BtcEthBollingerReversion4hTrendFilterLS1H

BTC and ETH 1H Bollinger Band Reversion with 4H Trend Filter Long-Short Portfolio

A LONG-SHORT, multi-timeframe, trend-aware mean-reversion strategy on BTCUSDT.BINANCE and ETHUSDT.BINANCE perpetual futures. Structurally distinct from every current portfolio strategy: the promoted CrossSectionalMeanReversion3Major is multi-instrument cross-sectional MR; the promoted ETH+SOL 1H+4H is long-only breakout; my pipelined MultiTimeframePullbackBTCBNBLong is long-only pullback-to-SMA; my pipelined BTCETHCointegrationPairsReversion is dollar-neutral pairs; my pipelined BTCVolatilityTargetedLong is long-only vol-overlay. This strategy is the FIRST in the portfolio to combine: (a) trend-following framework (4H SMA(50) filter), (b) statistically-grounded mean-reversion signal (1H Bollinger Bands at ±2σ), AND (c) LONG-SHORT direction. The mechanism: when the 4H trend is UP (close > 4H SMA(50) AND SMA slope rising), Bollinger -2σ touches on 1H represent oversold dips WITHIN an uptrend — buy the dip with mean-reversion thesis, exit on +1σ recovery (back to center). When the 4H trend is DOWN, Bollinger +2σ touches represent overbought rallies WITHIN a downtrend — short the rally. The trend filter prevents the strategy from fighting trend (which killed pure mean-reversion strategies on majors); the Bollinger threshold provides precise statistical entry timing (well-documented technical analysis tool, not curve-fit). This responds to the analyst's directive to pursue 'materially differentiated mechanism classes' beyond the cross-sectional MR class which has been declared closed. Uses ONLY BINANCE USD-M perp 1H + 4H bars (proven clean data; the recent 1H+4H success demonstrates this combination is testable). Fills FOUR critical gaps: (1) LONG-SHORT direction (14.8% → toward 45%), (2) trend-aware mean-reversion hybrid mechanism (zero in portfolio), (3) two-asset multi-timeframe (covers different instruments than the ETH+SOL success and BTC+BNB pullback), (4) OHLCV-only data dependency. Position sizing: 12% per instrument per direction; max 24% gross per instrument; max 48% gross portfolio total. Risk per trade: 1.0% of equity from per-trade 1.0% Bollinger-band-width stop.

Iteration-2 changes address the analyst's two blockers while preserving every layer that already passed (imports, multi-leg orchestration on the primary 1H clock, 1H-proxy trend fallback for Layer-2, continuous diagnostic signal). (1) CAPITAL DEPLOYMENT: bb_std loosened 2.0->1.7 so more dip/rally setups qualify, and the trend filter relaxed from strict close>SMA AND slope>0 to plain close-vs-4H-SMA (use_slope_filter defaults OFF) so UP/DOWN classification no longer collapses to FLAT most of the time — both raise signal frequency / time-in-market while keeping the trend-alignment that protected the edge. (2) PAYOFF SHAPE: profit target moved from the mid-band to mid ± exit_target_std(1.0)σ (a ~2.7σ favorable move from a -1.7σ long entry) so winners are no longer capped small, and the safety stop tightened 3.0%->1.5% so target distance >= stop distance — removing the structurally adverse avg_loss > avg_win asymmetry that previously only survived on a high win rate. All new magic numbers are parameters for the optimizer. Market-neutrality is structural (longs only in uptrends, shorts only in downtrends, symmetric sizing) and unchanged.

