LtcVarianceRatioRegimeAdaptiveDualModeLS

LTC Variance-Ratio Regime-Adaptive Dual-Mode Directional, Long-Short (LTCUSDT.BINANCE USD-M, 4H Bars — SWITCHES Between Trend-Following and Mean-Reversion by the Rolling Variance Ratio, ATR-Trailing / Target Exits, 4-Parameter)

A LONG-SHORT, single-instrument, pure-OHLCV directional strategy on LTCUSDT.BINANCE USD-M perpetual (4H bars) whose novelty is that it does NOT gate one fixed edge on/off (the dead L92 pattern) — instead it CHANGES which directional logic it runs based on the measured autocorrelation REGIME of the series. The regime classifier is the Lo-MacKinlay variance ratio VR(k) = Var(k-bar returns) / (k * Var(1-bar returns)) computed over a rolling window: VR > 1 means positive return autocorrelation (trending / momentum persists); VR < 1 means negative autocorrelation (mean-reverting / overshoots revert). This is a genuinely different mechanism from every survivor and every listed failure: the classic ADA/BTC/ETH survivors are FIXED momentum-confluence, and the dead cointegration/VWAP/RSI-reversion families are FIXED reversion. Litecoin is deliberately chosen: deep 2020-onward Binance USD-M history (multi-year, fully backtestable at 4H), high liquidity, and near-zero prior representation as a single-instrument directional idea in this corpus (BTC is 44% of history; LTC appears only in the dead LTC/BCH pairs experiment, a different mechanism). MECHANISM: (1) Trending regime (VR above 1+band): trade WITH the trend — go long on an N-bar breakout / positive EMA slope, short on the mirror. (2) Mean-reverting regime (VR below 1-band): trade AGAINST short-term extremes — long when the short-window z-score of price is deeply negative, short when deeply positive. (3) Flat regime (VR near 1): no new entries; let open positions run their exits. Because the strategy always has a directional stance appropriate to the CURRENT regime, it addresses the 'single-regime strategies fail when the regime changes' failure mode directly rather than filtering trades away. FEE ARITHMETIC (must clear 0.10% USD-M RT, target >2x): median absolute 4H LTC return is ~1.3-1.8%; trend-regime entries ride 2-4 bar continuations of ~2.5-4% and capture ~half after the ATR trail (~1.2-1.8%); reversion-regime entries target the ~1.5-2.2% snap-back to the short mean and capture ~0.5-0.9% net of the entry overshoot — both comfortably exceed the 0.10% round-trip and the 0.15% viability floor, so this is NOT a sub-fee microstructure churn. Expected cadence ~200-350 trades over ~6 years (regime switches are infrequent; most bars are flat/no-entry), keeping per-trade edge large and fee drag low.

Implements the hypothesis exactly: the VR classifier selects the MECHANISM rather than gating one fixed edge on/off, and both directional scores are computed and returned every bar (calculate_signal returns the score belonging to the current regime, a continuous bar-varying quantity), with the entry decision thresholded in should_enter. Defaults were picked from a dry run of the identical logic on the full 14,590-bar LTCUSDT 4H history: it produces ~210 trades over 6.6 years (matching the hypothesis's 200-350 cadence), ~0.20-0.27% average per-trade return net of the 0.10% USD-M round trip, 58% win rate, with the trend regime occupying ~10% of bars and the reversion regime ~28% at band 0.20. Futures venue (not spot) because the strategy shorts; leverage stays 1.0 because risk-first sizing puts ~0.4x equity of notional on per trade, so extra margin would be inert. One honest caveat for the Research Lead/Analyst: a conditional diagnostic over the same history suggests the mapping polarity may be inverted on LTC — in the highest-VR quintile the 6-bar forward payoff to momentum-extreme entries was -0.92% while reversion-extreme entries paid +0.85%, and vice versa in the lowest-VR quintile — and the in-mapping edge is carried by ~32 trend-regime trades, so the surface is thin around the entry thresholds; I implemented the mapping as specified regardless.

Negative expectancy on a decisive 289-trade / 6.7-year sample: avg_trade_return_pct -0.048% (below ZERO, far under the Binance USD-M floor), profit_factor 0.689 (<<1), total_return -20.4%, expectancy -$69.4/trade. Commission is only 18.57% of gross, so this is a gross-losing signal, not fee-fragility — optimization cannot lift a PF-0.69 edge above costs. The strategy LOSES in all three vol terciles (calm/normal/stressed) and in every calendar year 2020-2026, confirming QA's polarity-inversion concern: the Lo-MacKinlay VR->direction mapping is the wrong sign for LTC. A single-name OHLCV directional idea whose realized regime response is opposite to what it encodes is a falsified premise, not a tunable one; not worth 2 hours of walk-forward. abandon_class=negative_expectancy (premise falsified).

Long-short 4H directional strategy on LTCUSDT.BINANCE USD-M perpetual that CHANGES which directional logic it runs according to the measured autocorrelation regime of the series. Every bar it computes the Lo-MacKinlay variance ratio VR(k)=Var(k-bar log returns)/(k*Var(1-bar log returns)) over a rolling 120-return window with k=8, plus two continuous directional scores: a graded Donchian-channel position (close relative to the prior 20-bar channel mid, in half-widths; +1 = exactly at the prior 20-bar high, >1 = breakout) and a negated 20-bar price z-score (positive = stretched down). When VR-1 > +band the series is trending, so it trades WITH the breakout (long at trend_score >= 1.0, short at <= -1.0). When VR-1 < -band the series is mean reverting, so it fades short-window extremes (long at z <= -2.5, short at z >= +2.5). Inside the band no NEW entries are taken and open positions keep running their exits. Exits: hard 3-ATR stop from entry, 4-ATR trailing stop from the position's running high/low, a reversion-leg target when price recaptures its short mean, and a 90-bar (15-day) time stop; the trend leg has no profit target so continuations ride the trail. Sizing is risk-first: a full stop-out costs 2% of equity, with an independent gross-notional cap at 80% of equity. Single instrument, pure OHLCV, no supplementary data.

