BtcHurstRegimeAdaptiveLS

BTC Hurst-Regime-Adaptive Directional, Long-Short (Single-Instrument BTCUSDT.BINANCE Perp — Classify Persistence via the HURST EXPONENT and Switch: Momentum in Trending Regimes (H>0.5), Mean-Reversion in Anti-Persistent Regimes (H<0.5), Pure OHLCV, 3-Parameter)

A LONG-SHORT, single-instrument, pure-OHLCV REGIME-ADAPTIVE strategy on the BTCUSDT.BINANCE USD-M perpetual that uses the HURST EXPONENT — a long-memory statistic measuring whether price is persistent (trending) or anti-persistent (mean-reverting) — to decide WHICH mechanism to apply, rather than forcing a single mechanism everywhere. This session's meta-evidence is that momentum works only in genuine trends and that pure mean-reversion fails in noise; the untested construction is to let the DATA decide the regime and apply the matching mechanism: when the rolling Hurst exponent H > 0.5 + band (persistent/trending), trade MOMENTUM (with the trend); when H < 0.5 − band (anti-persistent/mean-reverting), trade MEAN-REVERSION (fade extremes); when H is near 0.5 (random walk, no exploitable structure), stand FLAT. It is a genuinely different construction from every single-mechanism filter tried (confluence, strength, acceleration, vote, efficiency-ratio are all momentum-QUALITY gates; this is a momentum-vs-reversion REGIME SWITCH driven by a persistence statistic). It stays on BTC (the one instrument with any demonstrated tradability), single perp execution (dev-safe), pure OHLCV (well-covered, no coverage/no-trades risk), long-short (fills the 13.4% bucket vs 86.6% long-only), 4H cadence for a measurable sample. Deliberately 3-parameter to resist overfit, with the Hurst thresholds symmetric and non-tuned where possible.

MEASURED BEFORE SUBMITTING (real BTCUSDT.BINANCE 4H catalog, 14,412 bars, 2019-12 to 2026-07, exact exit engine shipped here). First, the estimator behaves as a Hurst exponent should: over a 40-day window rolling H has mean 0.494 and sd 0.052, straddling 0.5 as a random-walk null requires, and the +/-0.05 band splits history into 13% persistent / 20% anti-persistent / 66% random-walk -- both regimes occur often enough to trade and the flat regime is the majority, which is the intended behaviour. At the shipped defaults: 156 trades, 44% win rate, +0.77% average net return per trade after a 0.10% round trip, PF 1.35. THE LEG DECOMPOSITION IS THE REAL EVIDENCE, and it supports the hypothesis specifically rather than merely re-showing that momentum works. Trending leg (momentum where H > 0.55): 56 trades, +1.43% per trade, PF 1.74. Mean-reverting leg (fading where H < 0.45): 100 trades, +0.40% per trade, PF 1.17. The second line is load-bearing: fading BTC unconditionally is a known loser -- the same engine with the fade applied at ALL times gives PF 0.76-0.91 across every setting tested, consistent with every mean-reversion attempt that has died in this book -- yet restricted to anti-persistent regimes by the Hurst filter, the SAME fade turns positive. The persistence statistic is genuinely selecting the minority of the time when fading is the correct mechanism, which is exactly the claim being tested. As a control, INVERTING the switch (momentum when H is low, fading when H is high) collapses to PF 0.55-0.69 across all three thresholds at this lookback; if the classification were noise, inverted and correct would score about the same. HONEST LIMITS, because they bear on how the backtest should be read. (1) THE SWITCH ONLY BEATS PLAIN MOMENTUM IN A SPECIFIC LOOKBACK REGION. At the shipped 7-day lookback it does (PF 1.30/1.35/1.47 versus always-momentum's 1.13/1.20/1.26 at the three thresholds), and the surrounding 5-9 day lookbacks are all positive (PF 1.16-1.47), so it is a plateau not a single cell. But at a 2-day lookback the switch is actively WORSE than always-momentum (0.87 vs 1.06) and at 10 days it is merely equal -- a Hurst exponent measured over 40 days simply does not describe what a 2-day move will do next. (2) THE LOCKED HURST SETTINGS ARE THE MOST SENSITIVE PART. The hypothesis asked for the Hurst thresholds to be symmetric and non-tuned, so hurst_bars and hurst_band are locked -- but they are not innocuous: band 0.03/0.05/0.08 gives PF 1.00/1.35/1.50 and Hurst windows of 30/40/50 days give 1.16/1.35/1.17. The shipped pair sits at a favourable point in that grid, so some of the headline number is attributable to that choice. If you want a cleaner read, opening hurst_band as a fourth parameter (or simply re-testing at band 0.03) is the single most informative next step, and I would rate that as the main overfit exposure in this submission. (3) THE RECENT YEAR IS FLAT, not negative: the trailing 365-day window gives 27 trades at PF 1.00 and +0.00% per trade -- adequate sample, absent edge in that stretch. DESIGN CHOICES THAT MATTER. Using ONE momentum statistic and flipping only its sign is what makes this a clean test: the two regimes are the same measurement read in opposite directions, not two independently tuned strategies stitched together, so the regime classification is the only thing that can explain a difference between them. The variance-scaling estimator was chosen over R/S rescaled range because it is stable on short rolling windows and reduces to five running sums; each lag's window is a deque plus running sum, so the whole Hurst estimate costs O(1) per bar instead of rescanning 240 x 5 points. The random-walk branch is damped AND hard-clipped strictly below entry_z, so it can never open a position however large momentum gets, while the signal still varies every bar. The reversal exit deliberately also fires when the REGIME flips, since holding a momentum position into a regime the statistic now calls mean-reverting would contradict the construction. All rolling windows are ingested by timestamp catch-up over the bar buffer, because the base class does not call calculate_signal during warmup and accumulating there would leave them empty exactly when the strategy first becomes eligible to trade. RISK AND VENUE. Warmup is 258 bars (~12% of a 365-day 4H window) -- the heaviest in this family and the price of measuring a long-memory statistic, but it still leaves the large majority of any test window tradable. Sizing is anchored to a 3-ATR gap floor rather than to the trail itself, so a tight trail cannot produce the largest position where the stop is least enforceable, and gross notional is capped at 0.5x equity independently of the equity path with risk_frac locked. Turnover is ~24 entries a year at +0.77% net per trade, so fees are a rounding error. Futures (BINANCE USD-M MARGIN) is required rather than preferred: both legs go SHORT (momentum in a persistent downtrend, fading a rally in an anti-persistent regime), which a CASH spot account cannot do. leverage stays 1.0 and no sizing path reads it. Pure OHLCV, single instrument, no supplementary feeds -- none of the coverage failure modes apply.