Overfit with no forward edge across every gate. walk_forward is_overfitted=TRUE: avg IS Sharpe 14.95 collapses to avg OOS 0.0 across ALL THREE windows — the optimizer inflated in-sample Sharpe only by over-filtering the signal (bb_std 1.98, bb_period 30, slope filter ON) into a near-non-trading config that fires just 16 trades and produces zero OOS activity. Multiple-testing gates concur: deflated_sharpe 0.0176 (<<0.95), is_significant=false, PBO 0.599 (>0.5), and the selected WF Sharpe (~15) sits right at the 225-trial expected-max luck bar (14.87) — indistinguishable from best-of-N noise. The holdout FAILED (0.0, ratio 0) and sensitivity FAILED with 3 cliffs on the CORE signal parameters (bb_std, bb_period, trend_sma_period), so the signal itself is fragile. The base was already marginal (Sharpe 0.76, sharpe_ci_low -0.336 straddling 0, profit_factor 1.163, CAGR 2.16%, IR -0.69), which is why optimization could not rescue it: trend-filtered Bollinger mean-reversion on BTC/ETH majors has essentially no exploitable edge net of fees (commission 11.1% of gross). Not iterate (attempt 1 of 2): OOS is uniformly zero, all three core signal params are cliffs, and there is no robust region to tune toward — a second best-of-225 sweep re-finds the same over-filtered, non-generalizing config. Not revise_hypothesis: this mechanism class has no promoted sibling stranded on a dead target; the base shows the mechanism lacks edge, not that the target was wrong. FAILURE PATTERN: trend-gated Bollinger-band mean-reversion on majors produces a marginal base (PF 1.16, CI through 0) whose only high-IS-Sharpe optimization region is an over-filtered, near-non-trading config that yields ZERO out-of-sample Sharpe across all windows — a PSR=0.97/DSR=0.018 split is the diagnostic tell that the in-sample number is pure best-of-225 selection on noise.

Long-short, trend-aware Bollinger-band mean-reversion on BTCUSDT+ETHUSDT 1H perps with a 4H SMA(50) trend filter. In a 4H uptrend (close > 4H SMA) a 1H lower-band (mid - 1.7σ) touch is an oversold dip -> BUY; in a 4H downtrend an upper-band touch is an overbought rally -> SHORT. Exits at a target band mid ± 1.0σ (wider than the mean so winners aren't capped), a tightened 1.5% per-trade stop, or a 30-bar time cap. 12% equity notional per instrument per direction, one position per instrument, leverage 1.0.

Healthy, statistically meaningful sample: 862 trades over ~6.5 years with clean entry diagnostics (441 signaled -> 441 submitted, 0 dropped) — optimization will fit signal, not noise

Mechanism matches the hypothesis: genuinely long-short and balanced (479 long / 383 short), long win-rate 63% / short win-rate 59% — a real long_short + trend-aware-reversion candidate the book lacks, not a closet long

Non-degenerate risk profile: Sharpe 0.76, Sortino 1.09, max_drawdown 7.06%, win_rate 61%, profit_factor 1.16, positive in 4 of 7 years, low exposure (18%)

Thin edge: profit_factor 1.16 and total_return only 15.3% over 6.5y — avg per-trade edge is close to the fee floor

Heavy cost drag: commission_pct_of_gross 11.1% plus impact_cost_pct 19.1% (~30% of gross) — fee-fragile; optimization must lift avg_trade_return by reducing trade frequency / widening bands

Small capacity (capacity_usd $2.7M) and three negative years (2022 -1.4%, 2023 -2.7%, 2026 -2.1%) — deployable size and robustness are limited

Genuinely long-short (479 long / 383 short) with low drawdown (7.06%) and low kurtosis (4.04) — not outlier-carried

Base is honest (end_unrealized 0.5%, no MTM mirage) and mechanism matches hypothesis (buy lower-band dips in 4H uptrends, short upper-band rallies in downtrends)

862 trades give ample sample; benchmark_meaningful correctly false for a market-neutral book

deflated_sharpe 0.0176 (<<0.95), is_significant=false — after 225-trial correction, indistinguishable from best-of-N noise; selected WF Sharpe ~15 sits right at the 14.87 luck bar

walk_forward is_overfitted=TRUE: avg IS 14.95 collapses to avg OOS 0.0 across ALL three windows — the optimizer over-filtered into a near-non-trading config (optimized backtest = 16 trades)

Holdout FAILED (holdout_sharpe 0.0, ratio 0); sensitivity FAILED with 3 cliffs on the CORE signal params (bb_std, bb_period, trend_sma_period)

PBO 0.599 (>0.5) — selection more likely than not overfit

Base itself is marginal: Sharpe 0.76 with sharpe_ci_low -0.336 (CI through 0), profit_factor 1.163, CAGR 2.16%, IR -0.69 — no real edge net of fees (commission 11.1% of gross)

