Event study + backtest Free public data 30-day diagnostic

Liquidation Cascade Mean Reversion

After a large, rapid liquidation event flushes leveraged positions in one direction, the forced flow is exhausted — and price should snap back. This site tests the core concept with an event study on real exchange data (Binance OI, Bybit klines, Hyperliquid candles) and a trade-level backtest with real fees.

The strategy idea

What we tested

The idea (from a practitioner strategy doc) has three parts:

Mechanism: forced market orders from liquidations overshoot fair value. Once the leveraged pool on the wrong side is exhausted (OI drops sharply), directional pressure disappears and price snaps back.
Entry: after a flush, fade it — buy the drop (long-liquidations) or sell the squeeze (short-liquidations).
Falsifiable elements: it fails if the flush is genuine fundamental deterioration, or in strong trends where over-leveraged late entrants keep getting flushed in waves.

Scope of this test: directionality of mean reversion after cascades — does price actually pull back, at what horizon, under which conditions? A trade-level build follows only if the concept survives.

Data availability

Can you pull liquidation data for free?

Verified live on 2026-09-10 — the honest answer is partially:

Exchange	Historical liq REST	Real-time liq WS	Historical OI	Historical OHLCV
Binance	GONE (404)	forceOrder WS	futures-data OI (30d)	klines (deep)
Hyperliquid	GONE (rejected)	per-user only	no deep free OI	candleSnapshot (deep)
Bybit	never existed	allLiquidation WS	1h OI (~4.5y)	klines (deep)

Binance removed allForceOrders (deprecated Apr 2026) — only the real-time forceOrder WS stream remains. Its 5m open-interest history is the fine-grained cascade detector (30-day cap).
Hyperliquid rejects the documented liquidations info request; only per-user WS events exist.
Bybit never had a liq REST endpoint — real-time allLiquidation WS only. But its 1h OI history walks back ~4.5 years with zero gaps.

Consequence: you cannot replay actual historical liquidation prints for free. This study uses the OI-proxy cascade — OI destruction + price shock — which is the strategy's own confirming condition. True-print forward validation is Phase 4 (collector not yet deployed).

Methodology

How cascades were detected

A cascade for symbol S is a 15-minute window where:

OI drop ≤ −1% (pooled p1 across the universe; looser 0.5% variant used for sample size)
Price shock ≥ 2 ATR (price-relative) and |move| ≥ 0.5% absolute
Side: price fell → long-liq flush (buyers liquidated); price rose → short-liq flush
6h cooldown; both OI samples within 15m of window edges (guards against coarse-OI artifacts)

No lookahead: cascade detection uses only past/current bars; forward returns are measured from the wick extreme (the flush low/high). Universe: top-30 cross-listed perps by 24h turnover. Realistic fills enter at the open of the first bar after the wick — buying the exact wick is not executable and inflates results ~9×.

Events (30 days)

Total cascades

OI-proxy, 0.5% variant

Long-liq flushes

buy the dip

Short-liq flushes

sell the rip

Real events captured

08-22 crash cluster: ARB −9.6%, USELESS −12.1%, WLD −10.8%, DOGE −8.4% (long-liq)
08-19 cross-market squeeze: SOL/TRUMP/AAVE/DOGE short-liq

Event study

Does price snap back after the flush? (H1)

Forward fade returns from the flush, by side and horizon. The tradable column is open-entry (next bar open after the wick); wick-entry is shown to expose the optimism gap.

Cohort	Horizon	n	Wick-entry	Open-entry (realistic)
Long-liq flush (buy the dip)	1h	32	+4.34%	+0.49%
Long-liq flush (buy the dip)	4h	32	+4.64%	+0.73%
Long-liq flush (buy the dip)	24h	30	+6.80%	+2.80%
Short-liq flush (sell the rip)	1h	31	+1.02%	+0.06%
Short-liq flush (sell the rip)	4h	31	+2.14%	+1.20%
Short-liq flush (sell the rip)	24h	31	+1.19%	+0.26%

Long-liq (buy the dip): net-positive at every horizon; +0.49% @1h, +2.80% @24h on realistic fills (net of 11bps taker). t=2.25 @1h.
Short-liq (sell the rip): flat-to-marginal on realistic fills — no robust edge.
Continuation rate (price moved through the wick within the horizon): long-liq 9.4% @1h (91% snapped back); short-liq 35.5% @1h.

Conditioning

When is the snap-back stronger? (H2 & H3)

H2 — exhaustion scaling: SUPPORTED. Quartile by |ΔOI|: mildest → +1.0%, most severe → +5.4% mean 1h fade (monotonic, Spearman +1.0). Bigger OI destruction → stronger snap-back.
H3 — trend filter: SUPPORTED for long-liq. Counter-trend long flushes +0.57% @1h vs trend-aligned −0.09% (n=28 vs 4). Short-liq inconclusive (n=9 counter).

