AVAX Volatility-Spike Wick-Rejection Intraday Fade, Long-Short (BINANCE USD-M Perp, 15m Bars, Pure-OHLCV, Fade Exhausted Stop-Run Overshoots Confirmed by Same-Bar Wick Rejection, Low-Parameter)

A LONG-SHORT, single-instrument, PURE-OHLCV intraday MEAN-REVERSION strategy on AVAXUSDT.BINANCE (USD-M perpetual) using 15-MINUTE bars. It is deliberately modeled on the factory's one demonstrated survivor in this family — the SOL 15m VWAP-reversion long-short that reached paper_stage — but uses a DIFFERENT, orthogonal reversion anchor: instead of distance-from-VWAP, it fades a single VOLATILITY-SPIKE bar whose OWN wick shows the overshoot is already being rejected. This fills the two most under-target quota dimensions after venue: short_1m_15m horizon (4.7% vs a >=10% target) and non-long-only direction (long_only is 84.1% vs a <=55% cap). It is routed to BINANCE (over-weight, but justified): a 15m strategy needs multi-year sub-hourly history, which Hyperliquid structurally cannot provide (~7 weeks at 15m), so routing it to an under-weight venue would guarantee an insufficient-history abandonment. Critically it is PURE-OHLCV with NO supplementary-data dependency — a direct response to the current data-layer reality where FIVE recent hypotheses died on unavailable liquidations, walled account-ratio history, or O(n)-per-bar supplementary scans that blew the 300s Layer-3 limit. There is no funding series, no liquidation feed, no options chain, no cross-venue leg: every input is a bar field, and every rolling statistic is maintained incrementally (O(1) per bar). It is NOT a liquidation-cascade strategy (those just failed on missing liquidation data) — it uses the bar's OWN range and wick as the forced-flow proxy, which is always available. Low-parameter (5 knobs) to resist the overfitting/deflated-Sharpe rejections that dominate the failure log. AVAX chosen over BTC/ETH for fee headroom: its typical 15m spike is ~2-3x a major's, so the ~0.10% round-trip futures cost consumes a small share of the target snapback.

Iteration 6 makes NO logic change, because QA's fee/edge-viability rejection is correct and I verified it independently rather than re-tuning around it. Replicating this file's exact entry/stop/target/time-stop rules directly on the catalog bars (AVAXUSDT.BINANCE 15m, 205,316 bars, 2020-09 to 2026-07, net of 0.10% round-trip) reproduces the backtest almost exactly at the shipped default: entry_threshold 1.5 gives -0.0321% per trade on 1,070 trades, t=-0.46 — matching QA's profit_factor 0.84 / Sharpe -1.11 / -4.2%. That confirms the iteration-5 sizing fix was sound and that the signal, not the plumbing, is what loses. Two diagnostics falsify the premise and neither is reachable by tuning. FIRST, the SHORT leg is significantly negative at every threshold: -0.3158% per trade (t=-3.35, n=360) at th=1.5 and -0.3483% (t=-2.22, n=134) at th=2.0, against a positive long leg. Fading a rejected up-spike on AVAX loses money persistently, so the LONG-SHORT construction the hypothesis mandates — the entire reason it was routed to fill the non-long-only quota — is the broken part; there is no symmetric wick-rejection reversion, only a one-sided drift artifact. SECOND, the one positive component is a single-year artifact: per-year net at th=2.0 is +1.276% in 2021 (n=61) versus -0.290% 2022, -0.142% 2024, -0.035% 2026, and excluding 2021 the full-sample result collapses to zero at every threshold (-0.069% at th=1.5, -0.002% at th=2.0, +0.024% at th=2.2, +0.064% at th=2.5). The long leg is not fading anything, it is collecting bull-market drift on a dip-buy. I deliberately declined to ship the best cell: a 48-cell sweep over entry_threshold x stop_mult x tp_mult x max_hold_bars tops out at t=+1.57 on 373 trades across 5.7 years, is driven by one leg in one year, is flat ex-2021, and at ~65 trades/year would fail the deflated-Sharpe/PBO gate — that is selection on a noise maximum, the exact failure mode dominating the log. My own iteration-5 criterion is met and QA invoked it correctly: RECOMMEND ABANDON on falsified premise, not iterate. Code is left byte-identical below the docstring since Layers 1-3 are green and there is no defect to repair.

