Why Stop-Losses Do Not Prevent Liquidation: Trigger, Mark and Fill

本文为原文前 6,000 字符的节选翻译，完整内容请查看原文。

Why Stop-Losses Do Not Prevent Liquidation: Trigger, Mark and Fill

A stop-loss feels like a promise. You set a price, the system watches for it, and your position closes before the damage spreads. Every part of that sentence is true except the guarantee, and the gap between “watches for it” and “closes it” is where accounts die. Three different prices are involved. None of them is the one most traders assume. A stop-loss is a conditional order, not a risk limit. The order sits in a conditional queue until a price condition is met. Meeting the condition does not close anything. It releases an order into the book, and that order then has to find a counterparty at some price. Between the condition and the fill there are at least four failure points: the condition may be evaluated on a different price feed than the one that triggers liquidation, the released order may not fill, the fill may be far from the trigger, and the whole pipeline may be slower than the price move.

止损单感觉像是一种承诺。你设定一个价格，系统会监控该价格，并在损失扩大前平仓。这句话的每个部分除了“保证”之外都是真实的，而“监控”与“平仓”之间的差距正是账户归零的地方。这涉及三个不同的价格，但没有一个是大多数交易者所认为的那样。止损单是一种条件订单，而非风险限额。该订单会停留在条件队列中，直到满足价格条件。满足条件并不会平仓，它只是将订单释放到订单簿中，而该订单随后必须以某个价格找到交易对手。在条件触发与成交之间，至少存在四个失败点：条件可能基于与触发强平不同的价格源进行评估；释放的订单可能无法成交；成交价可能远偏离触发价；整个处理流程可能比价格波动更慢。

Three prices, three jobs. The mechanics separate cleanly once you name them. The last traded price is the most recent print on that venue. It is fast, noisy, and thin books make it jump. The mark price is a reference price the platform computes to assess position risk and calculate unrealized PnL. Its construction is venue-specific, and it is what liquidation logic generally watches. The fill price is whatever the book pays when your order actually executes. A stop-loss is usually evaluated against the first. Liquidation is usually evaluated against the second. Your exit executes at the third. Three different series, three different moments, one position.

三个价格，三种职能。一旦命名，其机制便清晰可见。最新成交价是该平台上最近的一笔交易记录。它速度快、噪音大，且在流动性薄弱的订单簿中容易跳动。标记价格是平台为评估持仓风险和计算未实现盈亏而计算的参考价格。其构建方式因平台而异，也是强平逻辑通常监控的对象。成交价是订单实际执行时订单簿所支付的价格。止损单通常根据第一个价格进行评估，强平通常根据第二个价格进行评估，而你的平仓则在第三个价格执行。三个不同的序列，三个不同的时刻，却针对同一个仓位。

What the exchange documentation says outright. This is not a subtle implementation detail. One exchange’s Academy guide states it directly: “The stop-loss is triggered by the last traded price, while liquidation is based on the mark price. Liquidation occurs before the stop-loss is triggered.” That single sentence explains a class of support tickets. You set a stop 2 percent away, the mark moved 2 percent first, and the liquidation engine reached the margin boundary before the conditional order ever became eligible. The same guide lists the adjacent failure modes in the same language: an order triggered but not filled because no counterparty appeared or slippage was severe, price and liquidity gaps past the trigger level, cascading liquidations when many conditional orders fire together, and matching engine delays or API rate limits.

交易所文档中的明确说明。这并非细微的实现细节。某交易所的学院指南直接指出：“止损单由最新成交价触发，而强平基于标记价格。强平发生在止损单触发之前。”这一句话解释了一类客服工单。你设置了2%的止损，但标记价格先波动了2%，导致强平引擎在条件订单生效前就触及了保证金边界。同一指南以同样的语言列出了相关的失败模式：订单触发但因无对手方或滑点严重而未成交；触发价格水平后的价格与流动性缺口；当大量条件订单同时触发时引发的连锁强平；以及撮合引擎延迟或API速率限制。

Triggered and still open. The second failure is structural rather than adversarial. A stop-limit order that fires into a falling book can sit unfilled at a price nobody wants. A stop-market order fills, but at whatever the book offers. If your intent is exit, a limit price inside a stop is a decision to prefer a bad outcome over an uncertain one. That is a legitimate choice, but it is not the same as protection. There is also a queue position problem. When thousands of conditional orders fire on the same tick, they are not all first in line. Some fill well, some fill late, and the ones that fill late fill worse.

