EthPerpLiquidationCascadeOvershootReversalLS

ETH Perp Liquidation-Cascade Overshoot Reversal, Long-Short Event-Driven (Single-Instrument ETHUSDT.BINANCE USD-M — Fade a One-Sided Forced-Liquidation Spike Back Toward Pre-Cascade Price Using the Liquidations Feed, Reversion Take-Profit + Time/ATR Stop, 3-Parameter)

A LONG-SHORT, EVENT-DRIVEN microstructure strategy on ETHUSDT.BINANCE USD-M whose signal source is the LIQUIDATIONS supplementary feed — a genuinely different information source from OHLCV, funding, or IV, and one the meta-learner favors (strategies WITHOUT supplementary data survive only 3%). This is a materially different FAMILY from the collided short-strangle VRP idea (that whole options short-vol family died verification_loop): here there is no options leg, no vol-selling, no premium harvest — it is a forced-flow overshoot reversal. Mechanism: leveraged longs (or shorts) getting force-liquidated create a self-reinforcing cascade of MARKET orders that pushes price beyond fair value in seconds-to-minutes; once the forced flow exhausts, price mean-reverts. We detect a cascade from the liquidation feed (not from price), fade its direction, and exit on partial reversion. It is distinct from the abandoned 'Perp-Discount Capitulation Reversal' (that was LONG-ONLY SPOT gated on mark-vs-index PREMIUM) — this is long-short FUTURES gated on realized LIQUIDATION NOTIONAL. ETH (not BTC — BTC tag survival is 2%) is chosen for its deeper, more frequent cascades that produce larger overshoots. Only 3 tunable parameters to resist the overfit death that dominates the portfolio.

data_unavailable: backtest data for ? could not be collected after 5 retries: - supplementary:liquidations: got 0, needed 1 The data layer (not the strategy) is the blocker — re-evaluate this hypothesis once collection is fixed / the symbol is backfilled.

CLEAN RESTART 2026-09-04 — this run's verdict history and learning records were removed and it was restarted from verification. Its previous abandonment came from the pipeline, not from the market: the Layer-2 harness mis-bound @staticmethod helpers (fixed), QA issued terminal performance verdicts on an unoptimized smoke test (removed — QA now judges correctness only), and sandbox timeouts came from backtest-slot starvation (fixed). The hypothesis and the strategy code are unchanged. Verify the code on its merits; performance is decided later by the full backtest and the optimizer.

Verification failed (Layer 4 — QA review): - The core overshoot-reversal premise is falsified on the only data the strategy can trade, and the result is reliable (metrics_reliable=TRUE). Over the ETH liquidation window the fade loses: PF 0.19, Sharpe -1.44 with CI [-2.13, -0.59] entirely below zero, probabilistic_sharpe 0.0001, win_rate 0.44, avg_loss $693 vs avg_win $170 (4x), skew -12.4 / kurtosis 161 (a few cascades that kept running dominate; largest_loss -$1,910). By the developer's OWN chosen yardstick — avg_trade_return_pct and active-window stats — the edge is negative: avg_trade_return_pct is -0.043% (below the fee floor AND negative). The short side is worst (short_win_rate 0.167 on 6 trades): fading forced-buying squeezes in an up-trending ETH tape gets steamrolled. Polarity is implemented correctly (verified), so this is not a bug — the cascades in this sample CONTINUE rather than revert. - Structurally unmeasurable: the ETHUSDT liquidations feed spans only ~102 days (2026-04-21→today, per the developer's own data check) and produced just 16 trades over the 328-day sandbox. This hard-caps lifetime tradeable history at ~102 days and lifetime trades at ~16-25 — far below the ~100 floor. Worse, the main backtest runs the full ~6.5-year catalog over which the strategy is flat ~96% of the time (developer-acknowledged), so the 3-window walk-forward has OOS windows with zero trades and the search can't select out-of-sample; the 15-day holdout would carry only ~2 trades. The information source simply lacks the history to validate any edge.

Verification failed (Layer 4 — QA review): - The reversion premise is now falsified at BOTH horizons the hypothesis allows, and the developer's OWN pre-registered abandon criterion is met verbatim. They wrote: 'If iteration 2 still shows a negative avg_trade_return_pct, my recommendation is ABANDON ... the mechanism would then be falsified at both the peak and the exhaustion horizon.' It does: avg_trade_return_pct -0.056% (negative, below the fee floor) on a reliable sample (metrics_reliable=TRUE, 23 trades). Iteration 1 faded at the cascade PEAK and lost; iteration 2 defers to cascade EXHAUSTION (the hypothesis's stated precondition) and STILL loses: PF 0.156 (worse than it1's 0.19), Sharpe -1.58 with CI [-2.51, -0.01] not above zero, PSR 0.004. The geometry fixes did NOT repair the asymmetry — avg_loss $1,181 vs avg_win $240 is still ~5x despite the 1:1 designed R:R and intrabar stop, because cascades continue through the tight stop (skew -7.57, kurtosis 86, largest_loss -$3,173). ETH forced-liquidation overshoots do not mean-revert on this data at either horizon. - The measurability blocker stands in full and is now confirmed unfixable from strategy code. The developer re-verified at the codebase level that the ETHUSDT liquidations feed is forward-only live (pipeline_processor.py:110 'can't be backfilled'), PostgreSQL-only with no parquet path, and Binance's historical liquidationSnapshot is discontinued — lifetime history is capped at ~102 days and grows only in real time. 23 sandbox trades; over the ~6.5-year catalog the strategy is flat ~96% of the time, so the walk-forward OOS windows are near-empty and the 15-day holdout carries ~2 trades. The pipeline cannot validate an edge on this, independent of the negative result.