Archive / Crypto & DeFi / E224
On-chain forced selling adds nothing beyond the price drop
After large lending-protocol liquidation waves, a long ETH perpetual earned 54.3 bp less over 4 hours than after an equal price drop without a wave; exploits, wallet outflows and depegs showed nothing.
When a lending protocol liquidates borrowers, the collateral is sold at once whatever the price, so many traders expect a sharp drop followed by a bounce. We found 67 large liquidation waves on Ethereum and went long the ETH perpetual 15 minutes after each. The trade made +3.2 bp over 4 hours, but ETH had usually already fallen, and the same trade after similar falls without a wave made +57.6 bp, a dip-buying result of a rising sample rather than an edge. Event minus control was -54.3 bp, interval -132.7 to +26.0 bp. Exploits, wallet outflows and stablecoin depegs showed nothing usable either.
Forced on-chain selling pushes a price past its fair level, so a position taken after the liquidation or the exploit earns the rebound or the continuation.
Why people believe it
A lending protocol liquidates a borrower when collateral value falls below the loan. The liquidator receives the collateral and sells it to realise the profit, whatever the price. The same logic is applied to exploit proceeds being sold, to a long-dormant wallet starting to distribute, and to a stablecoin losing its peg. All of these events are timestamped on a public blockchain, so the forced seller is visible while a derivatives trader can still act.
Why it would persist The argument, as stated before testing
The seller does not care about price, liquidity on the day is thin because the market is already falling, and a position taken into a crash is uncomfortable, so the buyers who would absorb the flow may be slow to arrive.
Forced-flow events were collected from public blockchain data and public incident lists and priced on perpetual futures. The main test counted every liquidation on a large lending protocol, found the hours with unusually large liquidated value, and went long ETH 15 minutes after the hour ended. Each event was compared with five ordinary hours in which ETH had fallen by almost the same amount over the previous 24 hours but no liquidation wave happened. Only the difference to those hours counts, because any fall in ETH is followed by some rebound on its own. The same design was applied to altcoin collateral, and simpler versions to exploits, one wallet's outflows and depegs.
No rejection condition triggered, and the cell survives a Benjamini-Hochberg correction at q = 0.10 within its family (six cells for the main test, nine for alt-collateral, six for exploits).
Main test: the 4-hour event-minus-control interval includes 0, or the raw 4-hour net mean is at or below 0, or the last 20% window is at or below 0, or the result depends on the two best events. Alt-collateral: the 4-hour event-minus-control interval includes 0, or the raw 4-hour net mean is at or below 0, or the last window or the leave-two-out mean is at or below 0. Exploit: net mean at or below 0, interval includes 0, leave-two-out mean at or below 0, last window below 0, or placebo mean within one standard error of the event mean. The second test, with exact event times, fails if the new incidents have a mean at or below 0 or an interval including 0, or the +60 minute entry is at or below 0. Outflow and depeg tests: interval includes 0.
7 bp per side plus the real historical funding (paid by a long, received by a short); stress case 20 bp per side. Simple returns on notional with a floor at -100%.
Wave: for each event, the five non-event hours at least 48 hours from any event whose ETH return over the prior 24 hours was closest to the event's; the statistic is event minus the mean of its five controls, one cluster per event. Alt-collateral: the same with the same token and 72 hours of separation, clustered by token and month. Exploit: placebo decision dates 30 days before, 30 days after and 60 days after for the same tokens, plus the same trade minus a BTC short. Outflow: placebo at the same clock time 7 days earlier and later.
The liquidation wave, compared with equal price drops without a wave, adds -54.3 bp at 4 hours (interval -132.7 to +26.0 bp); the interval includes 0, the last window and the leave-two-out result are negative, and the other forced-flow tests failed their own pre-registered criteria.
- Liquidation waves: the raw 4-hour long ETH trade earned +3.2 bp net (interval -76.7 to +83.0 bp); 1-hour +14.0 bp, 24-hour +108.1 bp (interval -49.0 to +256.7 bp).
- ETH had fallen over the previous 24 hours before 60 of the 67 waves (median -7.3%). The matched hours, with an equal prior fall and no wave, averaged +57.6 bp over 4 hours (the controls overlap heavily and the event-clustered intervals understate the error), so the wave itself added -54.3 bp (interval -132.7 to +26.0 bp); at 1 hour -3.5 bp, at 24 hours -44.8 bp (interval -196.6 to +91.2 bp).
- The 4-hour difference by time window was +8.1 bp in the first 40 waves, -246.0 bp in the next 13 and -54.7 bp in the last 14. Dropping the two best events left -78.9 bp. The 10 waves above $20M of liquidated value gave -66.8 bp. No cell survived the multiple-testing correction.
- Alt-collateral liquidations (138 token-hour events, 95 token-month clusters): the long earned +48.9 bp at 4 hours (interval -57.0 to +171.3 bp) and +250.6 bp at 24 hours, but that is not an edge: matched controls with the same token and a similar prior fall averaged +301.5 bp at 24 hours, so event minus control was -50.8 bp (interval -261.9 to +175.2 bp); at 4 hours it was -38.6 bp (interval -133.5 to +61.4 bp), and -79.1 bp in the last window.
- Exploits, day-dated (60 events with a perpetual, one per token per 7 days): the 24-hour short earned +85.0 bp (interval -261.7 to +369.2 bp, which includes zero), but -9.0 bp without the two best events and -257.0 bp in the last 12 events. Placebo dates gave -54.5 bp (30 days before), -15.9 bp (30 days after) and +38.4 bp (60 days after) at 24 hours. The worst trade was -7,137 bp, a loss of 71.4% on the short; the token peaked 178% above entry inside the hold.
