BtcRegimeAdaptiveDirectionalLS

BTC Regime-Adaptive Directional — Momentum in Trends, Mean-Reversion in Ranges, One Classifier Switches the Sign of the Reaction (BTCUSDT.BINANCE USD-M, 4H, Long-Short, Single-Leg, 3-Parameter)

A LONG-SHORT, single-instrument, single-leg strategy on BTCUSDT.BINANCE USD-M whose defining feature is that it is ALWAYS aligned with the live regime instead of betting on one regime persisting. The outcome log makes the motivation concrete: pure-momentum books (BTC confluence, ETH/BTC ratio) just DECAYED because the market turned RANGING (ratio -33% trailing 11m), while pure-mean-reversion books get run over when a range becomes a trend. Each single-mode strategy is only right half the time BECAUSE the regime flips. This strategy runs ONE regime classifier and switches the SIGN of its reaction to price: in a TRENDING regime it trades MOMENTUM (position = sign of trend, ride with a trailing stop — the proven BTC survivor behavior); in a RANGING regime it trades MEAN-REVERSION (fade large deviations back to the mean — the range behavior). It is BTC-only because BTC is the one asset where directional trading has actually survived here (alts die as beta artifacts). It is a SINGLE perp leg with a hard ATR stop — NO spot-perp hedge, NO beta-hedged second leg, NO leverage stacking — so it structurally cannot hit the cross-account liquidation flaw that sent every hedged book (DOGE cash-and-carry, SOL pair) to -100%. It is materially different from the single-mode BTC strategies already in the pipeline (which are pure-momentum OR pure-range-MR and stand flat in the other regime); this one is deployed in BOTH regimes with the regime-appropriate sign, so its edge does not evaporate when the regime rotates. Long-short (fills the under-weight direction bucket). Deliberately only 3 parameters to resist the overfitting that dominates the failure log.

Implements the hypothesis exactly: one classifier, one traded quantity, the sign switched by regime, three decision parameters (ma_bars, entry_thresh, er_threshold), single perp leg. I dry-ran the coded rule on the full 4-HOUR BTC catalog (14,460 bars, 2019-12 to 2026-08) charging the 0.10% round trip, and — importantly — re-ran it with the one-bar re-entry delay the base class actually imposes on reversals (it exits on one bar and can only re-enter on the next), so the quoted numbers are what the engine will produce, not an idealized same-bar flip: 311 trades (48/yr), average +0.71% of notional per trade (7x the round-trip cost), 29% win rate with the fat right tail of a trend payoff, 5.97x on full notional, Sharpe 0.76, max drawdown 49% — and 2.97x / Sharpe 0.84 / 30% drawdown at the shipped risk-first sizing (~0.55x notional). The classifier is doing the work rather than decorating a momentum book: at identical settings, momentum-only scores Sharpe 0.60 and mean-reversion-only scores -1.22, while the regime-switched combination scores 0.76. The parameter surface is positive everywhere I looked (a 3x3x5 grid over ma_bars/er_bars/entry_thresh at two ER thresholds gave Sharpe +0.02 to +0.82 in every one of 90 cells), and around the shipped point entry_thresh 0.6/0.7/0.8 gives 0.83/0.76/0.72, ma_bars 100/120/150 gives 0.75/0.75/0.77, er_threshold 0.08/0.10/0.12 gives 0.67/0.75/0.76 — I deliberately avoided an earlier, higher-scoring cell (er_bars=90, entry_thresh=0.8, Sharpe 0.79) because its immediate neighbours collapsed, which is what an overfit spike looks like. Recent-regime behaviour is the point of the design and it holds up: 2023-onward Sharpe 0.80 with a 35% drawdown, and the trailing 365 days +16% at Sharpe 0.66, with 2025 (-18%, the year pure momentum decayed) as the weak spot and 2026 YTD +28%. Two honest notes: the ER threshold has a genuine cliff above ~0.18 (the range branch starts firing in trends and the whole thing goes negative), so that parameter should not be optimized upward without re-checking; and the 3-ATR stop essentially never binds in-sample because the target-flip exit fires first — I kept it as a tail guard for regimes not in the sample, not as a source of the measured performance.

