BtcFourHourPostVolSpikeRecoveryLong

BTC Post-Volatility-Spike Recovery Long-Only (Single-Asset BINANCE 4H, Parkinson-Volatility-Regime-Switching, Bars-Only)

Single-asset long-only strategy implementing the analyst's EXPLICITLY recommended future hypothesis class: 'BTC realized volatility regime strategies (volatility data is in bars, no supp_data needed).' This proposal addresses every documented failure mode from recent rejections: (a) no supp_data dependency (L/S ratio, HL funding, COIN-M, liquidations all confirmed unavailable or unreliable in factory infrastructure), (b) single-asset single-venue (no cross-venue execution realism gap, no multi-instrument backtest accounting bug), (c) long-only (no short-leg crypto alt-season blowup), (d) novel mechanism class never tested in pipeline. Mechanism: the Parkinson volatility estimator (high-low range, more efficient than close-to-close stdev — Parkinson 1980, Garman-Klass 1980) computed over a rolling 20-bar window detects the local volatility regime entirely from bar data. Volatility clusters (Engle 1982 ARCH, Bollerslev 1986 GARCH) — periods of high vol are followed by high vol, low vol by low vol — but TRANSITIONS between regimes are predictable: after a capitulation spike (vol > 75th percentile of 100-bar history), the subsequent reversion to normal vol is associated with the EARLY RECOVERY phase where price drifts up before the next vol cycle begins. Strategy: detect the moment when vol drops from the 'volatile' regime back into 'normal', confirm with a higher-high price pattern, go long, hold while normal-vol regime persists, exit when vol spikes again. This is a classic 'buy the calm after the storm' setup documented in volatility-trading literature (Sinclair 2010 'Volatility Trading'). NOT a chart pattern (uses statistical vol estimator), NOT a cross-venue strategy, NOT cross-sectional, NOT funding-driven, NOT positioning-extreme, NOT OI dynamics, NOT taker flow, NOT premium index, NOT volume breakout. Single-instrument BTC, single-venue Binance USD-M, 4H bars (safe data with multi-year history), 4 parameters (vol_lookback, vol_pct_high, vol_pct_normal, stop_pct) — minimal overfitting surface. Compliant with EVERY documented factory infrastructure constraint.

Addresses the iteration-2 feedback directly: the prior version had negative edge (PF 0.91) because it bought vol-normalization recoveries with no broader-trend filter, catching bull traps in the 2022/2023/2025 downtrends. The single, minimal change is a 200-bar 4H SMA trend gate in should_enter (close > SMA), implementing the 200-day falling-knife guard referenced in the hypothesis. All previously-passing logic (Parkinson vol estimator, regime state machine, continuous percentile-rank signal, sizing, exits) is unchanged so earlier verification layers stay green. min_bars_required and the buffer cap were widened to accommodate the 200-bar SMA. The vol-timing component is preserved for the backtest to measure whether it adds edge over a naive trend-filtered dip-buy; trend_sma is exposed as a parameter for that comparison.

Failed deflated Sharpe, sensitivity, and a NEGATIVE holdout on attempt 1. DSR=0.0057 (vs 0.95 bar) with the optimized Sharpe 0.2386 FAR below the 225-trial expected-max luck bar of 2.6567 (is_significant=false, sharpe_ci_low -1.699 deeply through 0, PBO 0.7302) — statistically indistinguishable from best-of-N noise. The time-ordered HOLDOUT FAILED and is NEGATIVE (holdout_sharpe -1.764, ratio -0.898): the untouched recent window loses money, directly contradicting the flattering WF-OOS of 1.964 (the strong-WF / negative-holdout split is a regime/holdout-luck artifact, and the base already decays — 2025 -7.4%). Sensitivity FAILED with 6 cliffs on the core signal-defining parameters (pct_window, hh_lookback, vol_lookback, vol_pct_high, max_hold_bars, vol_pct_normal) — the edge exists only at knife-edge values. Decisively, the optimizer DEGENERATED the mechanism by selecting vol_pct_high=1.018 (>1.0, an unreachable percentile), permanently disabling the 'volatile regime' detection that defines the hypothesis (no spike-cancel, no regime-ended exit) — so the 'optimized' config is not the strategy described, and it is economically dead anyway (CAGR 0.53%, PF 1.058). Not iterate (attempt 1 of 2): with 6 cliffs on the core knobs, a negative holdout, a degenerate best-params region, and a selected Sharpe an order of magnitude below the luck bar, there is no robust region to tune toward — a second best-of-225 sweep re-finds the same fragile/degenerate fit. Not revise_hypothesis: the Parkinson-vol regime-recovery mechanism has no promoted sibling and BTC is not a dead target, so this is a weak, novel single-asset edge that fails deflation, not a proven mechanism stranded on a dead instrument. FAILURE PATTERN: a clean, infrastructure-compliant, academically-grounded single-asset vol-regime 'buy the calm after the storm' produces a marginal base (Sharpe 0.50, CI through 0) whose best-of-225 optimization cannot clear deflation (DSR 0.006, Sharpe 0.24 vs 2.66 luck bar), whose defining parameters are all cliffs, and whose untouched holdout is NEGATIVE (-1.76). A non-overfit walk-forward with strong OOS is NOT evidence of an edge when the holdout is negative and the optimizer only 'wins' by driving a regime threshold above 1.0 to switch the hypothesized mechanism off — the flattering WF number is period-selection luck, not skill.

