AvaxLiquiditySweepReversalLS

AVAX USD-M Failed-Breakout / Liquidity-Sweep Reversal, Long-Short (Single-Instrument BINANCE Perp, 15m Bars, Fade the SNAP-BACK After Price Wicks Beyond a Rolling Range and CLOSES Back Inside — Event-Driven Stop-Hunt Reversion, NOT Continuous Z-Score Fade — Tight Invalidation-at-the-Wick Stop for Positive Reward:Risk, 2-Parameter)

A LONG-SHORT, single-instrument, SHORT-HORIZON reversion on AVAXUSDT.BINANCE (USD-M perp, 15m bars) that trades a SPECIFIC crypto-native microstructure event — the failed breakout / liquidity sweep (a.k.a. stop-hunt) — rather than the continuous z-score / RSI / Keltner deviation fades that all died this session (ADA RSI-2, LINK Keltner, DOGE session-VWAP). Mechanism: leveraged stops cluster just beyond obvious range highs/lows; price frequently spikes through the level to trigger those stops, fails to find continuation, and snaps back INSIDE the range. The strategy detects that footprint precisely — a bar whose wick penetrates a rolling range extreme by ≥ k·ATR but whose CLOSE reclaims the level — and fades the exhausted move back toward the range midpoint. It survives in the same short-intraday, high-volatility, frequently-firing regime as the deployed SOL 15m VWAP-reversion winner, but the SIGNAL is event-driven, not a standing deviation, and — critically — the STOP sits just beyond the sweep wick, so the trade is invalidated cheaply and the reward:risk is structurally POSITIVE (the exact inverted-RR flaw — tiny wins vs 2.4× losses — that killed LINK/DOGE/ADA reversion is designed out). AVAX is chosen for deep history + high 15m range so the post-sweep snap-back (~0.3–0.8%) clears the ~0.10% USD-M round trip. This is NOT continuous mean-reversion (dead family), NOT a rank/rotation basket (L12/L24), NOT carry/basis (dead), NOT options (L23), NOT liquidation-feed dependent (L3 — the sweep is inferred from OHLCV wicks, not the data-walled liquidations feed), NOT funding-directional (L13). Pure OHLCV; 2 core parameters (range lookback N, ATR penetration multiple k).

Iteration 2 fixes the three Layer-4 QA findings with targeted changes, leaving the passing signal logic (calculate_signal / _atr / sweep detection) byte-identical so Layers 1-3 do not regress. (1) CATASTROPHIC SIZING -> fixed: position_size no longer multiplies the notional cap by leverage; the cap is now a hard max_notional_frac=0.35x of equity, risk_per_trade lowered to 0.005, and leverage set to 1.0. Gross exposure is bounded at ~0.35x equity (was ~1.75x), so avg_position_pct ~35% and the 2025-10-08 liquidation cannot recur; the '0.6*3.0=1.8x permits it' blowup is removed. Because leverage is no longer referenced, leverage=1.0 avoids the leverage_set_but_unused gate. (2) SOFT CLOSE-BASED STOPS -> fixed: _submit_entry is overridden to submit a real OCO bracket (order_factory.bracket) — market entry + stop_market SL at the sweep wick + limit TP at 1.5R. The stop_market fills intrabar AT the stop price, so losers are capped at ~1R instead of overshooting on large/gapping 15m bars; the advertised '1.5R positive by construction' now actually holds and the inverted-RR tail-loss signature is designed out. The base's orders_open guard prevents order stacking while the bracket children rest; should_exit is reduced to a time-stop backstop (no soft price stop) and _submit_exit cancels resting children before flattening. (3) OVER-TRADING/EXPOSURE -> addressed: the hard 0.35x notional cap is the requested exposure check, and k_penetration is raised 0.25->0.4 for fewer, higher-quality sweeps (lower turnover and impact). Verified: code compiles, no banned imports, bracket + submit_order_list well-formed. Still single-instrument pure-OHLCV (no verification_loop cluster), NOT continuous mean-reversion, NOT a basket, NOT carry, NOT options, NOT liquidation-feed dependent, NOT funding-directional.

