BtcLiquidationCascadeFadeBidirectional5m

Liquidation Cascade Mean Reversion: Hyperliquid BTC Perp Bidirectional

Crypto-native microstructure strategy that fades extreme liquidation cascades on Hyperliquid BTCUSD perpetual. When forced liquidations cluster ($5M+ liquidated in a 5-minute window) and price has moved >1.5% in that window, the move is partially driven by margin-call forced flow rather than fundamental repricing. Hyperliquid's open liquidation feed gives a real-time view of cascade size that Binance throttles, creating a venue-specific data edge. Strategy enters contrarian (long after short-liquidation cluster + price down; short after long-liquidation cluster + price up) and targets mean reversion within 30 minutes. Bidirectional, single-symbol, short-horizon — fills 4 portfolio quota gaps (Hyperliquid venue, single-symbol scope, long+short direction, short horizon) simultaneously.

Iteration 2 fixes the Layer 3 zero-trades failure (should_enter never fired across 671k bars) with the smallest change that the diagnostic requested ('loosen the entry logic/thresholds'), keeping the fade mechanism, exits, sizing, and imports byte-identical so Layers 1-2 stay green. Root cause: the entry required, within a SINGLE 5-minute bar, both >=$5M liquidated AND a >=1.5% bar move AND directional dominance -- a near-impossible conjunction because a cascade's large move is cumulative over ~15 minutes while any single 5-min bar inside it is usually <1.5%. Three minimal edits: (1) a new window_bars parameter (default 3) widens the liquidation-accumulation window to ~15 min (window_ns = bar_interval * window_bars), so the summed liq USD can actually reach the dollar threshold; (2) the price move is now measured over that SAME window (close vs close window_bars ago) instead of a single open->close bar, aligning the move signal with the liquidation window so the conjunction is reachable; (3) defaults lowered to cascade_usd $1M and move_threshold 0.5% (clamp floors relaxed to 250k / 0.2%). min_bars_required grows to window_bars+2 so the window-return reference bar exists. Together these restore trade generation during the volatile, liquidation-rich periods covered by the feed while preserving the contrarian fade edge for the analyst to evaluate. Direction logic, take-profit/stop/time exits, and risk-based sizing are unchanged.

Do not optimize — the liquidation-cascade-fade mechanism is not backtestable with available data, and the sliver that exists is fee-destroyed. (1) DATA DEAD-END: despite 1,347,840 5-min bars (~6 years) processed, only 16 trades fired, all in 2026-04 to 2026-06 (data_days 7), because the `liquidations` supplementary feed has near-zero historical coverage. A 3-window walk-forward would be almost entirely empty, so the edge cannot be validated. The strategy's own docstring concedes the hypothesis venue (Hyperliquid) has NO backtestable historical liquidation-event feed (only a current-snapshot hl_liquidatable) and HL BTCUSD bars are quarantined, so it already pivoted to Binance USD-M — but the Binance `liquidations` catalog coverage is itself only ~2 months. (2) FEE-DESTROYED ECONOMICS: even in the covered window the result is a clear loser — commission_pct_of_gross 121% (fees alone exceed gross profit), profit_factor 0.24, avg_win $53 vs avg_loss $98, Sharpe -10.5, metrics_reliable=FALSE. A 5-minute fade targeting 0.75% with ~0.10% round-trip taker fees cannot clear costs regardless of parameter tuning. This is a structural data-availability + fee-fragility dead-end, not a parameter problem: no optimization manufactures a multi-year liquidations feed or makes a sub-1% 5-min fade fee-viable. Recommend filing a data-collection/infrastructure ticket for historical liquidation-event coverage (the same gap that the taker-flow and HL-funding strategies hit); until a multi-year liquidations feed exists, this mechanism cannot be evaluated and should be abandoned rather than optimized or iterated.

