AI-generated research, checked but not peer-reviewed. It may contain errors.

E230　 KILL　 Below costsTested · October 2026

Fading or bidding liquidation cascades earns nothing after costs

Buying crypto perpetuals after a liquidation cascade gains 7 to 20 bp before costs, which a 14 bp round-trip cost removes, and resting bids below the market do worse because they fill when the fall continues.

liquidation-cascade forced-flow microstructure adverse-selection perpetual-futures alt-tokens open-interest

Readable summary Full research record JSON recordFeedback on E230

A common belief says a wave of forced liquidations pushes a price too far, so buying the drop pays when it rebounds. We tested it in 30 crypto altcoin perpetuals from 2024 to 2026, buying after sharp drops with market orders and with resting bids below the price. Before costs the average gain was +7 to +20 bp. After a 14 bp round-trip cost it was -7 to +6 bp, and in the validation window that contains the October 2025 crash it was negative in nine of twelve variants. None of the 84 variants passed the multiple-testing screen.

Variants passing screen0 of 84altcoins; majors 0 of 84, stock perpetuals 0 of 60

Rebound before costs+7 to +20 bpbuy after a down cascade, 12 variants

Net after 14 bp cost-7 to +6 bppre-registered gate was +10 bp

Validation window-30.2 bpaverage net of the 12 variants; selection window +9.8 bp

01 · The idea being tested

After a cascade of forced liquidations the price overshoots, so buying the drop, or fading the squeeze, earns the rebound.

Why people believe it

People argue that a liquidation is a market order sent regardless of price. A burst of them pushes the price below where buyers would otherwise trade, so someone who buys into the burst, or waits with a bid below the market, is paid for providing liquidity.

Why it would persist The argument, as stated before testing

People argue that few traders want to buy while a market is collapsing, and that fast liquidity providers pull their quotes in exactly those minutes, so the overshoot would stay available to a patient buyer.

02 · How it was tested

We marked every minute in which price fell several standard deviations in 5 or 15 minutes on heavy selling by aggressive traders, and optionally required open interest to drop by at least 1% (positions being closed). Then we simulated three ways of buying: a market order right after the trigger, a resting bid placed 2 or 4 standard deviations below the price that counts as filled only if the market trades through it, and a standing bid far below the market re-set every minute. The mirror image, selling into up-squeezes, was tested too. Every variant was held for 15 minutes, 1 hour or 4 hours and charged realistic fees. All variants of a group were corrected for multiple testing.

Pass required

Criteria K1-K6, fixed before any return was computed (K6, a fill audit against individual trades, was pre-registered but not run). K2 and K3 and the gate: a variant needs a selection-window day-clustered p-value that survives Benjamini-Hochberg at q = 0.10, mean net above 0 in validation and in hold separately, a full-sample day-clustered 95% interval above 0, at least half its net with a one-minute delay, positive net at stress costs and after dropping the 5 best days, and a mean net of at least +10 bp.

Killed if

No variant survives the screen, or no mean net reaches +10 bp (below practical cost uncertainty). A variant with fewer than 100 events or 50 event-days gets no verdict.

Costs

Taker 7 bp per side (5 bp fee, 2 bp slippage), stress 20 bp per side. Patient and stub entries: 2 bp maker fee on entry plus 7 bp taker exit (stress: 20 bp exit). Funding charged only when a settlement falls inside the hold.

Controls

Benjamini-Hochberg false-discovery control (q = 0.10) over all variants of a group, day-clustered p-values; open-interest confirmation compared with all cascades; one-minute entry delay; results without the 5 best days.

03 · What happened

Why it was rejected

No variant passes the false-discovery control, none is positive in both validation and hold, and none reaches the +10 bp net gate; the gross gain is about the size of the 14 bp round-trip cost.

