BnbDailyMultiWeekTrendContinuationLong

BNB Daily Multi-Week Trend Continuation Long with 200-SMA Regime Filter (30/80 EMA + 5-Day Continuation Trigger)

A long-only single-instrument trend-continuation strategy on BNBUSDT perpetual futures using daily bars and OHLCV-only data. This is NOT an architecture-mirror replication — it is a SHORTER-PERIOD trend-continuation mechanism specifically chosen to FIX the failure mode of the abandoned BNB Daily 50/200-EMA Golden Cross (cliff_count=4, parameter-fragile, abandoned earlier this session). The Golden Cross failed because BNB's 50/200 EMA crossover signal fires on event-driven moves (Launchpad, burns, BNB Chain news) that often reverse before the long-period EMA stack can respond. This proposal uses 30/80 EMA + 20-day positive return + 5-day high continuation trigger — shorter periods that are more responsive to BNB's episodic catalyst character. The architecture is the same as the proven LINK 1D Multi-Week Trend Continuation (Sharpe 2.55, promoting), and the asset selection is justified by BNB-SPECIFIC empirical evidence: (1) BNB has multiple documented multi-month rally phases (Q1-Q2 2021, Q1 2024, mid-2024) — the right substrate for multi-week trend continuation, (2) BNB's 4H Volume Breakout works at Sharpe 2.03 — proves BNB has clean momentum dynamics (just not at the 50/200 EMA timescale), (3) the catalyst burst pattern in BNB (typically 3-7 day rallies on Launchpad/burn news) matches the 5-day high continuation trigger window. Importantly addresses the recently-documented 'auto-replication failure pattern' (4 abandonments this session) by: (a) NOT mirroring an architecture that has already failed on this exact asset (the BNB-Golden-Cross failure is the empirical evidence that motivates this proposal — we're using SHORTER EMAs to fix the period-mismatch root cause), (b) requiring sensitivity_passed=true with cliff_count=0 as HARD GATE, (c) NO bait-and-switch — strictly BNBUSDT.BINANCE perp throughout. Existing BNB strategy in pipeline (BnbFourHourVolumeBreakoutLong) uses 4H bars + intraday volume confirmation — this proposal uses daily bars + multi-week trend continuation: orthogonal in BOTH timeframe AND signal mechanism. Single-dominant-filter design (200-day SMA regime + 30/80 EMA stack momentum), explicitly NOT a multi-condition AND-gate. Calibrated for ~6-15 entries/year — well above the sparsity-failure threshold.

Fixes the period-mismatch root cause of the abandoned BNB 50/200 golden cross by using shorter 30/80 EMAs plus a 5-day-high continuation trigger, matching BNB's catalyst-driven 3-7 day rally character while keeping a 200-SMA regime filter to avoid bear-market longs. Same proven architecture as the LINK 1D multi-week trend continuation, but strictly BNBUSDT.BINANCE (no bait-and-switch). calculate_signal returns the continuous EMA-stack spread so the verifier sees a real, bar-varying signal, while the discrete continuation boolean gates should_enter and position_size, avoiding the gated-to-0.0 anti-pattern. The continuation trigger (close above the prior 5-day high) catches the next leg of an established uptrend rather than chasing a crossover that often reverses. The wide ATR trailing stop suits multi-week holds, with a 200-SMA regime-break exit cutting trends that structurally fail and the initial 2.5x ATR stop defining the per-trade risk for capital-relative sizing. Venue BNBUSDT.BINANCE USD-M daily is the landed BNB perp; futures RT taker (~0.10%) is cheaper than spot; long-only with leverage left at 1.0 (no amplification) so the leverage_set_but_unused gate does not trigger. ~6-15 entries/year clears the sparsity threshold while a wide trail keeps winners well above the round-trip fee.