触发后仍未平仓。第二个失败是结构性的，而非对抗性的。当止损限价单在下跌的订单簿中触发时，可能会因无人接盘而无法成交。止损市价单虽然会成交，但成交价取决于订单簿提供的价格。如果你的目的是平仓，在止损单中设置限价意味着你宁愿接受糟糕的结果，也不愿面对不确定的结果。这是一种合理的选择，但它不等同于保护。此外还存在队列位置问题。当数千个条件订单在同一时刻触发时，它们并非都能排在第一位。有些成交良好，有些成交滞后，而滞后的订单成交情况更差。

Latency is a measurable term, not a vibe. Recent quantitative work treats the delay explicitly. A 2026 paper in Forecasting models liquidation as a first-passage problem where the executable price is observed through a delayed or smoothed mark, funding drains collateral continuously, and liquidation fires when the isolated margin surplus hits its maintenance boundary. Two results are useful to anyone building or auditing this kind of system. Under a fixed observation delay, expected overshoot scales with the square root of that delay, and the probability of bad debt is governed by a latency-to-margin ratio. Their published 60 minute boundary-breach probabilities for a ten times position are 1.7 percent on BTC, 4.1 percent on ETH and 13.5 percent on SOL. Those numbers move higher when the mark is time-weighted rather than instantaneous. Read that as an engineering statement: a smoother mark reduces noise and increases the loss conditional on detection. Both choices are defensible. Neither is free.

延迟是一个可衡量的术语，而非一种感觉。近期的量化研究明确处理了这种延迟。2026年发表在《Forecasting》上的一篇论文将强平建模为首过时间问题，其中可执行价格通过延迟或平滑的标记价格观察，资金费率持续消耗抵押品，当逐仓保证金盈余触及维持保证金边界时触发强平。对于构建或审计此类系统的人来说，有两个结论很有用。在固定的观察延迟下，预期的超调量与延迟的平方根成正比，而坏账概率由延迟与保证金的比率决定。他们公布的十倍杠杆仓位在60分钟内的边界突破概率分别为：BTC 1.7%，ETH 4.1%，SOL 13.5%。当标记价格采用时间加权而非瞬时价格时，这些数字会更高。将其理解为工程陈述：更平滑的标记价格减少了噪音，但增加了检测时的损失。两种选择都有其合理性，但没有一个是免费的。

Cascades are subcritical more often than the narrative suggests. The other 2026 study measures an actual cascade in flight, using the on-chain fill log of a transparent venue during the October 2025 event. Its branching ratio stayed around 0.1 to 0.2, which is well below the critical threshold of one, and 88 percent of post-onset forced selling landed inside thirty minutes. The detail worth remembering is where the pressure went. 63 percent of that forced selling was absorbed off-book by the venue’s backstop rather than by the visible order book, which pushes the measured branching ratio down precisely at the climax. For a trader this means the book you can see is not the liquidity that is actually absorbing the event. For an engineer it means cascade metrics computed from visible liquidity will systematically overstate the danger at the moment it matters.

连锁反应的临界性往往低于市场叙事。另一项2026年的研究通过使用2025年10月事件期间某透明交易所的链上成交日志，测量了实际发生的连锁反应。其分支比率保持在0.1到0.2左右，远低于1的临界阈值，且88%的强制抛售发生在事件开始后的30分钟内。值得记住的细节是压力流向了哪里。63%的强制抛售是由交易所的后备机制在订单簿之外吸收的，而非通过可见的订单簿，这恰恰在最高潮时压低了测得的分支比率。对于交易者而言，这意味着你看到的订单簿并非实际吸收事件的流动性。对于工程师而言，这意味着根据可见流动性计算出的连锁反应指标，会在关键时刻系统性地高估风险。

What to check before trusting any stop-loss setup. Four things, in order of how much they usually matter. Confirm which price feed your stop reacts to, and which feed governs liquidation. If they differ, compute the historical disagreement rather than assuming it is small. A practical way to do it: pull one minute candles for the last month and take the maximum positive gap between mark and last price inside each hour. In quiet hours the gap is a few basis points. Around liquidation clusters it can pass your entire stop distance, and that is the hour the stop gets skipped. Decide whether your exit needs certainty or price. A market order buys certainty and pays spread plus slippage. A limit order buys price and risks not exiting. Size so the maintenance boundary is far from your stop, not adjacent to it. If liquidation sits 30 basis points behind your stop, the entire question becomes which feed updates first. Watch funding as a slow trigger. It moves the margin boundary between settlements, so a position that was comfortable at entry can drift into breach without.

在信任任何止损设置前需要检查的事项。按重要性排序有四点。确认你的止损对哪个价格源做出反应，以及哪个价格源决定强平。如果两者不同，请计算历史偏差，而不是假设其很小。一种实用的方法是：拉取过去一个月的1分钟K线，计算每小时内标记价格与最新成交价之间的最大正向缺口。在平静时段，缺口仅为几个基点。但在强平密集时段，缺口可能超过你的整个止损距离，而这正是止损失效的时刻。决定你的平仓是需要确定性还是价格。市价单买入确定性，但需支付点差和滑点；限价单锁定价格，但面临无法平仓的风险。调整仓位大小，使维持保证金边界远离你的止损位，而不是紧邻它。如果强平线位于你止损位后方30个基点处，整个问题就变成了哪个价格源先更新。将资金费率视为一种缓慢的触发器。它会在结算期间移动保证金边界，因此一个在入场时安全的仓位可能会在不知不觉中漂移至违约边缘。