LinkEfficiencyRatioGatedTrendLongLegOnly

LINK Efficiency-Ratio-Gated Trend — Long-Short, Trend-Quality Filtered (LINKUSDT.BINANCE USD-M, 4H — Take a Directional Position ONLY When the Kaufman Efficiency Ratio Confirms a CLEAN, Non-Choppy Trend; Ride with a Chandelier Trail; Stand Flat Whenever Price Action Is Inefficient/Choppy; 3-Parameter)

A LONG-SHORT, single-instrument, pure-OHLCV directional strategy on LINKUSDT.BINANCE USD-M whose defining feature is a TREND-QUALITY GATE (Kaufman Efficiency Ratio) rather than another entry-timing trick. The entire trend graveyard here dies the same way: the raw trend edge is real but WHIPSAW in choppy regimes drags avg_trade_return_pct below the ~0.10% fee floor. This design attacks that root cause directly — it only holds a position when the Efficiency Ratio (net directional move ÷ sum of absolute bar-to-bar moves over a window) confirms the market is TRENDING CLEANLY, and it stands entirely flat when price action is inefficient/choppy, which is exactly where trend systems bleed. It is NOT the pending BTC momentum-confluence survivor, NOT my pending BNB pullback / XRP structure-break / DOGE fat-tail / BTC spot regime-timer (different instrument and a different core filter — trend QUALITY, not entry location or dormancy), NOT the dead cross-sectional 'trend-smoothness RANKING' basket (L52 — this is a single-name self-gate, no cross-sectional ranking, no leverage/vol-target inflation), NOT a squeeze breakout (L54), NOT a fade (L53), NOT basis/carry/microstructure (all now confirmed dead across my prior batches). LINK is a deep, multi-year Binance USD-M major that alternates cleanly between strong trends and clear chop — the ideal profile for an efficiency gate — and it is absent from the failure and pending lists. Fills the under-represented long-short bucket (13.3% vs the ≥45% implied by the ≤55% long-only cap) with a simple 3-parameter design that resists overfitting (the #1 killer, 301 deaths).

The reported failure was Layer 1.5 and it was purely a config error: instrument_id was LINKUSDT.BINANCE_SPOT while the hypothesis declares LINKUSDT.BINANCE. I switched instrument_id and bar_type to the declared USD-M perpetual and its 4-HOUR bar type, confirmed LINKUSDT.BINANCE-4-HOUR-LAST-EXTERNAL exists in the catalog with 14,524 bars from 2020-01-17 to 2026-09-02, and left every line of executable logic byte-identical to the version that already cleared Layers 1, 2 and 3 — no import changes, no signal changes, no sizing changes. I had originally routed to spot under the venue rule that long-only no-leverage strategies belong there, but the hypothesis-consistency gate is the binding constraint and is right to be, since fees and settlement differ between contracts; the move is also strictly favourable because USD-M costs 0.10% round-trip against spot's 0.20%. Rather than assume my spot measurements carried over, I re-ran the coded rule on the declared futures series at futures fees. The long leg holds and is robust: all ten chandelier-width × threshold cells are positive at +1.58% to +4.17% per trade with compounded multiples of 1.24x to 5.80x, and er_period is equally flat across 12/16/20/24/30/36 (+3.63% to +4.59% per trade, every value positive), so the declared _param_bounds contain no sign flip. The short leg reproduces the spot finding on futures and stays out: its compounded multiple is below 1.0 in every one of the ten cells (0.23x to 0.96x) while per-trade means sit at statistical noise (every |t| ≤ 0.59), the wins-small-loses-big signature you get from shorting LINK capitulation flushes that snap back. I flag one thing the gate did not raise: the hypothesis is written long-short and this is long-only, so there is a direction-bucket mismatch. I named the class LongLegOnly rather than hide it, because adding the short leg to satisfy the bucket would knowingly convert a 5.80x result into a losing one. If the reviewer needs the long-short form, the honest fix is to re-scope the hypothesis, not to re-tune a leg I have measured dead in every parameterisation.

Single-name pure-OHLCV trend follower below the L82/L63 optimize threshold — not worth 2 hours of optimization. Over 94 trades in 6.6 years: Sharpe 0.74 with bootstrap CI [0.01, 1.48] essentially straddling zero and PSR 0.91 (<0.95), so it will not clear the deflated-Sharpe gate; base Sharpe 0.74 is the ceiling the ~225-trial deflation only lowers. L82's authorize-optimize bar is Sharpe >~1.3 with PF >2.0 and this is 0.74/1.63, matching the LinkFundingCrowdingFiltered (0.58) and BtcFast dual-TF (0.59) trend siblings abandoned this batch. The result is carried by ~5 outlier days (return_kurtosis 10.76: 2023-11 +32%, 2024-12 +24%, 2020-07/08 +17%, 2026-08 +17%) — strip them and it is flat-to-negative, and on only 94 trades that fat tail is not a tunable edge. Risk profile is punishing (max_drawdown_duration 1,183 days / >3 years underwater, DD 22% with CI to 41%, 7 consecutive losses, calmar 0.39), and the recent regime is flat by the developer's own admission (trailing 365d ~1.0x) with negative 2021/2022 and information_ratio -0.62 vs a meaningful buy-hold benchmark, so the holdout would sit in the weak 2025-26 window and lean on the single 2026-08-29 print. It clears the fee floor (avg_trade_return_pct 3.23%) and is cleanly built, so this is a no-edge/insignificance call, not a fee or code defect. Failure pattern: no_edge/insignificant single-name trend below the L82 optimize threshold, outlier-day-dependent, long underwater regime (L82/L63).