Failed deflated Sharpe: DSR=0.0 (vs 0.95 bar), expected-max=4.19 over 225 trials — the selected optimized Sharpe of 1.01 is FAR below the best-of-N luck bar, so it is statistically indistinguishable from selection noise (is_significant=false). This is the rare case where the design's one genuine win (sensitivity PASSED, cliff_count=0 — it truly fixed the BNB Golden Cross fragility) is overridden by every forward/multiple-testing gate: PBO 0.7328 (73% overfit probability), sharpe_ci_low -0.7495 (CI straddles 0), walk-forward is_overfitted=TRUE with avg IS 1.45 collapsing to avg OOS 0.0 and ALL three OOS windows producing ZERO trades, and a FAILED holdout (holdout_sharpe 0.0, zero trades). Root cause is structural sparsity: 17 base entries over 6.3 years, optimizer thinned to 10, with the headline carried by ~2 outlier days (kurtosis 207-360, 2024-04 +26.6% and 2025-10 +11.3%) — a signal this rare cannot populate the walk-forward OOS or holdout windows, hence the zero-trade collapse. Not iterate (attempt 1 of 2): with zero trades across all three OOS windows AND the holdout, no parameter change can manufacture out-of-sample trades from a structurally rare trigger; the 17-trade base is already the densest config, so a second best-of-225 sweep re-finds another sparse, front-loaded, sub-luck-bar fit. Not revise_hypothesis: the multi-week trend-continuation mechanism is already captured by the promoted LINK 1D sibling (Sharpe 2.55) — BNB is the sparse/front-loaded arm of that family's asset-selectivity, not a proven mechanism stranded on a once-working, now-dead target. FAILURE PATTERN: single-asset daily multi-week trend-continuation (EMA-stack + SMA regime + N-day-high trigger) fires too rarely on BNB (~17 trades over 6.3yr, optimizer thins to 10) to survive walk-forward/holdout — even a clean, cliff-free sensitivity grid (cliff_count=0) cannot rescue it because best-of-225 selection drives a config with zero-trade OOS windows and a zero-trade holdout, failing DSR (0.0 vs a 4.19 luck bar) with PBO 0.73 and is_overfitted=TRUE. A successful cliff-fix and a smooth in-sample surface are not edge; a daily trend-continuation signal whose frequency is structurally too low is unvalidatable out-of-sample, the same sparsity signature as the abandoned ETH Bollinger lower-bounce sibling. The PSR=0.99 / DSR=0.0 split is the diagnostic tell.

Long-only daily multi-week trend-continuation on BNBUSDT.BINANCE USD-M futures, pure OHLCV single feed. Enters LONG when BNB is in a bull regime (close > 200-SMA and 30-EMA > 80-EMA), the 20-day return is positive, and the close makes a fresh 5-day high (continuation trigger). Managed by a wide ATR trailing stop (initialised at entry - 2.5x ATR, trailed by 4x ATR) with a 200-SMA regime-break exit and a 120-bar time stop. The signal is the continuous EMA-stack spread (30-EMA - 80-EMA)/80-EMA recomputed every bar; the regime+momentum+continuation condition is the discrete entry gate. Sized at 20% equity notional, capped so initial-ATR-stop risk <= 1.5% of equity. leverage 1.0.

Proven architecture: identical to the promoting LINK Multi-Week Trend (Sharpe 2.55); BNB momentum validated by BNB 4H Volume Breakout (2.03)

Edge is realized (end_unrealized_pct ~0) and recent (2024 +23.1%, 2025 +8.6%), with PF 2.46 and a contained max_drawdown of 16.2%

Clean sizing: avg_position 15% capped at 20%, leverage 1.0, no artifact; the +26.6% day is a multi-week trade's P&L booked at exit (engine quirk), not a sizing bug

Code is correct and faithful to the hypothesis (17 longs, no shorts); single-asset single-feed, no stale-leg risk

Only 17 trades over 6.4 years (~2.7/yr, below the claimed 6-15/yr) — thin for walk-forward (~5-6 per window) and prone to curve-fitting

Result dominated by ~2 outlier trades (skew 18, kurtosis 360); the entire net return is essentially 2024

Sharpe 0.58 with CI [-2.09, 1.27] straddling 0; negative in 2021, 2022, and 2023

Modest CAGR 4.77% — the recency-concentrated edge may not survive deflated-Sharpe/holdout testing

Achieved its stated hard gate: sensitivity PASSED with cliff_count=0 — the 30/80 EMA + 200-SMA design genuinely fixed the parameter-fragility (4 cliffs) that abandoned the BNB 50/200 Golden Cross. The grid is smooth (Sharpe ~0.47-0.71 across all ±20% nudges).