- Why the day-dated short earned anything: the token fell by 168.4 bp on average over the hold (interval -461.1 to +182.1 bp, mean change of price, not return of the short), the short paid 69.4 bp in funding (interval -131.4 to -20.2 bp), and the 14 bp cost is the rest: 168.4 - 69.4 - 14.0 = +85.0 bp. Derived here from the three components.
- Exploits with a known attack-transaction time, short 30 minutes later for 4 hours: the price had already fallen 175 bp in the first 30 minutes in the earlier 9 incidents (interval -312 to -61 bp). The 7 new incidents averaged -166 bp (interval -354 to +39 bp), with a negative 4-hour mean at every entry delay, which met the rejection criterion of the second test, with exact event times.
- Outflows from one long-dormant wallet (12 event days): a 24-hour BTC short earned -42 bp (interval -206 to +114 bp); the same short at the placebo times earned -47 bp, so the outflow added about +5 bp, at a minimum detectable effect of about 250 bp.
- Stablecoin depegs (13 priced episodes in 12 clusters, 4 of them the same weekend): a short of the issuer's governance token earned +18 bp at 1 hour (interval -114 to +175 bp), +196 bp at 4 hours (interval -60 to +548 bp) and -328 bp at 24 hours (interval -699 to +61 bp), because the tokens rebounded within a day.
- Across about 110 cells (27 pre-registered primaries) none passed both the multiple-testing correction and an out-of-sample check.
The bounce is there, but so it is without a wave
Mean 4-hour net return of a long after the liquidation wave, and the wave minus matched hours with an equal prior fall and no wave. Both raw means are positive (+3.2, +48.9 bp); both differences are negative, with intervals that include 0 (-54.3 and -38.6 bp).
Chart data (bp net, 4-hour hold)
| Mean | 95% low | 95% high | |
|---|---|---|---|
| ETH long after wave (n 67) | 3.2 | -76.7 | 83 |
| ETH wave minus matched control | -54.3 | -132.7 | 26 |
| Alt token long after wave (n 138) | 48.9 | -57 | 171.3 |
| Alt token wave minus matched control | -38.6 | -133.5 | 61.4 |
Later windows are negative
ETH wavesAlt-collateral waves
Event-minus-control mean by time window (first 60%, next 20%, last 20% of events by date). ETH waves: 40, 13 and 14 events. Alt-collateral waves: 82, 28 and 28 events. Only the first ETH window is above zero, at +8.1 bp.
Chart data (bp, 4-hour event minus control)
| ETH waves | Alt-collateral waves | |
|---|---|---|
| First 60% | 8.1 | -16.1 |
| Next 20% | -246 | -64.1 |
| Last 20% | -54.7 | -79.1 |
Exploits, outflows and depegs: intervals around zero
Mean net return of the short after each forced-flow event with its 95% interval. All five intervals include 0. Holds are 24 hours unless stated.
Chart data (bp net per trade (short side))
| Mean | 95% low | 95% high | |
|---|---|---|---|
| Exploit, day-dated, 24 h (n 60) | 85 | -261.7 | 369.2 |
| Exploit, exact time, 7 new, 4 h | -166 | -354 | 39 |
| Wallet outflow, short BTC (n 12) | -42 | -206 | 114 |
| Depeg, governance token, 4 h (n 13) | 196 | -60 | 548 |
| Depeg, governance token, 24 h (n 13) | -328 | -699 | 61 |
Why it failed Interpretation · not established
Liquidation waves arrive in the middle of a fall: ETH was already down a median 7.3% before 60 of the 67 waves. The matched hours earned +57.6 bp, which looks like dip-buying in a rising sample, not an edge: the controls overlap heavily and the event-clustered intervals understate the error. The wave added nothing to it. Forced liquidation selling appears to be absorbed within the hour, before the first entry. For exploits, the market appears to price the news within about 30 minutes, and what is left is mostly paid away in funding, since the day-dated short earned 85.0 bp only after 69.4 bp of funding paid and a 14 bp cost. This is interpretation: the data do not show who bought, or when.
Lesson
A rebound after a forced-selling event only counts as evidence if equally large falls without the event rebound less. Here the matched hours earned more, so an event-day average reports dip-buying in a rising sample and attributes it to the event; that control rebound is itself not established as a real effect. Two method points follow. Event dates given only by day cannot test a hold of minutes or hours, because most of the reaction happens before a day-level entry. A result from a small sample (9 events) needs new events before it means anything; here the new events averaged -166 bp.
Scope of this result
Covers one large Ethereum lending protocol (two versions) for ETH collateral and 12 altcoins, perpetual futures hedged by no other leg, entries 15 to 60 minutes after the event hour or at day level, holds of 1 to 24 hours, December 2020 to October 2026. Exploit and depeg results rest on 60, 16 and 13 events at most, and the outflow result on 12 event days, with minimum detectable effects of 250 bp or more; they do not exclude smaller effects.
A rejected experiment applies to the tested conditions and does not prove an idea can never work. bp = basis point, 0.01%.
1. Test design
Pre-registeredDataUniverseSignalExecutionCostsControlsSample splitPass criterionKill criterionPower2. Result
ObservationsInstrumentsNet mean95% intervalMedianWorstFinding 1Finding 2Finding 3Finding 4Finding 5Finding 6Finding 7Finding 8Finding 9Finding 103. Interpretation
Why it failedLessonScope4. Reproducibility
StepsGrade