I Tested 8 Stop-Loss Brackets on 32,485 A-Share Trades. The '1:2 Risk-Reward' Rule Needed 33% and Got 24%
Every trading book hands you the same rule: risk 1 to make 2. Set the stop at -5%, the target at +10%, and you only need to be right one time in three. I had 32,485 A-share trades lying around, so I measured it. The rul
Every trading book hands you the same rule: risk 1 to make 2. Set the stop at -5%, the target at +10%, and you only need to be right one time in three.
I had 32,485 A-share trades lying around, so I measured it. The rule got nowhere near one time in three. It hit 23.8% and only came out flat because of a bucket almost nobody budgets for.
The setup
- 587 A-share names (538 Shanghai main board, 29 STAR, 6 ChiNext), 321 daily bars each, 2025-07-17 to 2026-11-12.
- Entry at the next day's open after a signal day. No same-bar fill, no lookahead.
- Barriers measured from the entry price against intraday high/low, horizon 20 trading days, sampled every 5 days. Because that overlaps, I re-ran everything at step 20 (8,131 non-overlapping entries) and flag any result that changes sign.
- When both barriers are touched on the same day I count it in its own bucket instead of silently picking a winner. Depending on width that bucket is 0.0%-1.7% of trades.
- Filling at the barrier is optimistic. I measured that separately instead of assuming it away.
The universe was flat over the sample: mean 20-day return -0.04%. Nothing here is beta.
Result 1: every "good" ratio missed its break-even rate
"Needed" assumes zero timeouts and a barrier fill. Those assumptions favour the rule, and it still loses.
| Stop / Target | Target hit first | Stop hit first | Timeout | Mean per trade | Hit rate needed |
|---|---|---|---|---|---|
| 3% / 3% | 47.0% | 49.5% | 1.9% | -0.08% | 50.0% |
| 5% / 5% | 43.8% | 44.2% | 11.7% | -0.08% | 50.0% |
| 3% / 6% | 29.6% | 63.5% | 6.4% | -0.07% | 33.3% |
| 5% / 8% | 30.0% | 50.5% | 19.4% | -0.03% | 38.5% |
| 5% / 10% | 23.8% | 52.7% | 23.5% | 0.00% | 33.3% |
| 8% / 16% | 14.7% | 35.6% | 49.6% | +0.05% | 33.3% |
| 10% / 10% | 29.2% | 24.5% | 46.3% | 0.00% | 50.0% |
| 10% / 20% | 10.7% | 26.6% | 62.7% | +0.04% | 33.3% |
The only rows that didn't lose money are the ones where timeouts did all the work. For 5%/10%, 23.5% of trades never touched either barrier, and those trades averaged +1.09% by day 20 β that bucket is what drags the mean to zero. The bracket itself is a 24%-hit-rate machine that needs 33%.
The wider the target, the faster the hit rate collapses. That's not a tuning problem, it's arithmetic: a 10% target inside a 20-day window in a market whose 20-day mean is -0.04%.
Result 2: the tight bracket is a coin flip with a fee
At Β±3%, timeouts are 1.9%. You're not trading a trend anymore, you're trading intraday noise, and the noise costs 0.08% per trade. 49.5% against 47.0% is not a strategy, it's a coin that pays the house.
Result 3: the population matters more than the bracket
Same Β±5% bracket, split by whether the signal day closed above its 20-day moving average:
| Population | Target hit first | Mean per trade |
|---|---|---|
| Close > MA20 | 46.0% | +0.03% |
| Close < MA20 | 41.8% | -0.17% |
The bracket didn't improve. The population did. Fix the population first, then tune the bracket β the reverse order just fits a rule to names you should not have been in.
Result 4: a stop's value is in the tail, not in the mean
Drop the target and use the stop alone, holding to day 20 otherwise:
| Stop | Mean | Std dev | 5th pct | 1st pct | Worst |
|---|---|---|---|---|---|
| none (hold) | -0.04% | 12.38% | -16.69% | -27.55% | -56.21% |
| -3% | +0.01% | 7.85% | -3.00% | -3.00% | -3.00% |
| -5% | +0.05% | 9.38% | -5.00% | -5.00% | -5.00% |
| -8% | +0.06% | 10.66% | -8.00% | -8.00% | -8.00% |
| -10% | +0.05% | 11.16% | -10.00% | -10.00% | -10.00% |
| -15% | +0.02% | 11.82% | -15.00% | -15.00% | -15.00% |
The mean barely moves β and honestly, it isn't robust. The paired difference is +0.09pp in favour of the stop on the overlapping sample (t = -1.98), but on the non-overlapping sample it flips to +0.19pp in favour of holding. If your argument for a stop is "it improves the average", you don't have an argument.
