Strategies

BNB Dual-Timeframe Momentum Confluence, Long-Short (Single-Instrument BNBUSDT.BINANCE USD-M Perp, 4H Primary + 1D Trend Confirm, Flat on Disagreement, ATR-Trailing Exit, 3-Parameter)

Outcome: Paper Trading

BnbDualTimeframeMomentumConfluence4HLS

Outcome Summary

This strategy carried the dual-timeframe momentum confluence archetype, already promoted on ADA and BTC, over to BNB as a long-short 4H/1D trend follower with an ATR trailing exit. After one iteration it backtested to +88.7% over about 6.7 years (Sharpe 0.77, 460 trades, 0.68% per trade). It also passed sensitivity checks with no cliffs and had a positive holdout (Sharpe ~1.06, t 1.24). The analyst promoted it even though pooled walk-forward OOS was essentially zero (Sharpe 0.038) and the edge depends on trend regimes, waiving a marginal profit-factor soft floor. It is now in paper trading; live trading is blocked until protective orders can rest at the exchange.

Strategy report

Hypothesis

A LONG-SHORT, single-instrument, pure-price momentum-confluence trend follower on BNBUSDT.BINANCE (USD-M perpetual), 4H primary bars with a 1D trend confirmation. It is a deliberate, evidence-driven re-application of the ONLY archetype that has survived this factory's full pipeline to paper: dual-timeframe momentum confluence long-short (promoted on ADAUSDT Sharpe 1.60 and BTCUSDT Sharpe 1.99). The mechanism takes a directional position ONLY when the fast (4H) momentum and the slow (1D) trend AGREE, and stands flat when they disagree. BNB is chosen precisely because it is NOT the over-mined BTC/ETH/SOL series and has an independent fundamental driver (exchange-token / BNB-chain flows), so its multi-day trend regime is not just another correlated copy of the majors' price path — yet it is deep enough in liquidity that market impact (the killer of the SOL/ETH dollar-neutral pairs trade) is negligible for single-instrument directional sizing. This is NOT the volume-confirmed variant (L91) and carries NO regime/efficiency/volatility gate bolted onto momentum (L92): the confluence of two horizons IS the mechanism, and it is the exact structure that already generalized across two distinct majors (ADA and BTC), which is the empirical warrant for a third liquid major.

Implementation

Long-short dual-timeframe momentum confluence on BNBUSDT.BINANCE USD-M perpetual. Primary 4H bars produce a continuous, volatility-scaled trend read spread = (EMA12 - EMA34)/ATR14; an extra 1-DAY BNB leg produces daily_dir = sign(18-day return). A position is taken ONLY when the two horizons agree (spread>0 and daily_dir>0 -> long; spread<0 and daily_dir<0 -> short) and the strategy stands FLAT on disagreement. Exits on any of: an ATR trailing stop (give-back of trail_atr_mult x entry ATR from the best close since entry), a daily-trend flip against the position, or a 4H spread flip against the position. Risk-first sizing: qty = equity * 1.5% / (trail_atr_mult * ATR), capped at 25% of equity x 2x leverage of notional, so per-trade risk stays ~1.5% of equity regardless of the equity path.

Implements the hypothesis exactly: single-instrument, pure-price, long-short momentum confluence on BNBUSDT.BINANCE USD-M perp with a 4H primary and a 1D trend confirmation, flat on disagreement, ATR trailing exit, three tunables. No volume confirmation and no regime/efficiency/volatility gate is bolted on (per L91/L92) -- the two-horizon confluence IS the mechanism, matching the structure that already generalized to paper on ADA (Sharpe 1.60) and BTC (Sharpe 1.99). The daily leg is supplied via extra_instruments/extra_bar_types with BNBUSDT.BINANCE listed positionally alongside the 1-DAY bar type (the resolver pairs them by index), and since 1-DAY is slower than the 4H primary the base template's same-timeframe alignment barrier does not stall primary bars. calculate_signal returns the raw ATR-normalized EMA spread every bar (continuous, never gated behind the entry condition); the entry decision lives in should_enter. Only ema_ratio, daily_lookback and trail_atr_mult are optimizer-visible -- the fast EMA, ATR period, risk fraction and notional cap are read under underscore-prefixed keys that the walk-forward and sensitivity engines skip, holding the effective search surface at 3 dimensions to keep PBO down; the declared _param_bounds mirror the in-code clamps exactly and sit on the robust plateau the ADA sensitivity map identified. leverage=2.0 is genuinely consumed: the notional cap is max_notional_frac * equity * self.config.leverage. Futures venue is correct because the strategy takes short positions. All state is incremental (EMAs, bounded ATR deque, deque of daily closes), so per-bar work is O(1) and the 300s smoke budget is safe; entry fires whenever the 4H spread and the 18-day daily return share a sign, which on 4H BNB bars is a routine, multiple-times-per-month event, not a rare conjunction.

Verification Results

Acceptable as-is. No change needed for measurement.

should_exit() infers _side from the live signal when _side==0 (restart recovery). In a clean backtest _side is always set by should_enter before an open position exists, so this branch is inert here; it is a defensive guard for live restarts and does not affect backtest correctness.

