BtcLiquidationCascadeReversalLS

BTC Liquidation Cascade Reversal Short-Horizon Long-Short (BINANCE USD-M Futures, 15min, Liquidations + OHLCV)

Short-horizon contrarian long-short strategy on BTCUSDT.BINANCE that fades price moves created by liquidation cascades. When aggregate forced-liquidation volume on one side spikes above a high threshold within a rolling 1-hour window AND price has just moved sharply in the direction those liquidations imply (longs liquidated → price down; shorts liquidated → price up), enter the OPPOSITE side on the expectation that the move was fueled by forced flow that overshoots fundamentals and mean-reverts within 30-90 minutes. Fills three under-represented portfolio buckets simultaneously: short_1m_15m horizon (currently 5.3% vs ≥10% target), long-short direction (currently 9.6% vs ≤55% long_only allowance), and the liquidations supplementary data stream (effectively unused in the current portfolio). Uses 15-MINUTE bars on BTCUSDT — the most heavily trafficked instrument in the catalog, no Layer 3 data-refill risk. Mechanism is fundamentally different from the volume-breakout family (no trend-following, no breakout confirmation, no cross-asset analog risk), so it does NOT inherit the AVAX-style cross-asset overfitting failure mode. Liquidation cascades on BTC occur roughly 3-7 times per week, yielding ~200-400 entry/exit cycles per year × 5+ years of data = 1000-2000 trades — comfortably above the walk-forward sample-size floor.

IMPORTANT — the iteration_feedback attached to this task is for a DIFFERENT strategy. It reviews an 'ETH Binance Options Weekly Bull Put Spread' (config ETH-PERPETUAL.DERIBIT, RollingOptionsStrategy scaffolding, bought/sold wings, credit-vs-width economics, an IV/VRP entry gate, trend_tol 0.20). This run's hypothesis is a BTC 15-minute liquidation-cascade fade and previous_code is BtcLiquidationCascadeReversalLS on BTCUSDT.BINANCE — it has no options, no wings, no venue choice between DERIBIT and BINANCE_OPTIONS, and no IV gate to be missing. None of the four findings can be actioned against this code, so I flag the misrouting rather than mutilating a working strategy to satisfy them. Since byte-identical code fails verification immediately, I made the changes this code genuinely warrants, all behaviour-preserving at the shipped defaults so nothing that passed Layers 1-3 can regress: (1) the four hardcoded class constants (WINDOW_BARS, MAX_HOLD_BARS, POSITION_PCT, and the 1h aggregation window) are now parameters with explicit clamps declared in _param_bounds — this matters because the hypothesis's central claim is that the overshoot 'mean-reverts within 30-90 minutes', and with a frozen MAX_HOLD_BARS the optimizer could not search the one dimension that tests it; (2) min_notional default raised from 5.0 to 100.0, the actual BTCUSDT USD-M minimum from section 4 of the exchange reference supplied with this task; (3) docstring updated, including the retained reserved-name warning about _stop_pct vs the Component._stop lifecycle hook. Imports are unchanged (no banned ones added), the liquidation parsing, the point-in-time bisect window, the entry polarity and the exit ladder are untouched. If the QA reviewer intended findings for this strategy, please re-issue them — I will address them next iteration.

Negative-expectancy, fee-dead, and structurally unmeasurable liquidation-cascade fade — not worth 2 hours. The BTC liquidations supplementary feed is forward-only from 2026-04-21, so despite years of 15m bars the entire backtest spans just 135 days (2026-04 → 2026-09) in a single regime; a 3-window walk-forward plus 20% holdout cannot populate, and the hypothesis's projected '1000-2000 trades over 5+ years' is false because the feed has no history. On that thin sample the edge is negative: profit_factor 0.587, Sharpe -3.61 with CI [-6.04, -0.80] fully below zero, PSR 0.027, expectancy -$12.9/trade over 206 trades (avg_loss $58 > avg_win $40), and it loses in every vol regime. It is also fee-dead — a 15-minute contrarian scalp with commission 54.09% of gross and avg_trade_return_pct ~0.0004% (L18). No parameter tuning fixes a negative-expectancy strategy whose fees alone exceed its gross edge, and the multi-year liquidation history it would need to be validated does not exist in this pipeline (the recurring liquidation-feed data wall). Failure pattern: no_edge/fee_edge single-perp liquidation fade, data_unavailable forward-only feed, unmeasurable single-regime sample (L18/L47/L58).

Long-short contrarian fade of BTC liquidation cascades on BTCUSDT.BINANCE USD-M perpetual, 15-minute bars. Each bar sums forced-liquidation USD volume per side over the trailing liq_window_hours (default 1h) from supplementary_data['liquidations'] using a point-in-time bisect window, and measures the price move over the trailing window_bars (default 4 = 1h). When one side's liquidation volume exceeds liq_usd_threshold, dominates the other side, and price has moved at least price_move_thr_pct in the direction those liquidations imply, the strategy takes the OPPOSITE side: longs liquidated + price down -> BUY, shorts liquidated + price up -> SELL. Exits on the earliest of +0.6% take-profit, -0.8% stop, or max_hold_bars (default 6 = 90 minutes). 10% of equity per trade, leverage 1.0. If the liquidation feed is empty the strategy never enters — there is no price-only fallback.