PSR 0.97 is the PSR-vs-DSR trap: it ignores trial count while DSR (0.018) rejects; optimized Sharpe 7.02 is an artifact of a 16-trade sample

The strategy is genuinely market-neutral (beta 0.0008) with a positive Sharpe (1.13, 70% win rate, 134 trades) but is non-contributory as configured and has a fragile payoff shape — fix both before optimization: 1) CAPITAL DEPLOYMENT: exposure is only 2.4% (CAGR 0.46%, total +3% over 6yr) because the +-2sigma entry + strict 4H trend-slope-sign filter fires rarely and reversion holds are short (~9.5h). Raise time-in-market to a useful level: loosen the entry (test bb_std ~1.5-2.0) and relax the trend filter (allow trend = close vs 4H SMA WITHOUT requiring a strict positive slope, or a gentler slope threshold) so more valid dip/rally setups qualify; the per-instrument position_fraction (12%) is fine, the signal FREQUENCY is the bottleneck. Target meaningfully higher exposure while keeping it market-neutral. 2) PAYOFF SHAPE: avg_loss ($247) is ~1.8x avg_win ($135) with tail_ratio 0.56 — the exit-at-mid-band caps winners small while the 3% stop allows bigger losses (the fragile MR asymmetry that only survives on the current 70% win rate). Widen the profit target beyond the mid-band (e.g., exit at the opposite band or a trailing target) and/or tighten the stop relative to the target so the win/loss ratio is not structurally adverse. 3) The edge is thin vs costs (commission 12.7% of gross) and the Sharpe CI deeply straddles 0 (sharpe_ci_low -1.21) on tiny 3.35% vol — so the re-config must demonstrate a more robust, better-deploying edge (higher exposure, less adverse payoff) before it is worth a 225-trial optimization. If the looser/wider-target re-run cannot deploy meaningful capital with a stable positive expectancy, abandon — a sub-1%-CAGR sleeve is not worth promoting even at a flattering Sharpe.

This strategy set out to fill portfolio gaps with a novel mechanism: a long-short, trend-aware Bollinger-band mean-reversion on BTC and ETH, buying 1H dips in 4H uptrends and shorting rallies in 4H downtrends. Its base backtest was honest and structurally on-thesis — 862 trades, balanced long/short, low 7% drawdown — but the edge was razor-thin (profit factor 1.16, CAGR 2.16%) with heavy fee drag, so the reviewer sent it to optimization rather than promotion. Optimization then exposed the fragility: the only high-Sharpe in-sample region was an over-filtered 16-trade config whose out-of-sample Sharpe was zero across every walk-forward window, and it failed the holdout, sensitivity (3 cliffs on core params), PBO, and deflated-Sharpe gates. On its second iteration the analyst abandoned it, concluding trend-gated Bollinger reversion on majors has no exploitable edge net of fees and no robust region left to tune toward.

Trend-gated Bollinger-band mean-reversion on BTC/ETH majors yields only a marginal base edge net of fees, and optimizing it merely over-filters into a near-non-trading config that fails to generalize — so a marginal base (PF ~1.16, Sharpe CI straddling 0) is not worth a best-of-N optimization sweep.

It cleared the backtest-review gate ('optimize') but the post-optimization analyst issued 'abandon': walk_forward was flagged overfitted (avg IS Sharpe 14.95 → avg OOS 0.0 on all three windows), the holdout failed (Sharpe 0.0, ratio 0), deflated Sharpe was 0.0176, PBO 0.599, and sensitivity failed with 3 cliffs on the core signal params (bb_std, bb_period, trend_sma_period).

A long-short, trend-aware mean-reversion strategy on BTCUSDT and ETHUSDT USD-M perpetuals that bought 1H lower-Bollinger-band dips when the 4H SMA(50) trend was up and shorted upper-band rallies when the 4H trend was down.

The base backtest ran 862 trades over ~6.5 years with a Sharpe of 0.76, 61% win rate, 7.06% max drawdown, and a genuinely balanced long-short split (479 long / 383 short), but the edge was thin: total return 15.3%, CAGR 2.16%, profit factor 1.16, and commissions were 11.1% of gross. Optimization's 16-trade selected config showed an inflated in-sample Sharpe (~15) that collapsed to 0.0 out-of-sample across all three walk-forward windows.