BTC 4H long-only Parkinson-volatility regime-switching 'buy the calm after the storm' strategy. Detects a volatile regime (Parkinson high-low vol percentile >= 75th of last 100 bars), then arms a recovery when vol reverts to normal (<= 50th pctile), entering long on a higher-high confirmation. Iteration 2 adds a 200-bar 4H SMA broader-uptrend filter so recoveries are only bought in established uptrends (falling-knife guard). Exits on vol re-spike, stop, take-profit, or time stop.

Trades faithfully implement the hypothesis: 105 long-only post-vol-spike recovery entries via Parkinson-vol regime transition + higher-high confirm + 200-bar uptrend gate — no mechanism mismatch

Healthy trade sample (105 over 6.5yr, ~35/walk-forward window) — well above the thin-sample threshold, so optimization can be statistically meaningful

Fee-viable and low-risk: commissions 4.02% of gross, max drawdown 12.9%, annualized vol 11.3%, Sortino 0.74 > Sharpe 0.50 (downside contained)

Novel, differentiated mechanism (vol-regime-switching) with a bounded parameter surface

Thin raw edge: profit_factor 1.17, Sharpe 0.50, sharpe_ci_low -0.81 straddles zero

Strongly regime-concentrated: return carried by 2024 (+16.7%), negative in 2022/2023 and most recent full year 2025 (-7.4%) — hints at decay or bull-regime dependence

Rolling Sharpe swings wildly (+7.6 to -5.1); information_ratio -0.69 (no alpha over buy-hold)

More than the advertised '4 parameters' effectively in play (vol params + trend_sma + hh_lookback + stop/tp/hold) — some overfitting surface

Infrastructure-clean: single-asset, single-venue, long-only, pure bar data (Parkinson vol estimator), no supplementary-data dependency — avoids every documented data/execution failure mode.

Base backtest is well-behaved on the surface: low max drawdown (12.9%), modest kurtosis (9.6), balanced win rate (53%), no outlier-carry.

Walk-forward reported non-overfit with positive OOS windows [1.81, 2.22, 1.86].

Fails the multiple-testing hard gate decisively: deflated_sharpe=0.0057 with the optimized Sharpe 0.239 FAR below the 225-trial expected-max luck bar of 2.657 (is_significant=false, sharpe_ci_low=-1.699 deeply through 0, PBO=0.7302).

HOLDOUT FAILED and is NEGATIVE: holdout_sharpe -1.764 (ratio -0.898) — the untouched recent window loses money badly, directly contradicting the flattering WF-OOS of 1.964. This WF-strong / holdout-negative split is a regime/holdout-luck artifact, not a real edge.

Sensitivity FAILED with 6 cliffs on the core signal-defining parameters (pct_window, hh_lookback, vol_lookback, vol_pct_high, max_hold_bars, vol_pct_normal) — the edge lives only at fragile knife-edge values.

Optimizer DEGENERATED the mechanism: it selected vol_pct_high=1.018 (>1.0), a percentile threshold that can NEVER be reached — this permanently disables the 'volatile regime' detection that defines the hypothesis (no spike-cancel, no regime-ended exit). The 'optimized' config is not the strategy the hypothesis describes.