Sandbox Sharpe is a modest 0.58 (CI [-0.19, 1.37] straddles zero) and win_rate is 0.35 with short_win_rate 0.31 vs long_win_rate 0.38 — the edge is asymmetric and right-tail dependent (largest_win 17959 vs avg_win 2851). The BNB port is the third re-application of the ADA/BTC dual-TF confluence archetype; per-trade return 0.94% clears the fee floor comfortably and 454 trades over 2399 days is decisively measurable, so this is not a blocker — flagging for the analyst to confirm the short leg carries its weight on full-history optimization rather than the long leg alone.

Analysis

Economics are well above costs. The full-history replay of the final parameters (ema_ratio 2.456, daily_lookback 18, trail_atr_mult 3.0) made 471 trades with an average trade of 0.626% against the 0.20% gate (the BINANCE USD-M fee floor is 0.15%). PF is 1.285, win rate 0.34 with a payoff ratio of 2.5, Sharpe 0.80, +94.6% total return and 19.2% drawdown against the registered 28% limit. Commissions are 5% of gross profit and modelled impact is 13% of gross.

The holdout (2025-06-16 to 2026-10-17, 488 days, 94 trades, not underpowered) is positive: Sharpe 1.06 to 1.14, t 1.24, consistent with the walk-forward OOS estimate (z +0.77). Its 19.1% drawdown is within the limit.

The sensitivity sweep found no cliffs. ema_ratio is flat across 2.2 to 3.4 (Sharpe 0.66 to 0.79) and trail_atr_mult is flat across 2.5 to 3.5. Every point sits within one Sharpe standard error (0.45) of the base.

All 7 calendar years are positive. It does not make its money in calm periods only: the stressed volatility tercile has Sharpe 0.82 and the calm tercile 1.06. Max loss on a single trade is about 3.0% of equity.

All 4 of the pre-registered predictions were met: 471 trades vs 150 predicted, drawdown 19.2% vs 28%, PF 1.285 vs 1.25, average trade 0.63% vs 0.5%.

This is a long-short strategy, not a closet long. Shorts make up 212 of the 460 trades, and the 19% drawdown is a fraction of BNB buy-and-hold's multi-year drawdown. Lesson 170's beta concern does not apply.

Pooled walk-forward OOS is essentially zero: Sharpe 0.038, t 0.08 over 12 windows from June 2021 to June 2025 (about 294 trades). 6 windows were positive and 6 negative. Re-tuning every four months added no edge over that period.

The edge depends on the trend regime. In the default-parameter replay, 2022 returned +2.3%, 2023 +3.2% and 2024 +0.8%. The profit comes from the 2021 and 2025 BNB trends, with 2021 alone carrying 47.9% of it (under the 60% floor). The holdout falls in a strong-trend period, so it flatters the long-run expectation.

The signal diagnostic shows almost no information on a one-bar horizon: IC 0.010, t 1.25. The strategy's value comes from the trailing-stop payoff shape, not from bar-level forecasting.

Inner PBO is high (0.43 to 0.98; a diagnostic only), and the walk-forward kept picking ema_ratio near its lower bound of 2.2. Parameter stability is moderate.

Two earlier BNB versions of this same confluence idea were abandoned as overfit. This run reaches the same near-zero walk-forward OOS, but it has a positive, consistent holdout and no hard failures.

PF 1.285 is just below the 1.3 soft floor.

Leveraged (2x) with no stop order resting at the exchange. That rules out live trading until venue-side protective orders exist; it does not block paper trading.

soft floor, waived below

inconclusive, t 1.24, consistent with OOS

HAC t>=2

advisory only

inconclusive (point estimate near zero)

Outcome Summary

A positive holdout taken during a strong-trend period can hide a near-zero walk-forward OOS edge. Most of this strategy's profit came from the 2021 and 2025 BNB trends, so paper results should be judged against how trend-dependent it is.

Backtest review sent it to optimization, and the analyst recommended promotion. The reasons: per-trade economics well above costs, no sensitivity cliffs, 7 of 7 positive years, all 4 pre-registered predictions met, and a positive holdout consistent with OOS. The profit-factor soft floor (1.285 vs 1.3) was waived, and the strategy moved to paper trading. Live trading is ruled out until venue-side stop orders exist, because it uses 2x leverage with no resting exchange stop.

A long-short trend follower on the BNBUSDT USD-M perp: it goes long or short only when 4H EMA momentum (scaled by ATR) and the 1D trend agree, stays flat when they disagree, and exits on an ATR trailing stop. It reuses the dual-timeframe confluence archetype that had already been promoted on ADA and BTC.

The initial backtest (2020-02 to 2026-10, 2x leverage cap) made 460 trades for +88.7% total return, Sharpe 0.77, 18.4% max drawdown, profit factor 1.27 and 0.68% average trade. Replaying the full history with the final parameters gave +94.6%, Sharpe 0.80, 471 trades and 0.626% per trade. The holdout (488 days, 94 trades) had Sharpe 1.06 to 1.14 at t 1.24, while the pooled walk-forward OOS Sharpe was only 0.038 (t 0.08).

Backtest Review

Sharpe: 0.77
Total return: 88.71%
Max drawdown: 18.35%
Trades: 460
Win rate: 33.7%
Profit factor: 1.27

Backtest and paper results are hypothetical. Trading involves risk of loss.