What a cascade chain looks like from the inside
A liquidation cascade is not a single event. It is a chain reaction. When price moves into a concentrated liquidation zone, the forced closure of leveraged positions generates additional sell pressure. That pressure pushes price downward into the next zone, which triggers another wave of liquidations. Each level adds forced selling to a market that is already under stress.
The speed at which this unfolds is what makes cascades consequential. A $200 price drop can become a $2,000 cascade in under 60 seconds. Traders who are watching individual levels miss the systemic picture. The risk is not in any single zone but in the chain that connects them.
This analysis uses data from the Cascade Learner—one of the core models within the Heatmap Engine—to break down how these chains propagate. We examine the structure level by level, measure propagation timing across exchanges, model the probability of continuation at each stage, and assess the aftermath in terms of open interest changes.
Cascade Waterfall: Cumulative Sell Pressure by Level
USD exposure at each cascade level with cumulative forced selling pressure. Levels ordered by trigger price descending.
Figure 1. Cascade waterfall showing forced sell pressure at each price level. The first three levels account for approximately 65% of total cascade volume.
Level-by-level breakdown
The waterfall above shows how a cascade chain distributes forced selling across price levels. It starts at the trigger price—the zone where the initial cluster of liquidations fires—and traces the cumulative sell pressure as the cascade works its way down.
The first three levels typically account for roughly 65% of total cascade volume. This front-loading is characteristic of most cascade events: the densest liquidation zones tend to cluster near each other, often within a 1–2% price range. Once the chain moves beyond these concentrated zones, the volume at each subsequent level drops off.
Beyond level 5, propagation probability drops below 40%. The cascade is running out of fuel. Liquidation zones at these deeper levels are typically smaller, more dispersed, and less likely to generate enough pressure to push price to the next threshold.
The shape of the waterfall carries diagnostic information. A front-loaded cascade—where levels 1 through 3 contain the majority of the volume—tends to be sharp and fast. It resolves quickly because most of the forced selling is concentrated at the top. A distributed cascade, where volume is spread more evenly across many levels, is slower and potentially more dangerous. It can continue propagating for longer, and each new level adds pressure to a market that has already moved significantly.
Cross-Exchange Propagation Timing
Cumulative liquidation volume ($M) across exchanges over 60 seconds from cascade trigger. Stacked by exchange showing propagation delay.
Figure 2. Cross-exchange cascade propagation. Binance leads at T+0, with Bybit and OKX following within 2–5 seconds. Hyperliquid and smaller exchanges lag by 8–15 seconds.
Cross-exchange propagation timing
Cascades do not hit all exchanges simultaneously. The exchange where the trigger occurs—where the initial cluster of liquidations fires—processes those liquidations first, at T+0. Connected exchanges follow based on two factors: their arbitrage latency relative to the trigger exchange, and the depth of their order books at the affected price levels.
In the composite data, Binance typically leads. Its share of perpetual futures volume means it is statistically more likely to be the trigger exchange, and when it is not the trigger, its deep liquidity pools mean arbitrageurs transmit price pressure quickly. Bybit and OKX follow within 2–5 seconds. These exchanges have sufficient liquidity and arbitrage connectivity to reflect cascade pressure almost immediately.
Hyperliquid, as a decentralized exchange, exhibits a characteristic delay of 8–12 seconds. This latency comes from on-chain settlement mechanics and the different structure of its liquidation engine. Smaller and less liquid exchanges lag by 8–15 seconds on average.
This timing differential matters operationally. The delay creates a window—typically 2 to 15 seconds—where a cascade is visible on one exchange before it has fully impacted another. Monitoring cross-exchange propagation in real time provides an early signal of cascade severity and reach.
Propagation Probability by Cascade Level
Probability that a cascade continues to the next level, with 90% confidence interval. Calibrated against 3,200+ historical events.
Figure 3. Propagation probability decays from 92% at level 1 to 8% at level 10. Shaded region shows the 90% confidence interval, which widens at deeper levels.
Propagation probability at each level
The Cascade Learner models propagation probability at each level of a cascade chain. Level 1—the trigger zone, where the initial cluster of liquidations fires—has a 92% propagation probability. Once a trigger fires, it almost always pushes price into the next zone.
By level 3, propagation probability drops to 68%. At this point, roughly one-third of cascades have been absorbed—either by organic buying, by thinner liquidation zones that do not generate sufficient pressure, or by exchange-level interventions such as auto-deleveraging.
By level 7, probability is below 25%. Cascades that reach this depth are rare events. The confidence band widens at each level because more variables affect whether the chain continues: order book depth at that specific price, the timing of arbitrageurs stepping in, whether new limit orders have been placed, and the behavior of market makers during the cascade.
These probability figures are calibrated against over 3,200 historically observed cascade events. The calibration is continuously updated as new events are recorded and fed back into the model.
Key Insight
Cascade depth as risk metric: A cascade with 8+ loaded levels and high propagation probability at levels 1–3 represents a qualitatively different risk than a cascade with 3 loaded levels. Depth and front-loading together determine whether a forced move will be contained or extend beyond initial expectations.
Open Interest: Before vs. After Cascade
Open interest by exchange before and after a composite cascade event. Reduction varies by exchange leverage profile and liquidation zone concentration.
Figure 4. Open interest changes after a composite cascade event. Binance shows the largest absolute reduction (~$1.6B), consistent with its higher share of leveraged perpetual futures volume.
The aftermath
After a cascade completes, the open interest landscape changes materially. The positions that were liquidated are gone. The surviving positions have different characteristics—they were either at safer leverage levels, at different price points, or on exchanges where the cascade arrived later and with less force.
The OI reduction is not uniform across exchanges. Exchanges with higher leverage allowances and more concentrated liquidation zones tend to see larger reductions. In the composite data, Binance typically sees an 8–13% OI reduction after a significant cascade, while exchanges with lower average leverage see reductions closer to 5–7%.
The post-cascade OI profile matters for what comes next. A market with significantly reduced open interest is less susceptible to an immediate follow-on cascade—the positions that were vulnerable have already been liquidated. However, as new positions are established in the aftermath, new liquidation zones form and the cycle begins again.
How the Cascade Learner models these patterns
The Cascade Learner is one of the specialized models within the Heatmap Engine. It tracks cascade events as they occur, measures propagation rates at each level, records cross-exchange timing, and uses this accumulated data to weight future predictions.
The model operates by identifying clusters of liquidation exposure at each price level and estimating the sell pressure each cluster would generate if triggered. It then chains these estimates together, modeling cascades up to 10 levels deep with decreasing confidence at each stage. The probability decay curve shown above is a direct output of this model.
Calibration is ongoing. Each observed cascade event is compared against the model’s prediction, and the weights are adjusted accordingly. The model has been calibrated against over 3,200 observed cascade events, with particular attention to the front-loading characteristics and cross-exchange propagation timing that determine cascade severity.
Methodology
Data derived from the Cascade Learner within Heatmap Engine. Cascade events identified across 22 exchanges with cross-exchange propagation tracking. Probability figures calibrated against 3,200+ historical cascade events. Timing data measured from exchange WebSocket event timestamps with sub-second precision. All figures represent modeled composites derived from aggregate data, not individual events.