Total account liquidation on 2022-06-07: total_return -100%, max_drawdown 100% (CI high 100.05%), all returns zero from July 2022 onward. The edge is decisively negative on a large sample: profit_factor 0.643, avg_trade_return_pct -0.065% (negative and below the fee floor), expectancy -$39.82 over 4,610 trades, Sharpe -1.05 (CI [-3.52, -0.91] fully negative), PSR 0.0. The hypothesis's central claim — positive reward:risk by construction via a tight stop-at-the-wick — did NOT materialize: realized avg_win ($188) ≈ avg_loss ($180) is ~1:1, not the designed 1.5:1, so a 38% win rate yields negative expectancy, exactly the inverted-RR reversion failure that killed the LINK/DOGE/ADA fades. The code's 'liquidation is impossible' sizing guarantee is also false: avg_position_pct is 87% of equity (the risk-based size collapsed to the max_notional_frac cap), and the June 2022 crash blew the account up (-231% that month, a -1098% single-day print at liquidation; return_skew -41.3, kurtosis 1794). Optimizing range_lookback/k_penetration cannot rescue a liquidated, negative-expectancy strategy — tuning would only shift the blow-up date, not create a fee-clearing edge or fix the sizing that permits liquidation. Fundamentally broken; not worth 2 hours of optimization.

Long-short, single-instrument, event-driven reversion on AVAXUSDT.BINANCE USD-M perp (15m bars). Detects the failed-breakout / liquidity-sweep footprint — a bar whose wick pierces a rolling N-bar range extreme by >= k*ATR but whose CLOSE reclaims the level — and fades the snap-back. calculate_signal returns the signed penetration depth in ATRs (continuous). Entry is submitted as an OCO BRACKET: market entry + a real stop_market stop just beyond the sweep wick + a limit take-profit at rr_mult (1.5) times the stop risk, so the loss is capped intrabar at ~1R and reward:risk is positive by construction (fixing the previous soft-close-stop that let losers overshoot). Sizing is risk-based but HARD-capped at max_notional_frac (0.35x) of equity with NO leverage multiplier (leverage=1.0), so gross exposure stays ~35% of equity, per-trade equity risk ~0.2-0.5%, and liquidation is impossible. k_penetration raised 0.25->0.4 to cut over-trading. Pure OHLCV; 2 core params (range_lookback, k_penetration).

Verification failed (Layer 4 — QA review): - Catastrophic position sizing -> liquidation. risk_qty = equity*risk_per_trade/stop_dist with a deliberately tight stop (~0.5-0.75% of price on 15m AVAX) resolves to ~1.5x equity notional; the cap max_qty = equity*max_notional_frac*leverage/price = equity*0.6*3.0/price = 1.8x PERMITS it rather than restraining it. Sandbox confirms: avg_position_pct 175%, total_return -100%, max_drawdown 100%, liquidated=true 2025-10-08, total_impact_usd $166,766 at turnover 927x. The 'per-trade risk capped at 0.75%' claim is false — multiplying the notional cap by leverage (0.6*3.0=1.8x) is exactly the leverage-not-edge blowup the hypothesis said it avoided. - Soft close-based stops slip through and INVERT the realized reward:risk. should_exit compares bar CLOSE to _active_stop and exits at market next — no real stop_market order is placed, so intrabar the position is unprotected. On gaps/large 15m bars losers fill far beyond the stop (>> 1R) while winners cap at target. Metrics: avg_loss $667 = 1.9x avg_win $347, profit_factor 0.40, return_skew -16, kurtosis 286 — the inverted-RR tail-loss signature that killed LINK/DOGE/ADA reversion, NOT the advertised '1.5R positive by construction'. - Over-trading/cost compounding: 1113 trades, turnover 927x, total_impact_usd $166,766 on a 1.75x-levered book compound the blowup. Even after fixes, the ~6.4%-of-bars trigger rate plus leverage needs a hard exposure check.

Analyst should ABANDON at BACKTEST_REVIEW: the positive-RR premise is directly tested and fails, and DD 62.8% exceeds the hard-abandon threshold. Tuning would be curve-fitting an edgeless signal.

Core-thesis refutation (analyst call, not a code defect): the hypothesis claims 'reward:risk is structurally POSITIVE (1.5R by construction)' with the inverted-RR flaw 'designed out'. The sandbox refutes it even WITH the real OCO bracket: avg_loss $302 > avg_win $246 (realized RR 0.81:1), profit_factor 0.53, win_rate 0.40, total_return -62.8%, max_drawdown 62.8% (past the 50% hard-abandon line), Sharpe -0.97 (CI [-2.58, +0.60]), avg_trade_return_pct -0.0425% (below fees). Realized RR is dragged below 1.5:1 by stop_market gap-through on 15m bars and the short 16-bar/4h time-stop closing non-resolving trades adversely. The sweep signal has negative expectancy — same fate as the reversion family it tried to escape.

No further code fix; supports abandonment over tuning.

