China Ashare Trading Taxes

China A-share trading taxes

Two taxes decide what an A-share round trip costs, and a flat symmetric basis-point model can express neither of them. One is charged to the seller only and halved in the middle of any sample that spans 2023-08-28. The other charges you for holding briefly.

These are modelling assumptions for a backtest, not tax advice, and everything here is for an individual resident investor. Institutions, funds, QFII/RQFII and Stock Connect investors are treated differently and none of that is in this file. Confirm current law with a qualified professional before relying on any of it.

✅ Measured comes from scripts/ashare_taxes.py — numpy + pandas, seed 20260909, one seeded 2021→2025 path that deliberately spans the rate cut, under 1 s. ✅ source-verified material was read on 2026-09-09 at 国家税务总局 (fgk.chinatax.gov.cn, www.chinatax.gov.cn) and 财政部 (www.mof.gov.cn).

1. ✅ source-verified — stamp duty is one-sided, and it moved

中华人民共和国印花税法, Article 3, in force from 2022-07-01, verbatim:

本法所称证券交易，是指转让在依法设立的证券交易所、国务院批准的其他全国性证券交易场所交易的股票和 以股票为基础的存托凭证。证券交易印花税对证券交易的出让方征收，不对受让方征收。

Securities transaction stamp duty is levied on the transferor, and not on the transferee. The attached 印花税税目税率表 sets the rate at 千分之一 (0.1%) of 成交金额, and Article 5 makes the transaction amount the tax base.

财政部 税务总局公告2023年第39号《关于减半征收证券交易印花税的公告》, issued 2023-08-27, verbatim: "为活跃资本市场、提振投资者信心，自2023年8月28日起，证券交易印花税实施减半征收。" 0.1% × 50% = 0.05%, still seller-side only.

财政部 release of 2008-09-19, verbatim: "调整为单边征税，即对买卖、继承、赠与所书立的A股、B股 股权转让书据的出让方按千分之一的税率征收证券（股票）交易印花税，对受让方不再征税。" That is the date the duty stopped being two-sided. ⚠️ The document number is widely reported as 财税明电〔2008〕2号; the MOF page itself does not carry it, so treat the number as secondhand and the substance as verified. (⚠️ The research this skill started from named 财税〔2008〕96号 — that is wrong.)

🚨 Before 2008-09-19 the duty was two-sided and the rate moved several times (it was raised to 0.3% in 2007). This library did not verify those rates, so stamp_duty_rate() raises for any date before 2008-09-19 rather than extrapolating the table backwards.

⚠️ ETFs and funds are outside the tax. Article 3 defines 证券交易 as transferring 股票 and share-based 存托凭证. A fund unit is neither. ✅ Measured: stamp_duty(1_000_000, "S", ..., instrument="etf") returns 0.00 where the same trade in shares returns 500.00. A strategy implemented through an A-share ETF has a genuinely different tax structure from the same strategy in the constituents — which is a real reason to prefer one implementation, and one no constituent-level backtest sees.

2. ✅ Measured — what a flat symmetric model gets wrong

Seeded path 2021-01-04 → 2025-09-05, 100,000 shares, round-tripped at several frequencies. "Flat 10 bps/side" is the model most A-share backtests use; "real stamp duty" is the tax alone.

🔑 The point is not that 10 bps is too big. A calibrated 10 bps a side is a defensible total cost — commission, 过户费, spread and impact all live in there too. The point is the structure:

• It is symmetric and the tax is not. The buy side pays zero duty. A model that charges both sides equally misprices any strategy whose buys and sells are not balanced — an accumulating book, a liquidating book, or any comparison between a long-only sleeve and a turnover-matched one.
• It is constant and the tax is not. ✅ Measured on the same monthly blotter, split at the cut date: 10.0 bps of sell notional before 2023-08-28, 5.0 bps from it. A sample that spans that Monday contains two cost regimes, and a single number is wrong on one side of it whichever number you choose. Backtests run before and after 2023 are not comparable on cost without saying which rate they used.