In the 30 altcoins, 4-sigma 5-minute down cascades occurred 20,908 times on 996 days; the stricter 6-sigma 15-minute rule gave 3,148 events on 481 days, of which 1,178 on 300 days came with a drop in open interest.
Buying with a market order after a down cascade gained +7 to +20 bp gross across the 12 hold-and-rule variants and -7 to +6 bp net after the 14 bp round-trip cost; averaged over the 12 variants, gross was +13.6 bp and net -0.4 bp.
0 of 84 altcoin variants, 0 of 84 majors variants and 0 of 60 stock-perpetual variants pass the Benjamini-Hochberg screen. 12 of 84 altcoin variants have a positive mean net, none reaches +10 bp, and none is positive over the full sample and in both validation and hold.
The validation window, which contains the October 2025 crash, breaks the rebound: the 12 buy variants average +9.8 bp in selection, -30.2 bp in validation and +0.8 bp in hold, and nine of the 12 are negative in validation (the worst, 4-hour hold on the 15-minute 6-sigma rule, -91.5 bp).
At the one-hour hold on the 15-minute 6-sigma rule (3,148 events) the mean net is +1.3 bp (t 0.06), the median trade +19.0 bp and the worst trade -3,436 bp: many small wins and a few very large losses. Without the 5 best days the mean is -15.8 bp.
That rule at 20 bp per side stress costs nets -29.9, -24.7 and -20.8 bp at 15 minutes, 1 hour and 4 hours; with one extra minute of entry delay it nets -1.0, 0.0 and +3.9 bp.
Selling into up-squeezes loses because they continue: across 18 variants with at least 100 events, net -103 to -16 bp (gross -89 to -2 bp).
Resting bids 2 or 4 sigma below the market filled on 21% to 57% of triggers, netted -12 to +2 bp and in 11 of 12 variants earned less gross than a market buy at the trigger; their validation-window results run from -161 to -6 bp. Deep fills are cascades that keep falling.
Open-interest confirmation changes nothing: confirmed cascades net +2.0, -1.7 and +1.5 bp at 15 minutes, 1 hour and 4 hours (t 0.1), against -3.9, +1.3 and +5.2 bp for all cascades; confirmed cascades in validation lost -33 to -97 bp.
No majors (Bitcoin and Ethereum) variant with at least 100 events passes the screen or survives stress costs. Stock perpetuals: the best of 60 variants is +4.0 bp (t 0.5).

Every gate removes the survivors

Altcoin variants (84) passing each gate in turn. 12 have a positive mean net; none stays positive in both validation and hold, none reaches +10 bp, none survives the multiple-testing screen.

Chart data (variants)

Stage	Count
Variants tested	84
Positive mean net	12
Positive in validation and in hold	0
Mean net of at least +10 bp	0
Pass false-discovery screen	0

Even the worst variants of each style lose

Worst net mean among altcoin variants with at least 100 events, by trading style. Market orders into either direction lose most; selling into up-squeezes loses -103.3 bp at worst and buying after down cascades -137.2 bp.

Chart data (bp)

	Worst variant
Market buy after down cascade	-137.2
Market sell after up-squeeze	-103.3
Resting bid below, after down cascade	-12.4
Resting ask above, after up-squeeze	-34.4
Standing bid far below	-8.5
Standing ask far above	-36.7

The rebound disappears in the crash window

Average net of the 12 market-buy variants after down cascades (altcoins), by period. The selection window shows +9.8 bp, validation (with the October 2025 crash) -30.2 bp, hold +0.8 bp; none reaches the +10 bp gate.

Chart data (bp)

	Average net, 12 variants
Selection 2024-01..2025-08	9.8
Validation 2025-08..2026-03	-30.2
Hold 2026-03..2026-09	0.8

04 · What it means

Why it failed Interpretation · not established

The average gain after a cascade, gross +7 to +20 bp, is about the size of the 14 bp round-trip cost, so the net is indistinguishable from zero. It also looks unreliable in the one window with a major crash, where the 15-minute 6-sigma variants lost 26 to 92 bp on average. For resting bids, a plausible reading is adverse selection: a bid several sigma below the market is filled when the price keeps falling, which would fit the lower gross than a market buy at the trigger; fills were not audited. Open interest adds nothing: cascades with a visible drop in open interest did no better than other sharp drops.

Lesson

An average gain that is no larger than the cost of taking it leaves nothing, and in this test the one crash window turned the average negative, so a single stress period can decide a result that looks flat elsewhere. A limit-order test has to simulate the fill rule as well as the exit, because the orders that fill may differ systematically from the ones that do not.

Scope of this result

Covers 1-minute data for 30 altcoin perpetuals, Bitcoin and Ethereum, and 14 US stock perpetuals (2026 only) on the largest USD-margined crypto perpetual venue, January 2024 to September 2026, with simulated fills on candle trade-through. Fills were not audited against individual trades. Tests around funding settlement times are a different mechanism and are not part of this record.

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-registered

Yes: rule and criteria fixed before the test data was examined

Data

Perpetual futures 1-minute candles with taker-buy volume — largest USD-margined crypto perpetual venue, public market-data archive; 1 minute; 2024-01..2026-09 (one extra month for warm-up); public; about 2.3 GB downloaded, aggregates and event windows kept | Open interest history — largest USD-margined crypto perpetual venue, public market-data archive; 5 minutes, fetched for event days only; 2024-01..2026-09; public

Universe

30 altcoins: the largest by December 2023 traded volume among all USDT perpetuals listed before November 2023, including contracts later delisted (no survivorship). Reported separately: Bitcoin and Ethereum perpetuals, and 14 US stock perpetuals (2026 only, short history).