The tail, though, is mechanical, not statistical. The 5th percentile goes from -16.69% to -5.00%. The 1st percentile from -27.55% to -5.00%. The worst trade from -56.21% to -5.00%. Std dev from 12.38% to 9.38%. A stop doesn't buy you return, it buys you a bounded distribution β and that's the thing that decides whether you get to keep trading.
Result 5: a third of your stops are wicks
- 56.83% of 20-day windows touched -5% intraday.
- 49.02% closed below -5%.
- The difference, 7.8pp, is trades that went through your stop and finished the day above it.
If you can only act on the close, your stop gets hit 7.8 percentage points less often, on the same population. That's free information.
Timing matters too: median day to touch is 6 of 20, and 59.6% of all touches happen in the first 7 days. The stop decision is a week-one decision. If a trade is going to shake you out, it usually does it before you've had time to build a story about it.
Result 6: gaps get through, and you should price that in
| Bracket | Stops filled through the barrier | Median fill | Mean fill |
|---|---|---|---|
| 3% / 3% | 6.0% of all stops | -3.47% | -3.86% |
| 5% / 5% | 7.8% | -5.53% | -6.00% |
| 5% / 10% | 8.0% | -5.58% | -6.08% |
| 10% / 20% | 9.9% | -10.65% | -11.14% |
A -5% stop is really a -5.08% stop once you weight the 7.8% that opened below it. A -3% stop is worse in relative terms. Limit-down opens in A-shares don't ask permission.
Result 7: the first week sorts trades into two different worlds
Of the 18,461 trades that touched -5%:
- mean 20-day return: -5.16%
- ended positive: 20.7%
- ended above +10%: 5.8%
- ended below -10%: 25.5%
Of the 14,024 that never touched it:
- mean 20-day return: +6.69%
- ended positive: 74.6%
This is hindsight selection β you cannot trade "names that never touched -5%", and I'm not pretending otherwise. What it does show is that the stop isn't merely a loss limit. Being stopped out is a strong signal that you're in the other distribution, and the two distributions barely overlap in outcome.
What I actually changed
- 8%+ stop distance for 20-day horizons. Anything tighter than the noise band is a coin flip with a fee. The Β±3% row is the cleanest evidence: 94% of trades resolve, and they resolve at chance.
- Never pair a wide target with a narrow stop and expect width to fix the hit rate. 5%/10% needs 33.3% and delivers 23.8%. The only reason people think it works is the timeout bucket.
- Filter the population before the bracket. Same Β±5% is +0.03% above MA20 and -0.17% below it. One line of code beats any amount of stop tuning.
- Close-based stops when you can't watch intraday. 7.8pp fewer false stops on identical data.
- Size for the fill, not the barrier. Budget ~0.1pp of slippage on a -5% stop; more on tighter ones.
Caveats, because this is one sample
- 587 names, mostly one exchange, one 16-month regime. Not a claim about A-shares in general.
- Barriers can be blown through; fills at the barrier are optimistic and I only corrected for gap-through, not intraday slippage or liquidity.
- 5-day sampling overlaps; every result I quote here I also re-ran at step 20 and said so where the sign flipped.
- The MA20 split is in-sample. It's a filter direction, not a fitted parameter.
Why this belongs in an agent's skill file
These rules live in the A-share skill my coding agent loads before it screens anything. The reason isn't the numbers β it's that an LLM will happily write stop_loss: 0.03, take_profit: 0.06 into a config file and call it risk management, because it looks like risk management. The calibration gate is what stops that from shipping.
The study scripts are plain stdlib Python against cached JSON β no pandas, no TA library, reproducible in about 1.4 seconds per run. The write-ups and tools I keep public are here:
- GitHub: https://github.com/Felixwang007/awesome-content-tools β the audit write-ups and the small verification tools (gate checks, idempotency probes, clock calibration) that come out of this habit.
- The A-share skill packages (screening, three-pillar scoring, and this stop calibration) are published on θΎθ― / xiaping for anyone wiring an agent into Chinese market data.
If you take one thing: measure the barrier before you trust it. The bracket you were taught needs a hit rate the market doesn't hand out.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.