BnbFundingConfirmedTrendLS

BNB Perp Funding-Confirmed Trend (Single-Instrument USD-M, Long-Short, Daily, Carry-Momentum Confluence, Flat on Disagreement)

A long-SHORT, SINGLE-INSTRUMENT, single-venue directional strategy on BNBUSDT.BINANCE USD-M perpetual futures using DAILY bars, that takes a position ONLY when the price trend and the funding rate AGREE in direction, and stays FLAT when they disagree. This extends the factory's ONE promoted mechanism — the ETH Perp Daily Funding-Confirmed Trend (Sharpe 2.74, PAPER) — to the most DIVERSIFYING liquid major with clean data. EVERY design choice respects this session's accumulated kills: (1) SIGNAL = FUNDING, the single safest crypto-native stream. This session HARD-BLOCKED open_interest, taker_ratio, and top/global L/S account ratios as primary signals (only ~30-56 days of catalog history → walk-forward degenerate, multiple identical abandons); funding_rates is the lone supplementary stream with proven multi-year coverage on majors (it cleared the promoted ETH strategy's data-span gate). BNB is a top-tier perp with full-history funding and price collection. (2) ASSET = BNB for genuine DIVERSIFICATION — its exchange-token / ecosystem driver produces idiosyncratic trends lowly correlated with the BTC/ETH/SOL funding-confirmed-trend siblings, so it adds a distinct return stream rather than duplicating major beta; and BNB is a fully-collected symbol, avoiding the alt-collection gaps (ARB/AAVE/NEAR/ATOM) that failed the reversal basket outright. (3) NOT a fade / mean-reversion — intraday VWAP-MR and ratio-MR were shown anti-predictive on trending crypto (excursions trend through fair value); this is a confluence-gated TREND, and BNB is lower-vol with cleaner trends than SOL — the right profile for trend-confirm (the opposite of the VWAP-fade that needs ETH's specific regime and died on SOL). (4) NOT naked trend/breakout (regime-dead across 14+ abandons) — the funding-confluence AND-gate plus flat-on-disagreement filters the unsupported whipsaws the 2024-26 chop punishes; this gate is the proven differentiator that carried ETH to PAPER. (5) NOT a closed family: NOT pairs-ratio (single asset), NOT delta-neutral funding CARRY (funding is a directional SIGNAL on a single price-creditable leg, never booked as cash flow), NOT cross-venue/HL, NOT COIN-M (engine-blocked), NOT supplementary-stream-primary. (6) NOT a rare-conjunction gate — continuous funding SIGN/level fires across all regimes, populating every walk-forward OOS window densely (anti-event-sparsity). Intentionally minimal: 3 tunables (trend_lookback, funding_lookback, funding_deadband).

The Layer-3 failure was 'No trades produced / should_enter never triggered' across 2268 bars. The signal logic is correct; the bug is in _build_funding, which filtered keys with 'if not isinstance(k, int): continue'. The funding_rates supplementary dict is delivered to the backtest subprocess via JSON, which stringifies all dict keys (and the source dict already carries dual int+str keys), so every key failed the isinstance(int) test, _fund_ts stayed empty, _funding_available was False, and should_enter returned None on every bar -- the classic funding-gated zero-trade signature. The minimal fix parses each key with int(k) inside a try/except (accepting string keys) and de-duplicates by timestamp via a dict (so dual int/str keys for the same event aren't double-counted), then builds the sorted ts/rate arrays as before. Nothing else changes: the daily-funding windowing, 20-day trend sign, deadbanded funding sign, confluence AND-gate, flat-on-disagreement and stop exits, the continuous confluence signal, and capital-relative sizing are byte-identical, so the passing Layer-1/2 checks and the promoted-ETH mechanism are preserved. Once funding loads, the confluence gate fires across the full multi-year BNB history (funding has proven multi-year coverage on majors), populating walk-forward OOS windows densely. The no-price-fallback design is retained: if funding genuinely were absent the strategy stands aside. leverage stays 1.0 (single directional leg, no amplification) so the unused-leverage gate does not apply.