Failed deflated Sharpe: DSR=0.0002, expected-max=1.74 annualized over 225 trials — the selected config's realized Sharpe (~1.15 sensitivity-base, 0.73 full-sample) sits at or below the best-of-N noise bar, so after multiple-testing correction the edge is statistically zero (is_significant false, sharpe_ci_low 0.125). Two hard gates also fail: PBO 0.5648 > 0.5 (CSCV says the selection is more likely than not overfit) and holdout Sharpe 0.571 vs walk-forward OOS 1.347 = ratio 0.424 < 0.70 (degradation on the untouched final window). Per the robustness rubric, low DSR combined with high PBO is a decisive abandon. This is not a re-tuning problem: the sweep already found its best and it is below the noise Sharpe, so re-optimizing the same 3-parameter mechanism cannot manufacture significance the signal doesn't contain. The positives are real but insufficient — the walk-forward is not flagged overfit (IS/OOS 1.68, all three OOS windows positive) and sensitivity is clean with an 8.8% drawdown — but surviving one coarse 3-window split is not significance across 225 trials. The failure store confirms the class: the same regime-switch mechanism (BTC Hurst-Regime-Adaptive Directional, similarity 0.79) was abandoned OVERFIT, alongside the vol-regime-band and alt-breadth BTC siblings. Abandon.

Long-short, single-leg BTCUSDT.BINANCE USD-M on 4-HOUR bars. One quantity is traded — the volatility-normalized deviation of price from its own 120-bar (20-day) mean, returned as the continuous signal — and one classifier decides the SIGN of the reaction to it: the Kaufman efficiency ratio over the last 60 bars (10 days). When ER >= 0.12 the tape is trending, so the position is +sign(deviation): ride the move (momentum). When ER < 0.12 the tape is ranging, so the position is -sign(deviation): fade the stretch (mean reversion). A side is taken or held only while |deviation| >= 0.7, and the position flips whenever the target changes — either because price crossed its mean or because the regime itself rotated — with a hard 3-ATR stop from the entry price as a tail guard. The book is therefore deployed in BOTH regimes with the regime-appropriate sign rather than standing flat in half of them. Sizing is risk-first: quantity = equity x 2% / (3 x ATR), so a stop-out costs a fixed 2% of equity regardless of volatility and losses cannot compound into size, capped at 1.0x equity notional with leverage 1.0 (no margin used). Single instrument, single account, no hedge leg — structurally no cross-account liquidation path.

Regime-switch er_threshold is the highest-risk overfitting axis with a developer-flagged cliff above ~0.18 (range branch fires in trends → negative). Optimizer must bound it below the cliff.

Pure-OHLCV book, 39 trades, Sharpe CI [-1.0, 2.13] straddles zero, kurtosis 24.4, outlier-carried wins. Positive but thin; regime-adaptivity claim needs walk-forward + outlier-removal verification.

3-ATR stop rarely binds in-sample (target-flip exit fires first), so OOS tail control is less tested. Risk-officer note.

Large measurable sample: 312 trades (166L/146S), well above the 100-trade bar

Genuine decorrelated alpha (not bull beta): alpha +0.127, beta 0.06, correlation 0.20; positive in most years including 2026 +17.9% YTD

Strong and controlled: profit_factor 1.85, Sharpe 0.88 with CI-low 0.29 > 0 (significant), max_drawdown 15.4% (best in batch), avg_trade_return_pct 0.92%, PSR 0.998

Win/loss shape (win_rate 0.29, avg_win 4.5x avg_loss, positive skew) matches the hypothesized trend-follow/mean-revert mechanism

Outlier dependence: return_kurtosis 29, skew 2.67, a few very large single days (2023-01-30 +36%, 2021-01-11 +26%); 2023's +68% leans on its January

Family/overfit risk: the BTC vol-regime-classifier sibling just collapsed in the same 3-phase run (IS 7.6 -> OOS -2.1, holdout -0.99); a sign-switching er_threshold is overfit-prone

