SuiFundingExtremePositioningWashoutReversalLS

SUI Funding-Extreme Positioning Washout Reversal, Long-Short (SUIUSDT.BINANCE USD-M, 4H Bars — Fade Over-Leveraged Crowds ONLY After Price Confirms Exhaustion, ATR Stop + Funding-Normalization Exit, 3-Parameter)

A LONG-SHORT, single-instrument, price-return-driven contrarian strategy on SUIUSDT.BINANCE (USD-M perpetual), 4H bars, using the perp FUNDING RATE as a structural positioning signal (NOT as the return source). The return comes entirely from the subsequent PRICE move, which the backtest engine models directly — deliberately avoiding the pure-carry trap where a strategy depends on the engine crediting funding cashflows (the reason the factory's surviving funding strategies use funding only as a FILTER/veto, not as PnL). This fills three of the most under-represented buckets at once: long_short direction (14.7% vs ≤55% — the whole book is 85% long-only), the funding_rate data family (the single best-surviving tag at 3%, vs pure_ohlcv/momentum at 1-2%), and an essentially UNTOUCHED instrument (SUI is a liquid Binance USD-M mid-cap that does not appear anywhere in the corpus, unlike the endlessly-retested BTC/ETH/SOL). It is NOT a momentum clone (the overfit graveyard), NOT a crypto pairs/cointegration spread (L96 dead), NOT a beta-neutral/residual reformulation (L95 dead), NOT an options VRP short (risk_reject/no_edge), NOT a Bybit/HL survivor-clone (L94 dead), and NOT a pure funding-carry PnL bet.

Implements the hypothesis exactly: funding as a structural positioning/filter variable (never the return source), price-confirmed exhaustion as the trigger, ATR stop plus funding-normalization exit, long-short on a USD-M perp (shorts require futures, so BINANCE USD-M is the correct venue; leverage stays 1.0 because risk-based sizing never needs more than ~0.4x equity notional, so declaring >1 would be inert). Three headline tunables — funding_lookback, entry_signal, atr_stop_mult — with every code clamp mirrored in _param_bounds. Design checks on the real SUIUSDT 4H catalog + fundingRate history (2023-05 to 2026-08): the composite entry with confirmation fires ~350 times over 3.3 years (no zero-trade risk, no over-selective conjunction — only two conditions), and net-of-0.10%-fees per-trade return is positive across the whole parameter grid I swept (entry 1.25-2.0, stop 1.5-3.0, confirm 2-4, lookback 45-90, holding 6-18), i.e. a region rather than a spike. Defaults sit mid-region, not at the grid maximum. One honest caveat for the analyst: the raw edge is stronger in the first half of SUI's history than the second (per-trade ~0.89% vs ~0.54% with the chosen defaults, still positive in both), and the long side (fading crowded shorts) carries more of it than the short side — a shorter funding baseline (~20 days) is what keeps the recent half positive, which is why the default lookback is 60 observations rather than a slower 90+. Performance is O(funding_lookback) per bar on bounded deques, so the 300s sandbox limit is not at risk.

analyst_decision

Long-short contrarian on SUIUSDT.BINANCE USD-M perp, 4H bars. The funding rate is used purely as a POSITIONING gauge: a rolling z-score of the latest 8h funding print against its own last `funding_lookback` observations (default 60 = ~20 days) says which side of the book is crowded and paying. Crowding alone is not traded — the strategy fades it only once PRICE confirms the crowd is being washed out: the last `confirm_bars` of price action, measured in ATR units, must already be moving in the fade direction. calculate_signal returns a continuous composite every bar, signal = -funding_z * conf, where conf = 2/(1+exp(-conf_gain * fade_dir * price_impulse_atr)) is a smooth (0,2) confirmation multiplier that amplifies the fade when price has rolled over against the crowd and damps it while price still runs with it. Entry when |signal| exceeds entry_signal AND the confirmation term is strictly positive: BUY when funding is extremely negative (crowded shorts) and price is turning up, SELL when funding is extremely positive (crowded longs) and price is rolling over. Exits: funding normalization (|z| < exit_funding_z — the dislocation is gone), ATR stop (atr_stop_mult x entry ATR), ATR take-profit, opposite-extreme flip, and a hard 12-bar (48h) time stop. All PnL comes from the subsequent price move; no funding cashflow is assumed. Sizing is risk-based: risk_pct of equity at the ATR stop, capped at max_notional_frac of equity (leverage 1.0).

Optionally capture entry ATR at the moment should_enter returns a side (or on the first bar the position appears with entry_atr==0), acceptable as-is.

_entry_atr is captured on the first bar where the position is observed as open (bars_in_trade increments), which is one bar AFTER entry, so the ATR used for the stop/TP distance is measured one 4H bar after fill rather than at fill. Minor and does not change the mechanism, but the stop distance is anchored to a slightly-later ATR than the entry price it is measured from.

Sandbox is net-positive (avg_trade_return_pct 0.79%, PF 1.42, WR 0.53 over 330 trades) but the calm-vol tercile is negative (-13.5%, Sharpe -0.39) and the developer's own honest caveat notes the edge is stronger in SUI's first half (~0.89%/trade) than second (~0.54%), sustained recently only by a shortened funding baseline. Sharpe CI [-0.52, 1.53] straddles zero. This is the single-perp funding/supplementary-gated directional class that has repeatedly died negative-expectancy on full history (e.g. AvaxFundingExtremeCrowdFadeLS PF 0.856, LtcCoinMDaily... PF 0.916). Not a code issue and not a block — the analyst should watch full-history per-trade expectancy and the calm-regime drag closely.