Signal

Down cascade at the close of minute t: return over W minutes at most -k times sigma times sqrt(W), where sigma is the standard deviation of 1-minute returns over the trailing 1,440 minutes; aggressive-sell share of volume over the window at least 55%; window volume at least 3 times its trailing baseline. W in {5, 15} minutes, k in {4, 6}. One event per symbol per 240 minutes. Up-squeezes are the mirror. Open-interest confirmation: open interest down at least 1% over the window.

Execution

Taker variant: enter at the open of minute t+1, exit at the close of minute t+H, H in {15, 60, 240} minutes (also with one extra minute of delay). Patient variant: resting bid at the trigger close times (1 - j sigma sqrt(W)), j in {2, 4}, live for 30 minutes, filled only if a later minute trades at least 2 bp below it (queue position then does not matter), exit H minutes after the fill. Stub variant: bid m sigma below the previous close (m in {6, 10}), re-set every minute, same fill rule. 84 variants per group (stocks: 60).

Costs

Controls

Sample split

Selection 2024-01..2025-08 (60%), validation 2025-08..2026-03 (20%, contains the October 2025 crash), hold 2026-03..2026-09 (20%), split by event time.

Pass criterion

Kill criterion

No variant survives the screen, or no mean net reaches +10 bp (below practical cost uncertainty). A variant with fewer than 100 events or 50 event-days gets no verdict.

Power

No minimum detectable effect was computed. The central altcoin variant (15-minute window, 6 sigma) has 3,148 events on 481 days; single trades range down to -3,436 bp, so the noise per trade is large.

2. Result

Observations

3,148

Instruments

Net mean

+1.3 bp

95% interval

Not reported

Median

+19 bp

Worst

-3,436 bp

Finding 1

In the 30 altcoins, 4-sigma 5-minute down cascades occurred 20,908 times on 996 days; the stricter 6-sigma 15-minute rule gave 3,148 events on 481 days, of which 1,178 on 300 days came with a drop in open interest.

Finding 2

Buying with a market order after a down cascade gained +7 to +20 bp gross across the 12 hold-and-rule variants and -7 to +6 bp net after the 14 bp round-trip cost; averaged over the 12 variants, gross was +13.6 bp and net -0.4 bp.

Finding 3

0 of 84 altcoin variants, 0 of 84 majors variants and 0 of 60 stock-perpetual variants pass the Benjamini-Hochberg screen. 12 of 84 altcoin variants have a positive mean net, none reaches +10 bp, and none is positive over the full sample and in both validation and hold.

Finding 4

The validation window, which contains the October 2025 crash, breaks the rebound: the 12 buy variants average +9.8 bp in selection, -30.2 bp in validation and +0.8 bp in hold, and nine of the 12 are negative in validation (the worst, 4-hour hold on the 15-minute 6-sigma rule, -91.5 bp).

Finding 5

At the one-hour hold on the 15-minute 6-sigma rule (3,148 events) the mean net is +1.3 bp (t 0.06), the median trade +19.0 bp and the worst trade -3,436 bp: many small wins and a few very large losses. Without the 5 best days the mean is -15.8 bp.

Finding 6

That rule at 20 bp per side stress costs nets -29.9, -24.7 and -20.8 bp at 15 minutes, 1 hour and 4 hours; with one extra minute of entry delay it nets -1.0, 0.0 and +3.9 bp.

Finding 7

Selling into up-squeezes loses because they continue: across 18 variants with at least 100 events, net -103 to -16 bp (gross -89 to -2 bp).

Finding 8

Resting bids 2 or 4 sigma below the market filled on 21% to 57% of triggers, netted -12 to +2 bp and in 11 of 12 variants earned less gross than a market buy at the trigger; their validation-window results run from -161 to -6 bp. Deep fills are cascades that keep falling.

Finding 9

Open-interest confirmation changes nothing: confirmed cascades net +2.0, -1.7 and +1.5 bp at 15 minutes, 1 hour and 4 hours (t 0.1), against -3.9, +1.3 and +5.2 bp for all cascades; confirmed cascades in validation lost -33 to -97 bp.

Finding 10

No majors (Bitcoin and Ethereum) variant with at least 100 events passes the screen or survives stress costs. Stock perpetuals: the best of 60 variants is +4.0 bp (t 0.5).

3. Interpretation

Why it failed

Lesson

Scope

4. Reproducibility

Steps

1) Rank USDT perpetuals listed before November 2023 by December 2023 volume; take the top 30 non-major altcoins plus Bitcoin and Ethereum. 2) Download 1-minute candles with taker-buy volume for 2023-12..2026-09 and 5-minute open interest for event days. 3) Compute trailing 1,440-minute sigma and volume baseline; flag cascades with the stated rule, one per symbol per 240 minutes. 4) Simulate taker, resting-bid and stub variants with the stated fill rule and costs. 5) Day-clustered p-values on the selection window, Benjamini-Hochberg over the variants, then validation and hold.

Grade

COMPLETE