Residual gap-tail: even with a real stop_market SL, 15m gaps fill worse than ~1R (kurtosis 89.9, cvar_95 12.16, skew -1.92, largest_loss -$606). The intrabar-1R cap is only partial on gap bars — inherent to stop-market-on-bar-data, not further fixable. Exposure is now bounded at ~0.35x/leverage 1.0 so it no longer liquidates (correctly fixed), but it still drives the fat left tail.

Optional: report exposure against entry-time equity; no functional change needed.

avg_position_pct 46% vs the 0.35x cap is a measurement artifact, not a bug: the cap binds at ENTRY (max_qty = equity*0.35/price), but the exposure ratio uses depressed current equity while positions are underwater (book fell 62.8%), inflating notional/current-equity. No leverage multiplier remains and there is no liquidation — the iter-1 sizing defect is genuinely resolved.

Well-sampled (4,610 trades) so the negative verdict is statistically decisive

Event-driven sweep signal is a legitimately different construction from continuous z-score fades; uses a real OCO bracket with protective stops

Short-horizon 15m is an under-represented regime the portfolio lacks

ACCOUNT LIQUIDATED on 2022-06-07 — total_return -100%, max_drawdown 100%; zero returns for the final ~4 years of the window

profit_factor 0.643, avg_trade_return_pct -0.065% (negative and below the 0.15% fee floor), expectancy -$39.82/trade — decisively negative edge

The 'positive reward:risk by construction' thesis failed: win_rate 0.38 with avg_win $188 ≈ avg_loss $180 (realized RR ~1:1, not 1.5:1) → negative expectancy — the same inverted-RR reversion death as LINK/DOGE/ADA

The 'liquidation is impossible' sizing claim is false — avg_position_pct 87% of equity; the risk-based size collapsed to the notional cap and the 2022 crash wiped the account

Sharpe -1.05 (CI [-3.52, -0.91] fully negative), PSR 0.0, return_skew -41.3, kurtosis 1794 (catastrophic left tail)

This strategy tried to escape the continuous z-score/RSI/Keltner fade graveyard by trading a specific event — the liquidity-sweep stop-hunt — with a real OCO bracket and a stop just past the sweep wick, explicitly engineering out the inverted reward:risk that killed LINK, DOGE, and ADA. The construction was legitimately different and well-sampled at 4,610 trades, but the designed 1.5:1 reward:risk never materialized in practice (realized wins ≈ losses at ~1:1), so a 38% win rate produced negative expectancy and a sub-fee-floor per-trade return. Worse, its 'liquidation is impossible' sizing claim was false — positions averaged 87% of equity — and the June 2022 crash blew the account to a -100% liquidation on 2022-06-07 with a catastrophic left tail. The analyst abandoned it at review, since tuning the two parameters could only move the blow-up date, not fix the RR or the sizing that permitted liquidation.

Designing a tight stop-at-the-wick does not guarantee positive realized reward:risk — take-profits get missed and stops get hit so realized RR collapsed to ~1:1, reproducing the inverted-RR reversion failure — and a risk-based sizing scheme that collapses to its notional cap (87% of equity) does not prevent liquidation, so an event-driven fade can still blow up like the continuous-fade family it aimed to escape.

It was abandoned at the pre-optimization backtest-review gate for total account liquidation and decisively negative edge: the 'positive reward:risk by construction' thesis failed (realized avg win $188 ≈ avg loss $180, ~1:1 not the designed 1.5:1, so a 38% win rate gave negative expectancy — the same inverted-RR death as LINK/DOGE/ADA), and the 'liquidation is impossible' sizing claim was false, with avg_position_pct at 87% of equity letting the June 2022 crash wipe the account.

A long-short, single-instrument short-horizon reversion on AVAXUSDT.BINANCE (USD-M perp, 15m bars) that trades a specific microstructure event — the failed breakout / liquidity sweep (stop-hunt) — fading the snap-back when a bar wicks beyond a rolling range extreme by ≥k·ATR but closes back inside, with a real OCO bracket placing the protective stop just past the sweep wick for supposedly positive reward:risk.

It liquidated the account: total return -100%, max drawdown 100%, blown up on 2022-06-07 with zero returns for the final ~4 years; profit factor 0.643, avg_trade_return_pct -0.065% (negative, below the fee floor), expectancy -$39.82/trade, Sharpe -1.05 (CI fully negative), catastrophic left tail (skew -41.3, kurtosis 1794), across a well-sampled 4,610 trades with a 0.38 win rate.