Negative-expectancy single-name intraday wick-rejection fade with no edge — not worth 2 hours of optimization. Over 1,088 trades: profit_factor 0.775, Sharpe -0.96 (CI [-1.84, -0.42] fully below zero), total_return -33%, expectancy -$30.5/trade, PSR 0.016, and it loses in EVERY year 2020-2026. avg_trade_return_pct is 0.031% — essentially zero and below the 0.15% Binance USD-M fee floor, with commission 15.6% of gross. The developer's own in-catalog re-measurement (205k bars) falsifies the premise on two independent counts: (1) the mandated long-short construction is the broken part — the SHORT leg loses ~0.3%/trade with real negative t-stats at every threshold; (2) the only positive component is a 2021 single-year bull-drift artifact — ex-2021 the edge is flat/zero at every threshold, i.e. the long leg is a dip-buy collecting bull drift, not a reversion edge. Iteration 6 makes no code change and explicitly recommends ABANDON on falsified premise, and iteration-5's own criterion ('if the per-trade edge remains below cost after this fix, abandon rather than re-tune') is met. This is the L55 zero-survivor intraday fade/reversion family; the cited SOL VWAP-reversion 'survivor' is a different anchor. No parameter tuning flips a negative per-trade edge, and selecting the best threshold cell (~65 trades/yr, t~1.5, flat ex-2021) is selection on a noise maximum that would die at the DSR/PBO gate. Failure pattern: no_edge/fee_edge intraday fade with one-sided bull-drift artifact (L55).

Verification failed (Layer 4 — QA review): - SIZING BLOWUP / LIQUIDATION — observed position size is ~13x the code's own documented cap, and the account was liquidated mid-sample. The docstring/config claim gross notional is 'equity * 0.50 * conviction, capped at 0.75x equity notional, comfortably inside a 1x margin cap,' and position_size does compute notional = equity * risk_frac(0.5) * conviction(<=1.5) * leverage(1.0) <= 0.75x equity per entry. Yet the sandbox reports avg_position_pct = 1038.6% with liquidated=true, liquidation_date 2026-06-01 (mid-sample), total_return -100%. These are mutually inconsistent: a single <=0.75x-notional position on a 1x-margin account cannot trigger a margin liquidation and cannot average 1038% of equity. The only mechanisms that produce this are (a) POSITION ACCUMULATION / PYRAMIDING — should_enter firing repeatedly in the same direction while a position is already open and orders stacking rather than entering only-when-flat — or (b) get_account_equity()/the sizing denominator not reflecting open-position losses. Either is a correctness defect that invalidates the backtest. This is NOT the clean fee/edge failure the caveat frames — a negative-edge strategy sized at <=0.75x bleeds slowly; it does not liquidate to zero mid-sample. - Even aside from the blowup, per-trade economics look negative: avg_trade_return_pct = -0.079% (below the ~0.10% round-trip cost), win_rate 0.47, avg_loss (558) ~2x avg_win (293), kurtosis 258 / skew 14.5 with tail losses dominating. This is the fee-fragile high-turnover fade family the developer flagged — but it cannot be adjudicated until the sizing defect is fixed, since the PnL distribution is contaminated by oversized/liquidated positions.

Verification failed (Layer 4 — QA review): - PREMISE FALSIFIED — confirmed and decomposed; abandon. My iteration-5 fee/edge rejection is verified by the developer against the ACTUAL coded rule on the full sample (AVAXUSDT.BINANCE 15m, 205,316 bars, 2020-2026, net of 0.10% round trip). The shipped default reproduces the sandbox almost exactly (entry_threshold=1.5 -> -0.032%/trade on 1,070 trades, matching PF 0.84 / -4.2%), confirming the signal, not the plumbing, loses. Two diagnostics kill it, neither tunable: (1) THE MANDATED SHORT LEG IS SIGNIFICANTLY NEGATIVE AT EVERY THRESHOLD — fading a rejected up-spike loses ~0.3%/trade persistently (th=1.5 SHORT -0.316% t=-3.35 n=360; th=2.0 SHORT -0.348% t=-2.22). The long-short construction (the reason it filled the non-long-only quota) is the broken part — no symmetric wick-rejection reversion, only a one-sided drift artifact. (2) THE POSITIVE COMPONENT IS A SINGLE-YEAR ARTIFACT — th=2.0 net is +1.276% in 2021 (AVAX's parabolic year) vs negative/flat every other year; ex-2021 the full sample collapses to zero at every threshold (all |t|<1.2). The long leg collects bull-market drift on a dip-buy. The 48-cell sweep tops at a t=+1.57 noise maximum on ~65 trades/year — would fail deflated-Sharpe/PBO.