BNB 1H Volume-Confirmed Strong-Close 3-Day-High Break, Fixed 24h Clock, Long-Only, Flat ~84% of the Time (BNBUSDT.BINANCE, pure OHLCV, single leg)

Hypothesis

Reject this registered BNB hourly implementation at fixed 90% equity sizing for failure of its risk and temporal-persistence contract: no risk-admissible candidate across nine search windows, 47.114% full-history drawdown versus the declared 20% limit, and only 3 of 7 exposed years profitable versus the supplied enforced 60% floor. Average net trade return 0.3463% clears the supplied 0.20% economics floor. Inconclusive significance is not the abandonment reason. No demonstrated code defect or validated risk-admissible parameter region supports developer iteration. This does not falsify volume-confirmed breakout generally or establish an accounting/data artifact.

Implementation

Long-only BNB USD-M perp. Enters when an hourly bar closes above the prior 72-bar high, closes with CLV > 0.7, and trades volume above 1.5x the median of the prior 168 bars. Exits on a fixed 24h clock counted from the position's open time. The signal is the % distance of the close above or below the prior 72-bar high, computed every bar.

Implements the hypothesis rule exactly: a close through the 3-day high, a strong close and volume confirmation, then a 24h hold, flat otherwise. The trading rules are unchanged in iteration 3. This is the second sandbox_timeout in a row, and it does not come from the strategy code. Run on its own, the per-bar logic took 0.11 s for 20,000 bars (about 5 us per bar, O(log n)). The strategy reads no supplementary data and keeps no buffers that grow over the run. Following L155(c), the only edit is a docstring recording this diagnosis. The likely cause is the catalog query, which earlier profiles of other strategies showed takes most of the 300 s, on a host with load around 38. That needs a pipeline fix; no change to the strategy code can shorten it.

Verification Results

Allow calculate_signal to run from the first bar, using its existing prior-window readiness checks to suppress entries until 168 prior bars exist, or deliberately seed its rolling windows from the framework history without double-processing the current bar.

The framework does not call calculate_signal until min_bars_required() bars have accumulated. Returning 169 therefore leaves the private rolling windows empty through bar 168; they then need another 168 prior observations, making the first eligible signal occur on bar 337 rather than bar 169. This discards otherwise eligible early signals and lengthens startup warmup, but does not make entries unreachable.

The smoke window reports -0.0501% average net trade return, PF 0.907 and 24.97% drawdown at approximately 90% equity deployment. Modeled impact is substantial ($9,721); full-history review should assess execution costs and the hypothesis's stated dependence on 2021. These are advisory profitability concerns.

Backtest Review

Sharpe

0.54

Total return

137.13%

Max drawdown

47.11%

Trades

399

Win rate

45.9%

Profit factor

1.14

399 closed long trades, all held 24 hours, and 16.57% exposure match the hypothesis. Code uses prior-bar channel and volume windows.

Closed-trade PnL reconciles to +$137,133.10. Average net trade return 0.3463% exceeds the 0.15% Binance review reference; PF 1.139 and Sharpe 0.540 support proceeding.

QA’s negative smoke-window expectancy is not borne out over full history, but its cost and concentration concerns remain: impact consumes 45.62% of gross PnL, 2021 dominates gains, and four of seven calendar years lose.

90% equity deployment produces 47.11% drawdown. Sampled buy-and-hold returns approximately 2,862.51% with 69.00% drawdown; negative reported alpha and information ratio warrant scrutiny during optimization.

Historical entry conditions were checked in code rather than independently reconstructed from signal-bar OHLCV. The one-hour forecast diagnostic does not directly test the hypothesized 24-hour payoff.

Analysis

All 399 closed trades are long and held exactly 24 hours, matching the hypothesis. Closed-trade PnL reconciles to $137,133.10, with no terminal unrealized gain.

Average net trade return 0.3463% clears the supplied 0.20% economics floor. This is not a fee-floor rejection.

Sensitivity shows broad positive development Sharpe; the strong-close cliff lies above the permitted search range.

None of nine search windows produced a risk-admissible candidate. All 270 window trials breached the registered 20% inner-training drawdown constraint; reported drawdowns ranged from 25.03% to 56.23%. Flat fallback does not establish negative traded OOS performance.

Final parameters remain the baseline: Sharpe 0.541 versus initial 0.540, unchanged 399 trades and 47.114% drawdown. No accepted improvement was produced, and equity_fraction 0.9 is contract-fixed.

Only 3 of 7 exposed calendar years are profitable versus the supplied enforced 60% persistence floor. 2021 contributes 118.6% of dollar profit.

The 487-day holdout has 79 trades, Sharpe 0.413 and drawdown 17.898%, but significance is inconclusive because required evidence is missing. Pooled traded OOS significance and holdout consistency remain unknown; this is not a statistically rejected holdout.

Full-history trade count, win rate and average trade return meet research predictions. PF 1.139 misses predicted 1.25; drawdown exceeds the declared 20% risk limit. Predicted OOS Sharpe cannot be assessed. Prediction misses are not additional gates.

Buy-and-hold return is approximately 2862.51%, with drawdown at least 69.00% on the sampled detail curve. Strategy return is 137.13%, with negative reported alpha and information ratio -0.9602.

The numeric-gate statement that no optimization report exists conflicts with the attached completed attempt and is not used as the scientific rejection reason.

soft promotion floor; separately enforced search risk constraint

inconclusive

supplied hard failure

Iteration History

sandbox_timeout: Verification failed (Layer 3 — sandbox backtest): smoke test exceeded the 300s wall-clock limit on a ~20000-bar window (833 days of BNBUSDT.BINANCE-1-HOUR-LAST-EXTERNAL). This almost always means per-bar work that scales with history — e.g. rescanning the full funding/supplementary series, or rebuilding a list and calling min()/sorted() inside calculate_signal()/on_bar() on every bar. Precompute sorted timestamp arrays ONCE in __init__ and use bisect, or cache lookups keyed by timestamp, so per-bar cost is O(log n) not O(n).

Outcome Summary

An economically positive breakout needs risk-admissible sizing and more consistent performance across years to satisfy this strategy's registered contract.

Backtest review recommended optimization, but the analyst abandoned the fixed 90% equity implementation because all 270 trials across nine search windows breached the registered 20% training drawdown constraint, and profitable-year share was 42.9% versus the enforced 60% floor. Trade economics passed; inconclusive significance was not the abandonment reason.

Buy BNB perpetuals after a strong-close hourly breakout above the prior 72-hour high with volume above 1.5 times the prior 168-hour median, then hold for 24 hours.

The full-history backtest returned 137.13% across 399 trades, with Sharpe 0.540, profit factor 1.139 and average net trade return 0.3463%. Maximum drawdown was 47.11%, and only 3 of 7 exposed calendar years were profitable.