Liquidation-Cascade Fade

VERDICT: INCONCLUSIVE — DATA TOO WEAK (cannot run the pre-registered test)

0/11 gates evaluable. This is NOT a no-edge finding and NOT a g7 (placebo) kill. The archetype could not be tested at the resolution it is defined on, because the required data does not exist on any free source. Honest prior was 30%; it remains unresolved. No deploy, no retire — the archetype is parked pending a real data source.

Why (the decisive data-feasibility result)

The strategy is defined on 5-minute liquidation cascades over 2023-2026. Two hard free-data walls, both probed live on 2026-06-26 against the Hyperliquid public /info API:

1. No free historical liquidation prints. HL exposes no liquidations endpoint (type:"liquidations" -> HTTP 422). recentTrades returns only a live snapshot with no liquidation flag and no history. So the “$5M liquidated in 5 min” trigger has no true source — it can only ever be a proxy (adverse k-sigma move + volume spike). This was a known, accepted caveat going in.

2. THE KILLER: no free historical sub-daily candles. candleSnapshot returns at most ~5000 candles and always the most recent ones — there is no backward pagination for fine intervals (setting endTime to 30/90/180/365 days ago returns EMPTY). Measured lookback ceiling per interval:

   interval	rows	earliest available	usable span
   1m	5159	2026-06-22	3 days
   5m	5032	2026-06-08	17 days (the resolution this strategy NEEDS)
   15m	5011	2026-05-04	52 days
   1h	5003	2025-11-29	208 days
   4h	5001	2024-03-14	833 days
   1d	1273	2023-01-01	1272 days (cannot represent a 5-min cascade)

   Only ~17 days of 5m data exist on the free API. The strategy needs ~3.5 years.

What I refused to do (integrity)

• Did NOT fabricate older 5m data or synthesize candles.
• Did NOT down-substitute daily candles as the cascade signal. A daily bar cannot represent a 5-minute forced-selling overshoot; using it would silently mutate the archetype into a different (daily dip-buy) strategy and invite look-ahead. That would be curve-fitting the test to produce a number, which is exactly what this program forbids.
• Did NOT report a PF as a result. The 17-day window yields only 20 proxy trades — below the 100-trade floor by 5x and far too few for the N>=200 placebo that is “the whole ballgame” here.

The one thing the available data CAN show (illustration only, NOT a verdict)

Proxy detector (adverse >4.0sigma 5m move + volume

3.0x trailing median, 30-min delay, 4h/+2%/-1% exit, 4.5bp taker + 15bp pessimistic post-cascade slippage each side) over the only available 17 days:

• cascades detected: 20
• trades: 20 (need >=100)
• PF: 0.545, win-rate 35.0%, net $-272.52 on $5k
• This sample is statistically meaningless. It is reported ONLY to demonstrate how data-starved the window is — 20 trades cannot pass or fail any gate, and the placebo (the decisive is-it-just-ETH-beta test) cannot be run at all.

11-Gate Table (BASE config, pre-registered before results)

Gate	Status	Detail
g1_min_100_trades	NOT EVALUABLE	20 trades (need >=100) — 5m history capped at 17d (5032 bars) by HL free API; strategy needs 2023-2026 5m. Onl
g2_PF_ge_1.20	NOT EVALUABLE	PF 0.545 on n=20 (NOT EVALUABLE: sample too small)
g3_Sharpe_ge_0.6	NOT EVALUABLE	NOT EVALUABLE: <17d window, no daily equity series
g4_MaxDD_le_12pct	NOT EVALUABLE	NOT EVALUABLE: insufficient history
g5_pos_months_ge_60pct	NOT EVALUABLE	NOT EVALUABLE: <1 month of data
g6_bootstrap_LB_Sharpe_gt_0	NOT EVALUABLE	NOT EVALUABLE: insufficient sample
g7_placebo_beat_p95	NOT EVALUABLE	NOT EVALUABLE: cannot run N>=200 random-entry placebo on 17d
g8_2x_cost_PF_gt_1.0	NOT EVALUABLE	NOT EVALUABLE: no valid base PF
g9_DSR_positive	NOT EVALUABLE	NOT EVALUABLE: insufficient sample
g10_no_component_gt_40pct	NOT EVALUABLE	NOT EVALUABLE: insufficient history
g11_walkforward_OOS_ge_0.9xIS	NOT EVALUABLE	NOT EVALUABLE: cannot split 17d into IS/OOS walk-forward

g7 (placebo) — the kill shot — could NOT be run. Detecting cascades vs random ETH entries over a multi-year 5m series is the only way to separate a real forced-selling-overshoot edge from generic ETH beta / dip-buying. needs N>=200 random entries over multi-year 5m; only 17d available. Until that test is possible, the “cascade fade” cannot be distinguished from “ETH went up over 2023-2026.”

Is it just ETH beta? — UNTESTABLE

Cannot test — beta vs cascade-conditioning requires the placebo, which the data cannot support.

Backtest-vs-live delta (noted, not measured)

Even with proper data, entering 30 min after a cascade means entering into the thinnest part of the book; the 15bp/side slippage assumption is a pessimistic placeholder and live fills could be worse. This compounds the data problem rather than relieving it.

Methodology / data concerns (first-class caveats)

1. No free true liquidation-$ feed -> trigger is a proxy by necessity.
2. No free historical 5m candles beyond ~17 days -> the pre-registered 2023-2026 backtest is impossible on free data. This is the binding constraint.
3. To actually validate this archetype you need a PAID/archival source of either (a) HL liquidation events or (b) historical 5m+ OHLCV 2023-2026 (e.g. a tick/candle data vendor, exchange data dumps, or self-collected forward 5m capture going forward). NOTE: cannot install packages or buy data in this env.

Recommendation

INCONCLUSIVE. Do not deploy, do not retire. If Brent wants this archetype resolved, the next step is a data-acquisition task (archival 5m OHLCV or HL liquidation feed), after which this exact harness (detector + 11 gates + N>=200 random-entry placebo, all already coded in liq_fade.py) can run unchanged. Forward-collecting 5m from now would take many months to reach a >=100-cascade sample.

---

Generated 2026-06-26. Free data only. No fabrication, no curve-fit.