Do liquidation cascades actually mark the bottom?
The largest BTC long-liquidation hours were followed by stronger returns, but the residual effect became uncertain once we matched on the size of the drop.
The direct answer
Not reliably from the liquidation print alone.
The largest five percent of BTC long-liquidation hours were followed by a +0.219% average return over the next 24 hours. The all-hour baseline was +0.110%. Read alone, that looks like evidence for a capitulation bounce.
Those hours also contained the largest price drops. After we compared hours with similar same-hour declines, the remaining liquidation effect was not conventionally significant in any drop-size quartile. The largest-drop group still pointed toward a bounce, while the other three groups pointed the other way.
Large liquidation hours are associated with stronger subsequent returns in the uncontrolled BTC sample. This test does not establish that forced selling adds a reliable signal beyond the price drop that accompanied it.
A cascade tells you something important did happen: leverage was closed by force. It does not tell you that every vulnerable position is gone or that the last liquidation print marked the low.
Picture the liquidation chain
A leveraged perpetual position is backed by margin. If losses reduce that margin too far, the position can no longer satisfy the exchange's maintenance requirement. The exchange's liquidation process takes control and closes some or all of it.
A long liquidation creates sell flow. A short liquidation creates buy flow. Unlike a discretionary exit, the order exists because the position crossed a risk threshold, not because its owner chose that moment as fair value.
The cascade comes from feedback:
- Price moves against leveraged positions.
- The first positions reach their liquidation thresholds.
- Their forced orders push price toward more thresholds.
- The next group liquidates into a market with less available liquidity.
The process can accelerate even if no participant changes their opinion. It is the position structure and the exchange's risk engine turning a price move into more market orders.
What a cascade looks like
On a chart, a long-liquidation cascade usually combines a fast downward price move with a cluster of forced-sell prints. A short cascade is the mirror image: price rises rapidly while forced buying appears.
The liquidation feed is only one layer. The more complete footprint includes:
- a sharp price move rather than an isolated liquidation print;
- liquidation notional large relative to that market's own recent activity;
- open interest falling as outstanding positions disappear;
- aggressive perp flow moving in the direction of the cascade;
- nearby book liquidity being consumed or pulled as price accelerates.
A wick and immediate reclaim can look like exhaustion. A close near the extreme can look like continuing pressure. Neither shape proves what happens next, but they describe different price responses to the same category of forced flow.
What liquidations are often confused with
A stop-loss wave
Stops can also turn into market orders and create a fast move. They are not necessarily liquidations. A trader can close voluntarily or through a stop before the exchange takes control.
Any decline in open interest
Falling OI says contracts closed on net. Liquidations can contribute, but so can voluntary exits and offsetting changes. OI measures the stock of outstanding contracts; the liquidation feed reports a subset of forced events.
A single large print
One liquidation can be large without producing a cascade. The defining feature is propagation through price and additional thresholds, not the label attached to one order.
A confirmed bottom
Forced selling can remove leverage and price can still continue lower. New positions can open, spot sellers can remain active, and resting bids can fail. "Leverage was removed" and "the market cannot fall farther" are different claims.
A practical way to read a cascade
The historical study below does not validate a cascade-entry strategy. Once the size of the price drop was controlled, none of the liquidation differences was conventionally significant. These steps organize the event you are seeing; they do not turn the print into a tested buy or sell rule.
1. Identify the forced side. Long liquidations are forced sells. Short liquidations are forced buys. Keep the order direction separate from any later interpretation.
2. Scale the event to its market. Raw notional is not comparable across BTC, ETH, SOL or different volatility regimes. Ask whether the event is unusual for that asset and horizon.
3. Measure the move that caused it. A large cascade usually comes with a large move, and large moves already have different forward-return behavior. The liquidation number must add information beyond what price tells you.
4. Check how much open interest left. A large liquidation print with little OI change describes a different event from a broad contraction in outstanding contracts. Neither measure alone captures the full process.
5. Separate perp mechanics from spot demand. If spot buyers absorb the forced perp selling, the price response differs from a cascade confirmed by spot selling. The liquidation print itself cannot make that distinction.
6. Watch the response after the burst. Reclaiming the cascade area, holding below it or finding no liquidity on a retest are observable outcomes. Do not declare exhaustion before price demonstrates it.
How we tested the claim
We summed Binance BTCUSDT liquidation prints into clock-hour buckets from 2023-02-01 through 2026-08-01 and joined them to BTCUSDT perpetual candles. The same-hour return runs from the hour's open to close. The outcome runs from that close to the price exactly 24 clock hours later.
Hours without the exact future endpoint were excluded, leaving 30,546 BTC hours. Long and short liquidation notionals were ranked separately.
The largest-five-percent and 90th-to-95th-percentile rows are disjoint. The middle row covers the 50th through 90th percentiles, and the quiet row is the bottom half.
The uncontrolled BTC comparison
Long liquidations, or forced selling:
| band | hours | avg forced selling | return that hour | return next 24h | higher 24h later |
|---|---|---|---|---|---|
| quiet half | 15,276 | $0.01M | +0.136% | +0.104% | 51.1% |
| 50th to 90th percentile | 12,216 | $0.21M | -0.022% | +0.095% | 51.9% |
| 90th to 95th percentile | 1,527 | $0.85M | -0.351% | +0.176% | 54.8% |
| largest 5% | 1,527 | $4.01M | -0.740% | +0.219% | 53.0% |
The largest row is followed by about twice the unconditional BTC return. It is also preceded by a much larger same-hour fall than any other row. This table describes the association, but it cannot assign the extra return to liquidations.
