What followed past liquidation cascades?

Event Study · Descriptive

A descriptive event study of forward price moves after large long- or short-liquidation bursts. It reports the observed distribution, sample size, confidence interval, and an unadjusted permutation p-value. It is not a trade signal.

Coverage: loading…
This is one growing realized-liquidation stream, not a market-cycle sample. Cells require at least 5 observations. The “p<0.05 screen” label additionally requires n≥8, but p-values are unadjusted for the many cells shown. Treat all results as descriptive.
Current context
Loading…
Realized-event method

Executed liquidation prints are grouped into 1-minute bins. A bin becomes a cascade event when its notional is at or above both $25,000 and the symbol’s 90th-percentile bin. The dominant side labels the event: long-liq means longs were force-sold; short-liq means shorts were force-bought.

Forward returns use the nearest derivatives.price observation within about 2,000 seconds at +5 minutes, +15 minutes, and +1 hour. The 95% interval is a normal approximation for the mean. The two-sided permutation p-value compares the absolute event-window mean with 4,000 samples drawn from random price windows at the same horizon.

The realized-event endpoint is cached for 5 minutes. Its price observations are typically spaced about 400 seconds apart. The historical proxy is a generated static study, served through a 1-hour cache, and changes only when the study file is regenerated.

Mean forward move by horizon & side
whiskers = 95% CI of the mean · dashed = 0%
All reportable cells (n≥5)
HorizonSidenmean %median % up-freq95% CIperm pverdict
How to read this. long_liq = a cascade where longs were force-sold (down pressure); short_liq = shorts force-bought (up pressure). A verdict of not distinguishable from random windows means the observed forward move is within what random windows produce — do not act on it. Only p<0.05 screen (unadjusted) cells (with n≥8) are worth a second look, and even those deserve skepticism: the p-values are unadjusted for the many cells shown.

Historical OI-cascade proxy

Proxy · 4h bars

This separate static study covers 14 symbols on a 4-hour grid from 2024-03-19 through 2026-03-19 using Bybit base-unit open interest and price. A long-liq proxy is a fresh 24-hour OI drop below −10% combined with a price move below −5%; a short-liq proxy uses the same OI drop with price above +5%. Returns begin at the next bar open and are measured at +4h, +12h, +24h, and +48h. These are inferred deleveraging events, not executed-liquidation records.

—
PROXY — read this. Cascade events here are inferred from sharp OI drops in 2 years of real OI+price data — they are NOT realized liquidation events, and this is 4h-bar granularity. Binance retired its historical liquidation API, so no free 2yr real-event history exists; the OI-collapse footprint is the honest stand-in for the same forced-deleveraging mechanism. Every cell below shows n + 95% CI + permutation p-value and is labeled not distinguishable from baseline when the drift does not beat a random-bar baseline. The live real-events section above is separate and still accruing.
Pooled mean forward move by horizon (all 14 majors)
whiskers = 95% CI of the mean · dashed = 0% · green outline = p<0.05 screen (unadjusted)
Pooled cells (both sides, all horizons)
HorizonSidenmean %median % up-freq95% CIperm pverdict
Regime split — real 30-day BTC trend at the cascade bar
bear = BTC 30d trend < 0 · bull = trend ≥ 0. This is a genuine 30-day regime, not the live 7-day proxy.
RegimeSideHorizonnmean %median % up-freqperm pverdict