates, percentages, and citations. But every one of those lines is a mistake that someone, somewhere, no longer has to make. If this table spares even a few people a wrong number, a bad trade, or a wasted week, then it has done what I hoped it would do on the day it was named — bring a little more good into the world, in her name.

It is free for everyone, forever. No rights reserved.

— David Li, 24 October 2026

关于名字

Marina 以我爱人 Marina 的名字命名。这个项目建立并首次发布于 2026 年 10 月 24 日，她的生日。

一张规则表是件很不起眼的东西：几百行日期、百分比和出处。但其中每一行，都是某个人在某个地方不必再犯的一个错误。如果这张表能让哪怕几个人少写错一个数字、少做错一笔交易、少白费一周时间，它就做到了命名那天我希望它做到的事——以她的名字，给这个世界多带来一点好。

它对所有人免费，永远如此。不保留任何权利。