Failed deflated Sharpe decisively: DSR=0.0249 (vs 0.95 bar), with the optimized Sharpe of 0.49 FAR BELOW the expected-max luck bar of 1.35 over 225 trials — after honest multiple-testing correction the edge is indistinguishable from best-of-N noise (is_significant=FALSE, PBO 0.549, sharpe_ci_low -0.62 straddling 0). This exactly confirms the pre-optimization warning: the kurtosis-166 / CI-through-0 base predicted a deflated-Sharpe failure, and the optimized config (return_kurtosis 166.9) is still entirely 2021-outlier-carried — the two 2021 single days (2021-04-17 +21.8%, 2021-02-24 +13.3%) drive a +69.5% 2021 while 2020 (-19.9%), 2022 (-4.5%) and 2026 (-2.4%) are negative. The walk-forward is is_overfitted=TRUE with avg IS 0.945 collapsing to avg OOS -1.00 (windows [-3.10, +1.21, -1.12], two negative), and sensitivity has a CLIFF on the core trend_lookback parameter. The holdout 'pass' is spurious (ratio 0 computed against a negative WF-OOS of -1.00). Not promote: DSR below the luck bar + is_overfitted with negative average OOS + CI through 0 + PBO>0.5 + a core-param cliff are each disqualifying. Not iterate (attempt 1 of 2): the pre-opt promote_notes pre-committed to abandon if the edge collapsed outside 2021, and it did (avg OOS -1.00); with the best-of-225 Sharpe below the luck bar and a cliff on trend_lookback there is no robust region to tune toward, so a second sweep re-finds the same 2021-outlier fit. Not revise_hypothesis: the funding-confirmed-trend confluence mechanism is validated and already promoted on ETH (Sharpe 2.74), so this is BNB-specific asset-selectivity, not a sound mechanism stranded on a dead target (and the SOL sibling of this same transfer also failed at pre-opt). FAILURE PATTERN: the ETH-promoted funding-confirmed-trend confluence gate does NOT transfer robustly across the major-asset spectrum — on BNB it produces a kurtosis-166 edge carried entirely by two 2021 outlier days, failing deflated Sharpe (0.025, optimized Sharpe 0.49 below a 1.35 luck bar), is_overfitted with a negative average OOS (-1.00), and a cliff on the core trend parameter. A validated-on-ETH mechanism with a marginal base (Sharpe 0.42, CI through 0, kurtosis 165) does not survive best-of-225 deflation when its edge is outlier-concentrated in a single bull year; the promoted ETH instance is the asset where this gate works, and transferring it to BNB/SOL only re-fits 2021 mania.

Long-short single-instrument BNBUSDT.BINANCE USD-M daily funding-confirmed trend strategy. It takes a position only when the price trend (sign of the 20-day return) and the funding rate (sign of the 3-day mean daily funding, with a small deadband) AGREE in direction -- long when both bullish, short when both bearish -- and stays flat on disagreement or a held-side agreement break, plus a 10% protective stop. Funding is a directional signal only, never booked as cash flow. Iteration 2 fixes the funding loader: keys are parsed via int(k) (handling the subprocess JSON-stringified timestamp keys) and de-duplicated, so the funding series actually loads and the confluence gate fires -- the previous isinstance(k, int) filter dropped every key and produced zero trades. Three tunables: trend_lookback, funding_lookback, funding_deadband.

Verification failed (Layer 3 — sandbox backtest): No trades produced Bar type used: BNBUSDT.BINANCE-1-DAY-LAST-EXTERNAL, Bars processed: 2289 Diagnostics: should_enter() returned a side 0 times over 2268 evaluated bars -> your ENTRY CONDITION never triggered. Loosen the entry logic / thresholds. Ensure your strategy produces trades with the given data and parameters.