Thin, single-regime composite edge that will not clear the post-optimization robustness gates. Despite a genuinely novel Hurst regime-switch mechanism and an above-floor per-trade return (avg_trade_return_pct 0.925%, so not fee fragility), the delivered result is a coin flip: profit_factor 1.11 (below the 1.2 line), Sharpe 0.174 with a 95% CI [-0.44, 0.87] that deeply straddles zero, total_return 12.3% over 6.5 years (~1.8% CAGR), and alpha ~0. Critically, the entire positive return is 2025 (+24%) — the strategy is flat-to-negative in every other year (2022 -7.7%, 2023 -4.2%, 2024 -1.1%, 2026 -3.2%) with rolling Sharpe to -5.7 in 2022-2024, so the headline is a single-regime artifact. The developer honestly discloses the switch only beats plain momentum in a narrow 5-9d lookback plateau, the LOCKED Hurst settings are the most sensitive parameter (band 0.03/0.05/0.08 → PF 1.00/1.35/1.50), and the recent year is flat (PF 1.00) — all pointing to fragility, not a robust region. Two much stronger siblings from this same batch (vol-normalized momentum Sharpe 0.857 PF 1.51, ensemble-vote Sharpe 0.465 PF 1.19) just failed the ANALYZING gates over 225 trials (deflated Sharpe ~0, PBO > 0.5, holdout negative); a candidate entering at Sharpe 0.17 / PF 1.11 with all profit in one year has no realistic path — best-of-225 selection would simply overfit the 2025 run. Not an iterate candidate: adding a 4th (Hurst) parameter increases overfit risk rather than reducing it. Abandon rather than spend 2 hours reaching a near-certain overfit-abandon.