Trade-level backtest

Phase 3 — engine results

Frozen rules: fade the flush (long-liq → long, short-liq → short); entry at first bar open after the wick; stop 1/2/3 ATR beyond the wick; target 1/1.5/2R; time stop 2/4/24h; first-touch exits (TP-first on same-bar ties); real Bybit fees (taker 11bps RT / maker 4bps RT); 1% risk per trade, $1,000 start, compounded.

Headline configuration

Config

s2 ATR / tp2.0R / 24h

long (fade)

Trades

win rate 50%

Net total (taker)

+15.2R

net PF 2.24

MC P(loss)

2.25%

median +15.3R

Direction split — the decisive finding

Cohort	Configs net+	Both train/test halves +	Best	Worst
Long-only fades	27/27	24/27	+15.2R	+0.2R
Short-only fades	23/27	0/27	+7.9R	-7.5R

Long-only fades are a parameter CLUSTER — 27/27 configs net-positive, 24/27 positive in BOTH train/test halves (24/27 of the long grid survives the time split). Short-only fades fail out-of-sample by time — 0/27 configs positive in both halves (the short edge lived only in the first 2/3 of the sample). Cluster-vs-spike across the full grid: 24/54 (44%) — a borderline cluster (thresholds: spike ≤20%, cluster ≥60%).

Top configurations (by net total R, taker)

Direction	Stop	TP	Hold	n	Win%	Net R	PF	Both halves+
long	2 ATR	2.0R	24h	32	50%	+15.2R	2.24	YES
long	1 ATR	2.0R	24h	32	50%	+14.9R	2.10	YES
long	2 ATR	1.5R	24h	32	56%	+13.7R	2.35	YES
long	3 ATR	1.5R	24h	32	53%	+11.8R	2.13	YES
long	1 ATR	1.5R	24h	32	53%	+10.4R	1.84	YES
long	2 ATR	1.0R	24h	32	62%	+9.9R	2.23	YES
long	3 ATR	2.0R	24h	32	47%	+9.4R	1.75	YES
long	3 ATR	1.0R	24h	32	59%	+8.7R	2.05	YES
short	1 ATR	2.0R	4h	31	48%	+7.9R	1.58	no
long	1 ATR	2.0R	4h	32	56%	+7.5R	1.73	YES
long	1 ATR	1.5R	2h	32	69%	+7.1R	1.99	YES
long	1 ATR	1.5R	4h	32	59%	+7.0R	1.77	YES

Honesty checks

What could be fooling us

Wick-entry is not tradeable. Buying the exact wick low showed +4.3% @1h; realistic next-bar-open entry gives +0.49% — a ~9× haircut. All headline numbers use open-entry.
Medians ≈ half the means (skew from big events). Edge survives 10%-trimming but is smaller.
Concentration: removing the 3 biggest events cuts the wick-basis 1h edge by ~40%; the open-basis edge stays positive.
Random baseline: event 1h +2.7% vs same-symbol random-bar +0.6% (wick basis) — the edge is above the baseline, not an artifact of symbol selection.
Real costs applied: Bybit VIP0 taker 11bps round-trip; maker 4bps improves net by ~7-8bps.

Verdict

What we concluded

Supported

Long-liq cascades (flushes down) do snap back: 91% within 1h, monotonic in OI destruction, net-positive after taker costs at 1h-24h.
Long-only fades survive a full parameter grid + train/test time split (24/27 both-halves+).
The doc's trend filter is directionally right: counter-trend flushes revert more.

Not supported / caveats

Short-liq fades (sell the squeeze) show no robust edge — fails time-split.
30-day diagnostic only (Binance 5m-OI cap): n=63 events total, no walk-forward OOS yet. Per house rules, treat as a promising diagnostic, not evidence.
Survivorship (current listings), no true liquidation prints (OI proxy), deep MAE (long −2.3%, short −6.1%) → wide stops required.

Reproduce

How to re-run everything

# 0. data (free APIs)
python3 phase0_universe.py
python3 pull_oi.py          # Binance 5m OI (30d) + Bybit 1h OI (4.5y)
python3 pull_klines.py      # 15m klines: Bybit, Binance, Hyperliquid (90d)
# 1. events
python3 detect_cascades.py --theta-oi 0.005
# 2. event study
python3 event_study.py && python3 run_analysis.py
# 3. trade-level
python3 trade_engine.py --tag v1
# 4. site
python3 build_site.py

Artifacts: events/cascades.json, events/forward_returns.jsonl, results/trade_results_v1.json, trades/trades_*.jsonl. Full write-up: PLAN.md + RESULTS.md.