Bidirectional liquidation-cascade fade on Binance USD-M BTCUSDT 5-minute bars. Over a rolling window_bars (default 3 bars = ~15 min) it sums forced-liquidation USD from the historical liquidations feed and measures the cumulative price move over the same window. When >= cascade_usd ($1M) is liquidated and the window move is >= move_threshold (0.5%), it fades the overshoot: long-liquidation cluster + drop -> BUY, short-liquidation cluster + rise -> SELL. Exits on take-profit (0.75%), stop (1.5%), or a 6-bar (~30 min) time stop. Capital-relative risk sizing capped at 20% notional.

Verification failed (Layer 3 — sandbox backtest): No trades produced Bar type used: BTCUSDT.BINANCE-5-MINUTE-LAST-EXTERNAL, Bars processed: 671040 Diagnostics: should_enter() returned a side 0 times over 671036 evaluated bars -> your ENTRY CONDITION never triggered. Loosen the entry logic / thresholds. Ensure your strategy produces trades with the given data and parameters.

Genuinely novel microstructure mechanism (liquidation-cascade fade) that would fill real portfolio gaps if testable

Clean bidirectional implementation with correct liquidation-side convention

Liquidations feed covers only ~2 months (2026-04 to 2026-06) of a 6-year backtest: 16 trades, data_days 7 — cannot be optimized or walk-forward validated

The hypothesis's named venue (Hyperliquid) has NO backtestable historical liquidation-event feed (code's own admission); the Binance fallback feed is itself only ~2 months in the catalog

Fee-destroyed even in the covered window: commission_pct_of_gross 121%, profit_factor 0.24, avg_win $53 vs avg_loss $98, Sharpe -10.5, metrics_reliable=FALSE

5-min 0.75%-TP fade economics cannot clear taker fees regardless of parameters

BtcLiquidationCascadeFadeBidirectional5m proposed a genuinely novel microstructure mechanism — fading forced-liquidation overshoots — to fill four portfolio quota gaps at once, but the data it depended on did not exist for backtesting. The named Hyperliquid venue has no historical liquidation-event feed, and the Binance fallback covered only ~2 months, so despite processing six years of bars only 16 trades fired, and even those lost money with fees at 121% of gross and a Sharpe of -10.5. Both blockers were structural: no optimization can manufacture a multi-year liquidations feed or make a sub-1% 5-minute fade fee-viable. On its second iteration the analyst abandoned the strategy at the backtest-review gate, recommending a data-collection ticket for historical liquidation coverage before the mechanism can be revisited.

A venue-specific data edge is only testable if the historical data actually exists — a liquidation-cascade fade cannot be evaluated without a multi-year liquidations feed, and a sub-1%-target 5-minute fade is structurally fee-fragile, neither of which any parameter tuning can fix.

The analyst abandoned it at the backtest-review gate before optimization on two structural dead-ends: the liquidations feed has near-zero historical coverage (the named Hyperliquid venue has no backtestable historical liquidation-event feed and the Binance fallback catalog spans only ~2 months), so the edge cannot be walk-forward validated; and even in the covered window the economics are fee-destroyed (commission 121% of gross), since a sub-1% 5-minute fade cannot clear taker fees regardless of parameters.

A crypto-native microstructure strategy that fades extreme liquidation cascades on BTC perpetual 5-minute bars — entering contrarian (long after a long-liquidation cluster with price down, short after a short-liquidation cluster with price up) when forced-liquidation volume and a sharp price move coincide over a multi-bar window, targeting mean reversion within ~30 minutes, bidirectional and single-symbol.

Although ~1.35M 5-minute bars (~6 years) were processed, only 16 trades fired (8 long, 8 short) — all concentrated in ~2 months (2026-04 to 2026-06, data_days 7) — for a -0.82% return, profit factor 0.24, avg win $53 vs avg loss $98, win rate 31%, Sharpe -10.5, commission at 121% of gross, and metrics_reliable=FALSE.