Genuine per-trade edge above costs: avg_trade_return_pct 0.499% vs the 0.15% BINANCE USD-M fee floor - NOT fee-dead, commissions are only 5.92% of gross.

Reliable, adequate sample: 328 trades over 1225 days, metrics_reliable=true; entry_diag shows 328 signaled = 328 submitted with 0 drops, so the mechanism functions as coded.

Trades implement the hypothesis: balanced 177 long / 151 short (real long-short), ~21h avg hold on 4H bars, contrarian fade cadence - not a momentum/pairs/VRP clone.

Risk is contained: max_drawdown 15.6% (< 20% futures default), positive in 3 of 4 calendar years, and it makes money in the normal (+11%) and stressed (+16.2%) vol terciles.

Thin edge: profit_factor 1.133 and Sharpe 0.18; Sharpe CI [-0.91, 1.26] straddles zero, so significance is far from assured.

Calm-vol tercile is negative (-13.6%, Sharpe -0.31) - the edge concentrates in normal/stressed regimes and drags in quiet markets.

This is the single-perp funding/supplementary-gated directional class that has repeatedly died in deflation (QA names AvaxFundingExtremeCrowdFadeLS PF 0.856, LtcCoinM PF 0.916); the 12-parameter surface risks overfitting.

Edge decays over SUI's history (developer's own caveat: ~0.89%/trade first half vs ~0.54% second).

Per-trade economics are fine: avg_trade_return_pct 0.7295% vs a ~0.10% USD-M round trip; commissions 5.94% of gross. NOT a fee-edge failure.

Adequate sample/span: 581 trades over 1226 days, metrics_reliable=true, no sizing or min-notional drops.

Code implements the stated mechanism honestly — funding is a point-in-time state variable, not the return source (credited funding is only +$4,476 of $101.6k).

HARD unwaivable: PBO 0.7923 > 0.50 (windows 0.849/0.925/0.603).

DSR 0.000 vs expected-max 2.587 over 225 trials (59 effective) — selected Sharpe 0.439 sits BELOW the best-of-N noise bar; CI [-0.672, 1.438]; programme FDR rejects.

Ten trades, not an edge: top 1 = 18.8% of PnL, top 10 = 90.6%, top 20 = 152.3% (other 561 net negative). Median trade -0.105% vs +0.73% mean.

Post-holdout forward period is decisively negative on 165 trades: -$36,490, PF 0.772, win 35.8%; 2026 -$16,542 with both halves negative.

Holdout 'pass' is window placement: the holdout window alone holds 66% of lifetime PnL; Sharpe 3.05 ± s.e. 3.35.

Sensitivity failed: 6 of 13 params are cliffs over a flat 0.24–1.25 noise field.

Optimizer refuted the premise: entry_signal pushed 1.5 → 0.993, below the whole grid, turning the 'extreme funding' filter largely off (328 → 581 trades, exposure 24% → 40%).

Benjamini-Hochberg at q=0.10 over 294 programme candidates keeps 4. A candidate that does not survive here is not distinguishable from the programme's own noise, however good its individual statistics look.

SuiFundingExtremePositioningWashoutR-22702fcb3c

SuiFundingExtremePositioningWashoutReversalLS faded crowded perp positioning on SUIUSDT 4H bars, using funding purely as a structural crowding state variable and taking its return from the subsequent price move only after price confirmed exhaustion. Its first backtest looked economically live — 328 trades, +17.74%, 0.499% per trade against a ~0.10% round trip, 15.6% max drawdown — so the analyst waved it through to optimization, which roughly doubled trade count and lifted headline return to +97.17% at Sharpe 0.44 with a holdout that nominally passed. Post-optimization scrutiny dismantled it: PBO 0.7923, deflated Sharpe 0.000 versus a 2.587 best-of-N noise bar, six parameter cliffs, PnL concentrated in ten trades, and a decisively negative post-holdout forward period; the holdout 'pass' turned out to be window placement, with that window holding 66% of lifetime PnL. The analyst abandoned it after one iteration; risk review and promotion were never reached.

A headline per-trade edge and a passing holdout mean little when PBO is ~0.79, DSR is zero and 90% of PnL comes from ten trades — and when the optimizer has to disable the core filter (entry_signal 1.5 → 0.993) to find profit, the premise, not the parameterization, is what failed.

The analyst abandoned it at ANALYZING on hard robustness grounds: PBO 0.7923 (windows 0.849/0.925/0.603), deflated Sharpe 0.000 against an expected-max of 2.587 over 225 trials, programme FDR rejection, and a sensitivity phase that failed with 6 of 13 parameters flagged as cliffs. Supporting evidence was that the top 10 trades carried 90.6% of PnL (median trade -0.105%), the post-holdout forward period lost $36,490 on 165 trades at PF 0.772, and the optimizer pushed entry_signal from 1.5 to 0.993 — below the tested grid — effectively switching off the 'extreme funding' filter the thesis rested on.

A long-short 4H SUIUSDT.BINANCE USD-M strategy that used the perpetual funding rate as a point-in-time crowding gauge (never as PnL) and faded extreme positioning only after an ATR-scaled price impulse confirmed the crowd was being hurt, exiting on funding normalization, ATR stop/take-profit, signal flip, or max holding time.

The initial backtest over 1225 days produced 328 trades, +17.74% total return, Sharpe 0.18, profit factor 1.13, max drawdown 15.6% and avg_trade_return_pct 0.499% (well above the 0.15% fee floor), earning an 'optimize' verdict. After optimization it showed 581 trades, +97.17% total return, Sharpe 0.44, profit factor 1.25, avg_trade_return_pct 0.7295%, max drawdown 19.4%, and a holdout that passed on paper (Sharpe 3.05 over 184 days, 85 trades).