Single-instrument, pure-OHLCV trend-capture strategy on LINKUSDT.BINANCE USD-M perpetual, 4-hour bars, whose defining feature is a trend-QUALITY gate rather than an entry-timing trick. Every bar it computes the signed Kaufman Efficiency Ratio — net directional travel over the last er_period bars divided by the total path walked — so one continuous number in [-1, +1] carries the trend direction in its sign and the trend cleanliness in its magnitude. It enters long only when that ratio exceeds er_threshold (the market is walking a near-straight line rather than chopping) AND price sits above a 240-bar (~40-day) trend average, so clean counter-trend bounces inside a downtrend are declined. It stands entirely flat whenever price action is inefficient, which is precisely where trend systems bleed to whipsaw and fees. The only exit is a chandelier trailing stop at the highest close since entry minus chandelier_atr_mult × ATR; an ER-collapse exit was tested and measured to destroy the edge by fragmenting developing trends into fee-paying pieces. Sizing is risk-unit based rather than equity-compounding: a full trip to the initial stop costs risk_pct of equity, with gross notional hard-capped at equity × leverage, and leverage stays at 1.0. All state is maintained incrementally over fixed-length deques, so per-bar cost is O(1).

Verification failed (Layer 1.5 — hypothesis/config consistency) [class=hypothesis_mismatch]: - config.instrument_id is LINKUSDT.BINANCE_SPOT but the hypothesis declares LINKUSDT.BINANCE — same asset, DIFFERENT venue/contract type (fees, settlement and the mechanism's premise differ). Trade the declared instrument.

Clears the USD-M fee floor comfortably: avg_trade_return_pct 3.23%, commission only 1.46% of gross — not fee-fragile

Clean, honest 3-parameter pure-OHLCV construction; developer transparently dropped the measured-dead short leg rather than pad the bucket

Positive full-history record (Sharpe 0.74, +110% return, PF 1.63) with positive alpha 0.052

Sharpe 0.74 with bootstrap CI [0.01, 1.48] essentially straddling zero and PSR 0.91 (<0.95) — will not survive best-of-N deflation

Below the L82 optimize threshold (Sharpe >~1.3 / PF >2.0); this is 0.74 / 1.63, matching two dual-TF siblings abandoned this batch

Edge concentrated in ~5 outlier days (return_kurtosis 10.76); remove them and it is flat-to-negative

Punishing risk profile: max_drawdown_duration 1,183 days (>3 years underwater), max_drawdown 22% (CI to 41%), 7 consecutive losses, calmar 0.39

Recent regime flat by developer's own admission (trailing 365d ~1.0x); negative 2021/2022; information_ratio -0.62 vs buy-hold — holdout would sit in the weak 2025-26 window

LinkEfficiencyRatioGatedTrendLS

Verification failed (Layer 4 — QA review): - FEE/EDGE FAILURE in the evaluable regime. The sandbox (OOS-representative recent window) loses decisively: avg_trade_return_pct -2.09% (negative, below the 0.15% floor), total_return -14.33%, Sharpe -0.85, PF 0.53, PSR 0.156 over 34 trades (metrics_reliable=true — a genuine losing window, not a thin-sample fluke). The full-sample edge (+14.5% CAGR, +1.38%/trade over 226 trades) is real but regime-concentrated in 2020-2024; the developer concedes 2025 is negative and the sandbox (~Aug 2025-Aug 2026) confirms it. The walk-forward OOS windows and 15-day holdout all sit in this recent negative regime, so it cannot clear the min-OOS-Sharpe >= 0.5 or holdout gates.

Verification failed (Layer 4 — QA review): - Recent window still negative after the trend-agreement fix. The filter improved iteration 1's -14.33% to -1.25%, but the sandbox still loses: total_return -1.25%, Sharpe -0.08, PF 0.945, PSR 0.462. avg_trade_return_pct reads +0.29% but USD expectancy is negative (~-$50/trade), so the account loses. Near-breakeven-negative in exactly the regime the OOS windows and 15-day holdout evaluate — cannot clear the OOS Sharpe >= 0.5 gate. - Overfit cliff — the full-sample edge is a single-cell artifact, measured by the developer. At chandelier 4.0 the full sample compounds +42%, but at 3.0/5.0/5.5 it compounds to ~-95%: a +-25% move in one parameter flips the sign — a textbook cliff Phase-1 sensitivity will flag and Phase-2 walk-forward will reject. Nothing is statistically significant (|t|<1.3 over six years), and the recent-window improvement is concentrated in trend_sma_period 240-300, so selecting 240 partly fits the very window used to reject iteration 1. - The long-short construction is broken — the short leg is systematically negative full-sample (-0.93%/trade over 109 trades, t=-0.85), so the design is carried entirely by long-leg LINK beta during 2020-2024, not a genuine two-sided edge. Pure-OHLCV efficiency-gated trend is additionally a ~0%-survival family (L7).

LinkEfficiencyRatioGatedTrendLongOnly