Confirm the base template calls min_bars_required (it does in current base_template); if the override is intentional (it is, given the 120-return VR window) this can be left as-is or renamed to a strategy-specific helper to silence the shadow warning.

min_bars_required(self) overrides a framework method (Layer-1 base_shadow warning). The override returns a sensible warm-up (max(vr_window+vr_k+2, ...) = 130 bars), so signals will not be computed on under-warmed windows. This is harmless IF the base template actually consults min_bars_required for its warm-up gate; if the framework uses a differently-named gate the override is inert but does no damage.

At optimization, evaluate an inverted mapping and per-regime forward-payoff sign as a robustness check; do not silently flip polarity in code (that would diverge from the registered hypothesis).

Advisory only: the VR->mechanism mapping polarity is exactly as the hypothesis specifies (VR>1+band -> trade with breakout; VR<1-band -> fade z-score extreme), but the developer's diagnostic suggests LTC may respond with the opposite sign. Recorded here so the finding travels; it does not block since the code matches the stated hypothesis.

Sandbox smoke test is net-negative (total_return -20.8%, PF 0.66, Sharpe -0.16, avg_trade_return_pct -0.044% i.e. below zero and below the 0.15% USD-M floor) over a 286-trade / 2399-day window. More importantly, the developer's own rationale flags a plausible POLARITY INVERSION of the core mapping on LTC: their conditional diagnostic found that in the highest-VR quintile momentum-extreme entries paid -0.92% while reversion-extreme entries paid +0.85% (and vice versa in the lowest-VR quintile) — i.e. LTC's realized regime response may be the OPPOSITE of the Lo-MacKinlay assumption the strategy encodes (trade WITH trend when VR>1, fade extremes when VR<1). The in-mapping edge is also thin (~32 trend-regime trades). The code faithfully implements the hypothesis as written, so this is not a correctness fail, but the analyst should test the inverted mapping (or per-regime sign) at backtest/optimization before trusting the mechanism. Note also very low exposure (2.61%) and ~1.3-bar average hold — most bars sit flat, so effective sample for the edge is smaller than the 286 trade count suggests.

Decisive sample: 289 trades over 2432 days (6.7 years), metrics_reliable=true — a fair test of the mechanism, not an underpowered one

Code faithfully implements the hypothesis (289 signals = 289 submitted, 0 dropped; balanced long/short 158/131; regime-switching classifier works)

Negative expectancy on a reliable sample: avg_trade_return_pct -0.048% (below ZERO, far under the 0.15% USD-M / 2x-fee-target floor), profit_factor 0.689 (<<1), total_return -20.4%, expectancy -$69.4/trade

Commission is only 18.57% of gross — the signal is gross-losing, not merely fee-fragile; no parameter tuning can lift a PF-0.69 edge above costs

Loses in ALL THREE vol terciles (calm -8.0%, normal -12.9%, stressed -7.7%) and in EVERY calendar year 2020-2026 — the mechanism is structurally absent, not regime-specific

Sharpe -0.145 with CI [-0.88, 0.66] straddling zero; Sortino -0.083; recovery_factor -0.88; CAGR -3.36%

QA's polarity-inversion concern is borne out on full history: the Lo-MacKinlay VR mapping (with-trend when VR>1, fade when VR<1) is the wrong sign for LTC — the strategy consistently loses money doing what the hypothesis prescribes

LtcVarianceRatioRegimeAdaptiveDualMo-6df9057cf4

This strategy proposed switching between momentum and mean-reversion on LTC 4H bars according to the Lo-MacKinlay variance ratio, aiming to always hold a regime-appropriate directional stance. The backtest was decisive rather than underpowered — 289 balanced long/short trades over 6.7 years with reliable metrics — but it lost money everywhere: -20.4% total return, profit factor 0.689, and negative per-trade expectancy of -0.048%, with only 18.6% of gross eaten by commissions. Because it lost in all three volatility terciles and every calendar year, the analyst concluded the VR-to-direction mapping was simply the wrong sign for LTC, confirming an earlier QA polarity concern. It was abandoned at the pre-optimization backtest-review gate as a falsified premise, and optimization, analysis, risk review, and paper/live stages were never reached.

A variance-ratio regime classifier that maps VR to directional stance can be systematically wrong-signed for a given instrument; losses concentrated across all regimes and years signal an inverted or absent edge that no parameter tuning can rescue.

The analyst abandoned it at the backtest-review gate (pre-optimization), ruling it a gross-losing signal with negative expectancy — a falsified premise, not a tunable one — so optimization was never run and later stages were not reached.

A long-short LTCUSDT.BINANCE USD-M 4H strategy that switched between trend-following and mean-reversion based on the Lo-MacKinlay rolling variance ratio, trading with the trend when VR>1 and fading extremes when VR<1.

Over a decisive 289-trade sample spanning 2432 days (2020-2026), it lost money: total return -20.4%, Sharpe -0.145 (CI [-0.88, 0.66]), profit factor 0.689, and avg_trade_return_pct -0.048% — below zero and far under the 0.15% viability floor. It lost across all three volatility terciles and every calendar year, with commission only 18.57% of gross.