Liquidation cascades: why a market can drop 5% in ninety seconds

A candle that drops 5% in ninety seconds and recovers most of it within the hour is one of the most misread events in a leveraged market. Traders search for the headline that caused it and often find nothing, because in many cases nothing external happened. The move was a mechanical consequence of leverage meeting a thin order book.

The mechanics, step by step

A leveraged position is held against collateral. As price moves against it, the margin ratio falls. When it crosses the maintenance-margin threshold, the venue's liquidation engine takes over the position and closes it — not with a limit order at a considerate price, but with an order designed to remove the exposure before the collateral is exhausted. The trader's opinion about a fair price is no longer part of the transaction.

1. Price moves against a band of leveraged positions and reaches their maintenance-margin level.
2. The liquidation engine closes them as market orders, adding pressure in the same direction.
3. That pressure consumes resting liquidity, so the next order fills further away.
4. The new price reaches the next band of maintenance levels, and step 2 repeats.
5. The loop stops when either the resting book absorbs the flow or the leverage in that price region is exhausted.

The feature that makes this violent is that the flow is price-insensitive on both counts. The liquidation engine does not negotiate, and the traders whose stops sit in the same region are also sending market orders. Meanwhile, market makers widen or withdraw precisely when volatility spikes, so the book is thinnest at the moment demand for liquidity is highest.

Why leverage tiers concentrate the damage

Maintenance-margin distance is roughly the inverse of leverage. A 100x position is liquidated about 1% away from entry, a 50x position about 2%, a 25x about 4%, a 10x about 10%. That is why the first 1% of an adverse move is the busiest: it clears the entire top leverage tier at once, and the resulting orders are what carry price into the 2% band.

These are approximations: real maintenance requirements vary by venue and rise with position size through tiered margin schedules, and fees and accrued funding pull the level slightly closer. The shape is the point. Leverage does not scale risk linearly with position size; it scales how little the market has to move before your position stops being yours.

Distinguishing a cascade from a repricing

Both produce a large candle. They resolve differently, and the tells are usually visible within the hour.

The open-interest column is the strongest single tell. A cascade destroys positions, so open interest falls with price. A genuine repricing brings new participants, so open interest holds or grows. If price fell 4% and open interest fell 9%, the market did not change its mind — it ran out of collateral.

What this changes about a trade plan

The practical consequences are about placement and size rather than direction. If cascade risk is real in an instrument, a stop sitting inside the densest estimated liquidation band is a stop that will be tested by mechanics rather than by information. Widening it is not automatically the answer either — the honest response is usually a smaller position with the stop outside the noise, which is the arithmetic in position sizing.

Worked through with numbers: an account risking $200 on an idea whose invalidation is 1.1% away can carry roughly $18,000 of notional. If cascade behaviour in that instrument routinely produces 2% wicks, the 1.1% stop is inside the noise. Moving the invalidation to 2.6% — outside the observed wick range — and keeping the same $200 risk means about $7,700 of notional instead. Same risk budget, a stop that tests the thesis rather than the plumbing, and less than half the exposure.

Cascades are also why funding and estimated liquidation structure are worth watching together. Crowded one-sided funding tells you which side is carrying the leverage; a liquidation heatmap suggests where that leverage would be forced out. Neither predicts the trigger. Together they describe how much amplification is available if a trigger arrives.

The reporting problem

Published liquidation totals are almost always understated. Several major venues throttle their liquidation feeds to roughly one message per second, so a burst of hundreds of closes in the same second reports as one. Figures aggregated from those feeds should be read as a floor and as a relative indicator across time, never as a measured total. That caveat belongs in your notes every time a large number circulates.