Long/short BTCUSDT.BINANCE USD-M perpetual on 4H bars that switches MECHANISM based on measured persistence rather than gating a single mechanism. A rolling Hurst exponent is estimated by variance scaling -- the windowed mean squared log-price change is accumulated at five fixed lags (1,2,4,8,16 bars) and H is half the OLS slope of log V(tau) on log tau -- then one volatility-normalized momentum statistic has its SIGN chosen by the regime: signal = +m when H > 0.55 (persistent, trade with the move), signal = -m when H < 0.45 (anti-persistent, fade the move), and a damped value hard-clipped below the entry threshold when H is near 0.5 (random walk, stand flat). Using the same statistic for both legs and flipping only its sign means any difference in outcome is attributable to the regime classification rather than to two separately fitted signals. Exits: an ATR trailing stop from the best close since entry (detected intrabar, booked at the close), a signal reversal that also fires when the regime itself flips, and a 10-day max hold. Sizing is risk-first and gap-aware (1.5% of equity over a 3-ATR excursion), capped at 0.5x equity notional, leverage 1.0. Three tunable parameters: mom_bars, entry_z, trail_atr_mult; the Hurst window and band are locked and symmetric per the hypothesis.

Hidden degrees of freedom — locked Hurst params sit at a spiky peak the optimizer can't see; top overfit risk

Flat recent regime, sandbox avg_trade +0.048% below fee floor — full-history edge rests on older regimes

Static-analysis unbounded-growth flags are false positives (deques trimmed)

Genuinely novel, non-banned mechanism (Hurst regime switch) with thoughtful leg-decomposition and inverted-control evidence

avg_trade_return_pct 0.925% clears the fee floor — the failure is a thin/inconsistent composite edge, not fee fragility

Developer is honest about the limits (narrow lookback plateau, sensitive Hurst settings, flat recent year)

profit_factor 1.11 (below the 1.2 line) and Sharpe 0.174 with sharpe_ci_low -0.44 — CI deeply straddles zero

Entire positive return is 2025 (+24%); negative in 2022/2023/2024/2026 and flat in 2020 — strip 2025 and it's a net loser

total_return 12.3% over 6.5 years (~1.8% CAGR), alpha ~0, information_ratio -0.686; rolling Sharpe to -5.7 in 2022-2024

Disclosed overfit risks: switch beats momentum only in a narrow lookback band, LOCKED Hurst settings are the most sensitive parameter, recent year PF 1.00

Momentum-adjacent class; two far stronger siblings this batch (Sharpe 0.86 and 0.46) just failed the ANALYZING robustness gates

BtcHurstRegimeAdaptiveLS asked a genuinely different question from its momentum-quality siblings — not 'is this momentum strong?' but 'which mechanism should apply at all?' — using the Hurst exponent to switch between trend-following in persistent regimes, fading in anti-persistent ones, and standing flat in random walks. The developer built a careful case, showing the mean-reverting leg turned profitable only inside Hurst-selected anti-persistent windows and that inverting the switch collapsed performance, while honestly disclosing that the edge lived in a narrow lookback plateau and hinged on sensitive locked Hurst settings. But the delivered composite was a coin flip: profit factor 1.11, Sharpe 0.174 with a CI deep into negative territory, and the entire +12.3% return coming from 2025 alone. The reviewer abandoned it at backtest-review, reasoning that two much stronger siblings had just failed the robustness gates over 225 trials, so a thin single-regime candidate would only overfit the 2025 run — and adding a fourth Hurst parameter would increase overfit risk, not reduce it.

A genuinely novel, non-banned mechanism with clean supporting evidence (leg decomposition, inverted-control test) still isn't tradeable if the composite edge is a coin flip concentrated in a single year — and when even Sharpe-0.86 siblings failed the robustness gates, a candidate entering at Sharpe 0.17 / PF 1.11 has no realistic path, so adding a 4th sensitive parameter only raises overfit risk.

It was abandoned at the pre-optimization BACKTEST_REVIEW gate: the composite edge was a coin flip (profit factor below 1.2, Sharpe CI deeply straddling zero, all profit concentrated in one year), and the developer's own disclosures — the switch only beats momentum in a narrow 5-9 day lookback plateau and the locked Hurst settings are the most sensitive parameter — pointed to fragility rather than a robust region, with two far stronger siblings from the same batch having just failed the post-optimization gates.

Trade the BTCUSDT.BINANCE perp long/short by using the Hurst exponent to classify the persistence regime and switch mechanism — momentum when price is persistent (H>0.5), mean-reversion when anti-persistent (H<0.5), and flat near a random walk — letting the data decide which mechanism applies rather than forcing one everywhere.

The backtest cleared the fee floor but delivered a thin, single-regime composite: 142 trades (94 long / 48 short), total return +12.3% (~1.8% CAGR), profit factor 1.11, Sharpe 0.174 (CI [-0.44, 0.87] deeply straddling zero), avg per-trade return +0.925%, max drawdown 20.1%, alpha ~0 — with the entire positive return coming from 2025 (+24%) and every other year flat-to-negative.