⚠️ Not modelled here, because they are fees rather than taxes: 佣金 (broker commission, with its per-order minimum), 过户费 (transfer fee, charged on both sides), 规费, and the T+1 constraint that changes what a "round trip" even means. Those belong to ../../../fin-china/skills/china-trading-stack/SKILL.md.

3. ✅ source-verified — the dividend tax is a step function of holding period

财税〔2015〕101号《财政部 国家税务总局 证监会关于上市公司股息红利差别化个人所得税政策有关问题 的通知》, issued 2015-09-07, effective for record dates from 2015-09-08:

🚨 Both boundaries are inclusive on the taxed side. ✅ Measured — the boundary is one day wide:

🔑 Calendar months, not 30 days. holding_bucket() uses pd.DateOffset, so a purchase on 2024-01-31 sold on 2024-02-29 is inside one month while a 30-day counter would say otherwise.

⚠️ The timing is the part that breaks a cash-flow model. 财税〔2015〕101号 defers the withholding: the listed company withholds nothing at payment, and 中国证券登记结算公司 computes and withholds the tax when the shares are sold, using the holding period on a first-in first-out basis. So the dividend arrives gross, the tax comes out of a later sale, and a backtest that nets the tax on the ex-date has both the amount and the date wrong.

4. ✅ Measured — the turnover penalty is in the tax code

Same seeded path, same 2.5% dividend yield, same one dividend a year, same shares. The only difference is how often the position is round-tripped.

🚨 48 basis points a year, on a 2.5% yield, purely for rebalancing monthly instead of not rebalancing. That is larger than the stamp duty on the same strategy, it does not appear in any transaction-cost model, and it does not shrink as you trade more cheaply — it is not a cost of trading, it is a cost of not holding.

🔑 The step is where the design lives. Moving a monthly rebalance to a 32-day one takes the dividend rate from 20% to 10%; moving an annual rebalance one day later takes it from 10% to 0%. If a strategy's holding period is anywhere near a boundary, the boundary is worth more than most of its signal work — and it is a schedule change, not a signal change.

⚠️ A dividend-capture strategy is the extreme case and the arithmetic is brutal: it collects the dividend, pays 20% of it, and eats the ex-date drop in full. Model both halves before concluding anything about it.

5. What this skill does not cover

⚠️ Dated claim, re-check before use. Everything above was verified on 2026-09-09. The stamp duty rate in particular is an active policy lever — it was cut in 2008 and halved again in 2023, both times as market support — so it is exactly the kind of number a model's training prior will get wrong. The most recent related instrument this library found is 财政部 税务总局 中国证监会 公告2026年第8号 (2026-01-14), which extends the innovative-enterprise CDR pilot policies through 2027 by reference to 财税〔2015〕101号 and changes neither the dividend brackets nor the duty rate.

6. Scripts and where this sits

scripts/ashare_taxes.py — stamp_duty / stamp_duty_rate (dated table, seller side only, raises before 2008-09-19, zero for funds), rate_history_table, holding_bucket, dividend_tax_rate / dividend_tax, flat_symmetric_cost, real_stamp_duty, compare_cost_models, dividend_drag, and the seeded make_ashare_path / make_turnover_blotter. numpy + pandas, seed 20260909, under 1 s.

• A-share data, T+1, price limits, suspension and the broker stack — ../../../fin-china/skills/china-trading-stack/SKILL.md and ../../../fin-china/skills/china-ashare-data/SKILL.md.
• Reporting the result with the jurisdiction and rate attached, so it is comparable to a US number — ../after-tax-backtesting/SKILL.md, whose guard exists for exactly this.
• The US analogues: lot matching (../tax-lot-matching-and-cost-basis/SKILL.md), the wash sale (../wash-sale-rules/SKILL.md), and 60/40 (../section-1256-and-derivatives-tax/SKILL.md) — none of which has an A-share equivalent, which is itself the reason to keep the jurisdiction in the report line.
• Turning any of these drags into a breakeven and a capacity limit — ../../../fin-core/skills/backtest-validation/scripts/cost_curve.py.