Recent regime choppy: 2025 -7.5%, rolling_sharpe hit -8.7 (Sep 2024) and -3.8 (early 2026), though 2026 recovered

information_ratio -0.41 vs buy-hold (set aside given true decorrelation and positive alpha)

Walk-forward NOT flagged overfit: IS/OOS ratio 1.68, all three OOS windows positive (0.55, 2.58, 0.91, avg 1.35) — genuinely better than the vol-regime sibling's collapse

Clean sensitivity (0 cliffs, smooth ~1.0-1.3 surface); optimized config has a tidy 8.8% max drawdown, PF 2.34, and 2026 +22%

Deflated Sharpe 0.0002 with expected-max-of-N 1.74: the selected Sharpe is at/below the best-of-225-trials noise bar — statistically zero edge after multiple-testing correction (is_significant false, CI-low 0.125)

PBO 0.5648 > 0.5 (hard fail): parameter selection more likely than not overfit; low DSR + high PBO is a decisive-abandon combination

Holdout failed (hard fail): 0.571 vs WF OOS 1.347, ratio 0.424 < 0.70 — real degradation on the untouched window

Soft floors missed: Sharpe 0.73 < 1.5, win_rate 0.31 < 0.45, deflated_sharpe 0.0002 < 0.95; 2025 still a losing year (-6.6%)

Family prior: the same regime-switch mechanism (BTC Hurst-Regime-Adaptive, sim 0.79) was abandoned OVERFIT, alongside the vol-regime and alt-breadth siblings

BtcRegimeAdaptiveDirectionalLS aimed to stay aligned with whichever regime was live — trading BTC momentum in trends and mean-reversion in ranges via one efficiency-ratio classifier that flipped the sign of its reaction — so its edge would not evaporate when the regime rotated. It was the best candidate in its batch: +273%, PF 1.85, Sharpe 0.88 with a significant CI-low, a 15.4% drawdown, and genuine decorrelated alpha, which earned it a full optimization run. That run gave a genuinely mixed picture — the walk-forward was not flagged overfit and all three OOS windows were positive — but the decisive gates failed: deflated Sharpe of 0.0002 at the best-of-225 noise bar, PBO 0.565, and a holdout that degraded to a 0.42 ratio. The analyst abandoned it at the ANALYZING stage as statistically zero edge after multiple-testing correction, noting the same regime-switch mechanism had already been abandoned overfit; it reached optimization and analysis but not risk review or promotion.

Surviving one coarse walk-forward split with positive OOS windows is not the same as significance across a 225-trial sweep — when the selected Sharpe lands at or below the expected-max-of-N noise bar (DSR ~0) with PBO above 0.5 and a failed holdout, a sign-switching regime classifier has been fit to noise, and the whole regime-switch family shows the same overfit signature.

It earned an 'optimize' verdict at backtest review and ran the full 3-phase optimization, but the analyst abandoned it at the post-optimization ANALYZING stage on three robustness failures: the selected Sharpe sits at/below the best-of-N noise bar (DSR 0.0002, not significant), PBO 0.565 > 0.5 says the selection is more likely than not overfit, and the untouched holdout degraded to a 0.42 ratio — a decisive low-DSR/high-PBO combination that re-tuning cannot fix.

A long-short, single-instrument, single-leg strategy on BTCUSDT.BINANCE USD-M (3 parameters) that ran one Kaufman-efficiency-ratio regime classifier and switched the sign of its reaction to the same volatility-normalized deviation signal — trading momentum (ride the move) in trending regimes and mean-reversion (fade the stretch) in ranging regimes, with a hard ATR stop and no hedge leg.

The initial backtest (BTCUSDT 4H, 2409 data days) was the strongest of its batch: +273% return over 312 balanced trades (166L/146S), profit factor 1.85, Sharpe 0.88 (CI-low 0.29 above zero), max drawdown 15.4%, decorrelated alpha +0.127 (beta 0.06). But optimization exposed statistical insignificance — deflated Sharpe 0.0002 against an expected-max-of-225-trials bar of 1.74, PBO 0.565, and a failed holdout (0.571 vs walk-forward OOS 1.347, ratio 0.424), though the walk-forward itself was not flagged overfit (all three OOS windows positive, IS/OOS 1.68).