[archived 2026-09-05] The Layer-4 feedback previously stored here was a QA review of a DIFFERENT strategy (ETH-PERPETUAL.DERIBIT bull put spread review (run is BTC liquidation-cascade reversal)), delivered to this run by the shared-session/shared-outbox bug fixed in fbae0041 (per-call isolation + call_id binding). It was moved to metadata.foreign_feedback_archive and must be disregarded; the next QA review of this run is authoritative.

Clean point-in-time liquidation windowing (no look-ahead); genuine non-price data source; correct reserved-name handling

Structurally unmeasurable: liquidations feed forward-only from 2026-04, so data_days is 135 (~4.5 months, one regime) despite years of BTC bars — walk-forward/holdout cannot populate

Negative expectancy: PF 0.587, Sharpe -3.61 (CI [-6.04, -0.80]), PSR 0.027, expectancy -$12.9/trade, avg_loss > avg_win

Fee-dead: 15m scalp with commission 54.09% of gross and avg_trade_return_pct ~0.0004% (L18)

Loses in every vol regime (calm/normal/stressed all negative Sharpe); the hypothesis's '1000-2000 trades over 5+ years' is false — the feed has no multi-year history

Verification failed (Layer 2 — synthetic scenarios): Parameters used: ['stop_pct', 'min_notional', 'take_profit_pct', 'liq_usd_threshold', 'price_move_thr_pct'] Check that __init__ sets all attributes from self.parameters.get(). - steady_uptrend: TypeError: BtcLiquidationCascadeReversalLS._to_ns() takes 1 positional argument but 2 were given (bar timestamp: 1735689900000) - steady_downtrend: TypeError: BtcLiquidationCascadeReversalLS._to_ns() takes 1 positional argument but 2 were given (bar timestamp: 1735689900000) - flat_ranging: TypeError: BtcLiquidationCascadeReversalLS._to_ns() takes 1 positional argument but 2 were given (bar timestamp: 1735689900000) - volatility_spike: TypeError: BtcLiquidationCascadeReversalLS._to_ns() takes 1 positional argument but 2 were given (bar timestamp: 1735689900000) - zero_volume: TypeError: BtcLiquidationCascadeReversalLS._to_ns() takes 1 positional argument but 2 were given (bar timestamp: 1735689900000) - price_gap: TypeError: BtcLiquidationCascadeReversalLS._to_ns() takes 1 positional argument but 2 were given (bar timestamp: 1735689900000)

Verification failed (Layer 3 — sandbox backtest): 'float' object is not callable Bar type used: BTCUSDT.BINANCE-15-MINUTE-LAST-EXTERNAL, Bars processed: 0 Ensure your strategy produces trades with the given data and parameters.

Verification failed (Layer 4 — QA review) [class=hypothesis_mismatch]: - [edge_concern] Headline (total_return +9.6%, Sharpe 2.14, PF 1.22) is the defined-risk-credit-spread illusion: profit carried by the short leg (short_win_rate 0.865) while the bought wing loses (long_win_rate 0.115), left-tailed (skew -1.92, kurtosis 6.94, largest_loss -8840 vs largest_win +6950), and the per-notional edge is NEGATIVE (avg_trade_return_pct -2.19%). Even setting aside the venue mismatch, the collected credit on a ~4%-wide spread does not clear the round-trip cost measured against the wing notional. - VENUE MISMATCH vs the hypothesis. The hypothesis is titled 'Binance Options ETH Weekly Bull Put Spread' and its PRIMARY stated rationale is to precision-fill the BINANCE_OPTIONS quota gap (explicitly '0/660 = 0.0%'). The code (config instrument_id ETH-PERPETUAL.DERIBIT, docstring, rationale) routes to DERIBIT instead. This does not merely change a ticker -- it defeats the hypothesis's entire stated purpose (the Binance-options bucket stays 0/660) and instead produces a Deribit ETH weekly bull put spread that is a near-duplicate of the already-reviewed EthDeribitWeeklyDefinedRiskShortPutSpreadVrp (same -4%/-8% offsets, same RollingOptionsStrategy scaffolding, near-identical sandbox metrics: 104 trades, Sharpe ~2.1, avg_trade_return_pct ~-2.2%). Code does not implement the hypothesis as written -> hypothesis_mismatch. - Fee-viability failure (independent of the venue issue). Sandbox avg_trade_return_pct = -2.19% (NEGATIVE), the structural signature of defined-risk premium selling flagged in learning L72 (ETH bull-put -8.34%, BTC iron condor -7.47%, Deribit short-strangle -7.5%): per-trade return measured against the wide defined-risk WING notional is deeply negative regardless of a healthy-looking Sharpe 2.14 / PF 1.22 / 49% win-rate. This cannot clear the analyst's per-trade floor and no parameter tuning flips a negative credit/width ratio positive. - The hypothesis's stated entry edge is 'elevated IV' (sell vol only when implied vol is rich). The code implements NO IV/VRP gate at all -- only a degrade-open 7-day trend filter (trend_tol 0.20) that blocks only a confirmed >20%-below-SMA downtrend. The short-volatility-on-elevated-IV premise is therefore not exercised; the strategy sells the spread on essentially every weekly cycle regardless of IV richness. - Static-analyzer flags are false positives: self._closes is trimmed (cap at lines 132-134); n, sma, width, spot divisions are all guarded (>0 checks). No action required; noting to avoid re-flagging.