Stop-Loss Placement: ATR, Structure, and the 3-Sigma Trap

Stop-Loss Placement: ATR, Structure, and the 3-Sigma Trap

22 十一月 2026, 02:01
Xiang Jun Jiang
0
1
Almost every "why did my EA fail" conversation ends up here: the stop was too tight for the volatility of the instrument, or it was placed at a level that had nothing to do with how the price actually moves. This article goes through three ways to size and place a stop, when each one is appropriate, and the specific mistake I see most often — sizing the stop to a fixed number of points on instruments with wildly different volatility.

WHY "200 POINTS" IS NOT A RISK NUMBER

On gold, 200 points is roughly two dollars of movement. On a 5-digit EURUSD feed, 200 points is 20 pips. On indices quoted in whole points, 200 points can be less than the spread. The number tells you nothing until you express it in the same unit as the volatility you are trading against.

The practical fix: convert everything to "how many average ranges is this stop". If the instrument's average true range on your trading timeframe is ATR, then a stop of 1.2 x ATR means something consistent across instruments, and 200 points on gold and 200 points on EURUSD stop being treated as if they were the same risk.

A worked comparison. Gold at 3,530 with M15 ATR of 6.5 dollars: 200 points is about 30% of one average range. The price will cross that distance many times inside a normal session, which means the stop is effectively a coin flip against noise. On EURUSD with M15 ATR of 12 pips, 20 pips is 1.7 average ranges, which is a reasonable "give it room to breathe" level. Same number of points, completely different quality of stop.

STRUCTURE-BASED STOPS

The second approach is to place the stop where your trade thesis fails, not where your risk tolerance runs out. For a breakout, that is just below the broken level. For a pullback continuation, it is below the swing low the pullback respects. For a range fade, it is beyond the extreme the range has been tested at.

The advantage is that the stop has an economic meaning: if the price gets there, the reason you entered is gone. The disadvantage is that structure-based stops have variable size. A swing low might be 15 points away on one entry and 90 points on the next. If you size each trade at a fixed lot, your risk per trade will swing with it.

The correct pairing is structure-based stops with risk-based sizing: compute the stop distance in money, then set the lot so that the loss at that stop is your fixed fraction of equity. This is the one place where position sizing and stop placement cannot be separated.

If you must keep lot size fixed, use a volatility-based stop instead, and accept that you are trading a fixed risk rather than a fixed thesis.

THE 3-SIGMA TRAP

A third approach that gets recommended a lot is to set the stop at 3 standard deviations (or 3 ATR) to make the probability of being stopped out "small". The logic sounds statistical and is wrong in two ways.

First, the distribution is not the problem; the sequence is. A stop at 3 sigma has a low probability per trade, but you are exposed to it on every trade, and the losses arrive clustered with the volatility spikes that also invalidate your entry. The trades you lose this way are not random — they are exactly the ones where the market moved faster than your model.

Second, 3-sigma stops make the losing trades enormous relative to the winning ones. If your average win is 1.5 R and your stop is 3 ATR wide, a single stop-out erases the results of several good trades, and the sample needed to distinguish edge from noise grows quickly. Most retail accounts do not survive the sample size required.

The version of this that does work is asymmetric: a wide stop placed at a structural invalidation level, combined with a small position and a first target far enough away that the reward-to-risk stays above 1.5. Wide stop, small size — not wide stop, normal size.

WHAT TO CHECK IN A BACKTEST BEFORE TRUSTING A STOP

1. Average stop distance in ATR units. If it is below 0.8 ATR, expect a high proportion of noise stop-outs whatever the win rate says.
2. Distribution of losing trades. Look for clusters of maximum-loss events; they tell you the stop is not covering volatility regimes.
3. Time in trade. Very short average trade duration with tight stops usually means you are being cycled out by spread rather than by price.
4. Performance by volatility regime. Split the sample by ATR quartile; an EA that only works in the low-volatility quarter is a stop that is too wide for low vol and too tight for high vol.
5. Stop-outs that end in profit. Count trades where the stop was hit and price later reached the target. A high number is a direct measurement of how much you are paying for the stop being too close.

COMMON IMPLEMENTATION MISTAKES

- Comparing the stop to the current price instead of to the entry-derived risk. On a wide-spread instrument, a 10-point stop can be inside the spread on one broker and outside on another.
- Forgetting the stop level minimum. Brokers reject stops closer than the stop level; if your code silently ignores the reject, the position stays unprotected.
- Using indicator values from an incomplete bar. An ATR computed on the forming bar can change every tick, so the stop distance changes during the trade.
- Applying the same ATR multiplier to every instrument. Gold, EURUSD, and indices have different volatility clustering; the multiplier has to be tested per family.
- Ignoring swap and commission when computing R. A "1.5 R" target that is really 1.2 R after costs is not the same trade.

A PRACTICAL DEFAULT

For manual trading and for EA defaults, this combination survives most instruments:

- Stop at 1.2 to 1.5 times ATR on the entry timeframe, or at the structural level if it is farther and no more than twice the ATR distance.
- Position size computed from the stop distance, not fixed.
- First target at 2 to 2.5 times the stop distance; second target optional.
- Never a stop closer than three times the spread.

The stop is not a number you set once. It is the translation of "what would prove me wrong" into price, and the translation has to be done per instrument, per timeframe, and per volatility regime. Any EA that hardcodes points without that translation is not risk management — it is a random number generator with a chart attached.