Short liquidations, or forced buying:
| band | hours | avg forced buying | return that hour | return next 24h | higher 24h later |
|---|---|---|---|---|---|
| bottom 90% | 27,492 | $0.11M | -0.049% | +0.102% | 51.8% |
| 90th to 95th percentile | 1,527 | $0.82M | +0.331% | +0.290% | 51.5% |
| largest 5% | 1,527 | $3.08M | +0.644% | +0.071% | 49.2% |
The largest short-liquidation hours have the lowest forward return and the only directional hit rate below 50% in this table. The 90th-to-95th-percentile row, however, has the highest mean return. The relationship is not monotonic.
Matching on the size of the drop
We restricted the control to down hours, sorted those hours into four quartiles by the magnitude of the same-hour decline, and then compared the highest and lowest long-liquidation quartiles within each drop-size group.
| drop-size group | high-liq hours | low-liq hours | high-liq next 24h | low-liq next 24h | difference | Newey-West SE | t |
|---|---|---|---|---|---|---|---|
| Q1, biggest drops | 943 | 944 | +0.350% | +0.194% | +0.156 pp | 0.162 pp | 0.97 |
| Q2 | 943 | 944 | -0.048% | +0.121% | -0.169 pp | 0.124 pp | -1.37 |
| Q3 | 943 | 944 | -0.029% | +0.068% | -0.097 pp | 0.105 pp | -0.92 |
| Q4, smallest drops | 943 | 944 | +0.016% | +0.187% | -0.171 pp | 0.117 pp | -1.47 |
The 24-hour outcomes overlap, so the uncertainty estimates use Newey-West standard errors with 24 hourly lags.
The biggest-drop point estimate is consistent with a capitulation bounce, but t = 0.97 is not enough to distinguish it from noise in this sample. The other three point estimates are negative. None crosses the conventional two-sided 5% significance threshold.
One more thing had to be checked before trusting any of this. Ranking hours by raw dollar liquidations quietly favors recent ones, because open interest and prices grew across the sample. Only 1.7% of 2023 hours land in the largest-five- percent band, against 9.0% of 2026 hours. Sorting on dollars therefore sorts partly on the calendar.
So we repeated the controlled comparison ranking each hour against its own trailing 30-day average instead, which asks whether forced selling was unusual for that period rather than large in absolute terms. The biggest-drop estimate is effectively unchanged at +0.16 percentage points. The other three quartiles shuffle between -0.14 and +0.05 points and stay small. The conclusion does not depend on which scale is used.
The control is still not a full causal design. Quartiles contain variation in volatility, trend, time of day and market regime, and the drop being matched on is measured over the same hour that contains the liquidations, so part of it is the cascade's own aftermath. It is enough to show that the uncontrolled table was doing more work than the liquidation claim could bear.
What repeated across BTC, ETH and SOL
The following is the same uncontrolled largest-five-percent comparison in each market:
| market | all-hour baseline | after largest 5% long-liq hours | after largest 5% short-liq hours |
|---|---|---|---|
| BTC | +0.110% | +0.219% | +0.071% |
| ETH | +0.070% | +0.078% | +0.151% |
| SOL | +0.194% | +0.520% | +0.388% |
The long-liquidation row is above baseline in all three markets, but by only 0.008 percentage points in ETH. The short-side mirror fails: short squeezes are followed by below-baseline returns in BTC and above-baseline returns in ETH and SOL.
These are descriptive cross-market checks. Without the same drop-size control and uncertainty estimate in each market, they are not three confirmations of a liquidation mechanism.
A liquidation print only means something next to the size of the move that caused it and the open interest it removed. QuantumFlow shows liquidations on the chart alongside both, which is the comparison this article turns on.
Common mistakes
- Reading the feed as complete. Binance's public liquidation stream does not report every forced close, so printed notional is not total forced flow.
- Calling every spike capitulation. A liquidation burst can occur before more leverage is forced out.
- Ignoring the price move. Most of the apparent bounce can come from the size of the decline that selected the event.
- Treating an OI drop as a liquidation total. OI includes voluntary closing and does not identify every exit mechanism.
- Assuming the short side mirrors the long side. The cross-market table did not support that symmetry.
- Calling a better average return the bottom. A 24-hour endpoint does not reveal the lowest price reached along the path.
- Buying the first forced-sell print. The event is not finished merely because liquidations have started.
What this study does not establish
- Binance's public liquidation feed is not a complete record of every forced close, so printed notional is not total market liquidation notional.
- The analysis uses one exchange and fixed hourly buckets for events that can unfold in seconds.
- Point-to-point 24-hour returns do not show whether the event-hour price was the local low or high.
- None of the controlled quartile differences is conventionally significant.
- The analysis shows association, not that liquidations caused the subsequent return.
---
*Written by Dom at QuantumFlow. Published 5 August 2026. Figures use Binance liquidation prints and perpetual candles for BTCUSDT, ETHUSDT and SOLUSDT from 2023-02-01 to 2026-08-01.*