Optimization improved headline risk metrics (max_DD 16.2%->5.9%, Sortino 2.63->4.52, profit_factor 2.46->6.46) and trades as specified (long-only, 17 base entries, regime+momentum+continuation logic matches the hypothesis).

Code correctly implements the hypothesized mechanism (continuous EMA-spread signal cached, discrete decision read in should_enter — avoids the gated-to-0.0 anti-pattern).

Deflated Sharpe = 0.0 (is_significant FALSE): optimized Sharpe 1.01 is far below the expected-max luck bar of 4.19 over 225 trials — statistically indistinguishable from best-of-N selection noise.

PBO = 0.7328 (~73% probability of backtest overfitting, decisively > 0.5).

Walk-forward is_overfitted = TRUE: avg IS 1.45 collapses to avg OOS 0.0 with ALL three OOS windows = [0.0, 0.0, 0.0] (zero trades out-of-sample).

Holdout FAILED: holdout_sharpe 0.0, ratio 0 (zero trades in the holdout window).

sharpe_ci_low = -0.7495 — bootstrap 95% CI on Sharpe straddles 0.

Structurally sparse and outlier-dependent: 17 base trades over 6.3 years, optimizer thinned to 10; return_kurtosis 207-360, headline carried by ~2 days (2024-04 +26.6%, 2025-10 +11.3%). probabilistic_sharpe 0.99 is the PSR trap (ignores trial count).

BnbDailyMultiWeekTrendContinuationLong was an intentional repair of the abandoned BNB 50/200 golden-cross, swapping in shorter 30/80 EMAs plus a 200-SMA regime and a 5-day-high trigger to match BNB's episodic, catalyst-driven rallies. It succeeded at its stated goal — the sensitivity grid passed cleanly with zero cliffs, fixing the predecessor's fragility — and passed backtest review to enter full optimization, where headline metrics looked strong (Sharpe 1.01, profit factor 6.46, drawdown cut to 5.9%). But those numbers rested on just 10–17 trades dominated by two outlier days, and every robustness gate failed: deflated Sharpe of 0.0 against a 4.19 luck bar, 73% overfitting probability, and zero trades in all walk-forward OOS windows and the holdout. The analyst abandoned it as the classic sparse single-asset trend-continuation failure mode — a structurally too-rare signal that cannot be validated out-of-sample no matter how smooth its in-sample surface, with the multi-week mechanism already captured by the promoted LINK 1D sibling.

A smooth, cliff-free sensitivity grid fixes fragility but is not edge: a daily trend-continuation signal that fires only ~17 times in 6.3 years is structurally too sparse to populate walk-forward OOS or holdout windows, so best-of-225 selection only yields an overfit, sub-luck-bar fit.

It cleared the backtest-review gate (verdict: optimize) but was abandoned by the analyst after optimization for failing nearly every forward/multiple-testing gate: deflated Sharpe 0.0 vs a 4.19 expected-max luck bar over 225 trials, PBO 0.73, sharpe_ci_low -0.75, and a walk-forward (is_overfitted=true) plus holdout that produced zero out-of-sample trades. Its one genuine win — sensitivity passed with cliff_count=0 — could not override these.

A long-only BNBUSDT daily multi-week trend-continuation strategy combining a 200-SMA + 30/80-EMA regime filter, a 20-day positive-return momentum check, and a 5-day-high continuation trigger, deliberately using shorter EMA periods to fix the parameter-fragility that had abandoned the earlier BNB 50/200 golden-cross.

The base backtest took 17 long trades over ~6.3 years with a 29.4% win rate, profit factor 2.46, Sharpe 0.58 (CI [-2.09, 1.27]), max drawdown 16.2%, and a total return of ~22.9x. Optimization thinned it to 10 trades and improved headline risk metrics (Sharpe 1.01, profit factor 6.46, max drawdown 5.9%), but the gains were concentrated in ~2 outlier days (return kurtosis 207–360).