The optimized result is economically dead: total_return 6.2% over the span, CAGR 0.53%, profit_factor 1.058, Sharpe 0.239 — barely distinguishable from no edge after fees (impact_cost 14.6%).

Base is already marginal and regime-fragile: Sharpe 0.498 (CI low -0.815), PF 1.17, negative 2022/2023/2025 with recent decay (2025 -7.4%).

Do NOT optimize yet — the edge is negative (PF 0.91) due to a specific structural omission, not a parameter problem. The strategy buys vol-normalization recoveries with NO broader-trend filter, so it catches bull traps in downtrends (2022 -20.7%, 2023 -6.2%, 2025 -11.4%) while only making money in bull years. The hypothesis itself referenced the 200-day falling-knife guard but the code doesn't implement it. Fix: 1. ADD A TREND FILTER (primary): gate entries on the broader uptrend — require close > a long SMA (e.g. 200-bar 4H, ~33-day) at entry, so the 'buy the calm after the storm' recovery only fires in established uptrends. This directly targets the bear-year losses (~-38% combined). Re-run and report annual returns and PF. 2. CONFIRM THE VOL-TIMING ADDS EDGE (critical): after adding the trend filter, compare against a naive 'long after any dip in uptrend' baseline — if the vol-regime-recovery timing does not improve PF/Sharpe over the simple trend-filtered dip-buy, the Parkinson-vol machinery adds no alpha and the strategy is redundant with the existing dip-buy hypothesis (abandon then). 3. TARGET: trend-filtered version should reach PF > 1.2 with the edge present in OOS years, not just the bull markets. If even with the filter PF stays < 1.2, abandon — the vol-recovery signal is just long exposure with no timing edge. Resubmit once the trend filter is added and the vol-timing component is shown to contribute genuine edge.

BtcFourHourPostVolSpikeRecoveryLong was a deliberately infrastructure-clean, academically-grounded strategy that used Parkinson volatility regime-switching to buy BTC's early recovery after a volatility spike settled, on pure 4H bar data. Its base backtest looked tidy — Sharpe 0.50, 105 trades, 12.9% max drawdown, no outlier-carry — so it advanced to optimization, but the raw edge was thin and regime-concentrated (negative in 2022, 2023, and 2025). Optimization failed decisively: deflated Sharpe 0.0057 against a 2.657 luck bar, 6 sensitivity cliffs on the defining parameters, and a negative holdout (-1.764) that flatly contradicted the positive walk-forward OOS, while the optimizer degenerated the mechanism by choosing an unreachable >1.0 volatility percentile that turned the regime detection off. After two iterations it was abandoned as a weak, novel single-asset edge whose flattering forward number was period-selection luck rather than skill.

A non-overfit walk-forward with strong OOS is not evidence of an edge when the untouched holdout is negative; a clean, infrastructure-compliant, academically-grounded mechanism can still be regime-luck, and an optimizer that only 'wins' by pushing a threshold past a value it can never reach has switched the hypothesized strategy off rather than improved it.

The analyst abandoned it because it failed the multiple-testing gate (deflated Sharpe 0.0057 with optimized Sharpe 0.239 far below the 225-trial luck bar of 2.657, PBO 0.7302), sensitivity failed with 6 cliffs on the core signal parameters, the time-ordered holdout was negative (Sharpe -1.764, ratio -0.898) contradicting the flattering walk-forward OOS of 1.964, and the optimizer degenerated the mechanism by setting vol_pct_high to 1.018 (>1.0, an unreachable percentile that disables the defining volatile-regime detection).

A single-asset, single-venue, long-only 'buy the calm after the storm' strategy on BTCUSDT 4H futures that used the Parkinson high-low volatility estimator to detect vol reverting from a spike back to a normal regime, entering long on the early-recovery drift (confirmed by a higher-high and a 200-bar SMA uptrend gate) using pure bar data with no supplementary feeds.

The base backtest was well-behaved but thin: Sharpe 0.50, total return 17.26, profit factor 1.17, 53% win rate over 105 trades, with a low 12.9% max drawdown and no outlier-carry. Optimization actually worsened it to Sharpe 0.239, 6.2% total return, CAGR 0.53% and profit factor 1.058 over 51 trades — economically near-dead after fees.