Validated mechanism: the funding-CONFIRMED-trend confluence AND-gate that carried the promoted ETH sibling to PAPER (Sharpe 2.74) — NOT the dead funding-as-direction family; correctly uses full-history funding, not the data-walled OI/L-S feeds

Clean testable profile: 226 trades over 6.4 years, two-sided (107L/119S), capacity $24M, impact 6.5%, commission 3.46%, max DD only 10.3% — no data/capacity/inverted-edge blocker

Two positive windows: 2021 (+25.1%) and 2025 (+12.9%); the 2025 gain is distributed (not one spike), so not purely a 2021 artifact; diversifying exchange-token driver

Marginal base: Sharpe 0.42 with CI straddling 0 (-0.73 to +1.18), PF 1.20 (borderline), negative in 4 of 7 years

Outlier-influenced: return_kurtosis 165.8 — two 2021 days (+19.7%, +13.0%) are ~32 of the 41.6% total return; the deflated-Sharpe gate (which penalizes kurtosis) is a real risk

Mechanism transfers weakly from ETH (Sharpe 2.74) to BNB (0.42) — the optimizer must find a genuinely robust BNB region, not curve-fit the 2021 spikes

deployable

<~20 (not outlier-carried)

Validated mechanism lineage (the funding-confirmed-trend confluence gate is promoted on ETH at Sharpe 2.74); clean funding-only signal, two-sided, full-history data, deployable capacity

One walk-forward OOS window positive (+1.21) and a positive holdout Sharpe (0.89) — the gate does fire correctly in some regimes

Failed deflated Sharpe decisively: DSR 0.0249, with optimized Sharpe 0.49 FAR below the expected-max luck bar 1.35 — the 225-trial search found nothing above noise; is_significant=FALSE, PBO 0.549, sharpe_ci_low -0.62 (straddles 0)

Walk-forward is_overfitted=TRUE: avg IS 0.945 -> avg OOS -1.00, windows [-3.10, +1.21, -1.12] (two negative)

Sensitivity cliff on the CORE trend_lookback parameter (is_cliff=TRUE)

return_kurtosis 166.9 — the edge is entirely the two 2021 outlier days (2021-04-17 +21.8%); annual returns 2021 +69.5% carry everything, with 2020/2022/2026 negative

Holdout pass is spurious (ratio 0 against a negative WF-OOS of -1.00)

BnbFundingConfirmedTrendLS extended the factory's one promoted strategy — the ETH daily funding-confirmed trend — to BNB, choosing it for genuine diversification via its exchange-token driver and respecting every data and mechanism kill of the session (funding signal only, full-history data, confluence AND-gate, flat on disagreement). The base backtest was clean and low-drawdown but only marginally profitable, with a Sharpe CI through zero and a return kurtosis of 166 flagging that two 2021 days carried most of the gain. Optimization confirmed the warning: the best-of-225 Sharpe (0.49) fell far below the 1.35 luck bar, the deflated Sharpe collapsed to 0.025, the walk-forward was overfit with a negative average OOS, and a cliff appeared on the core trend parameter. The analyst abandoned it on its second iteration, concluding the ETH-validated confluence gate is asset-selective — on BNB it merely re-fits 2021 mania — and that the promoted ETH instance, not BNB or SOL, is the asset where this mechanism actually works.

The ETH-promoted funding-confirmed-trend confluence gate does NOT transfer robustly across the major-asset spectrum — on BNB (and SOL) it produces a kurtosis-160+ edge carried entirely by two 2021 bull-mania days, so a validated-on-ETH mechanism with a marginal base (Sharpe 0.42, CI through zero) does not survive best-of-N deflation when its edge is outlier-concentrated in a single year; ETH appears to be the specific asset where this gate works.

On its second and final iteration the analyst issued an 'abandon' verdict at the post-optimization ANALYZING stage: it failed the deflated-Sharpe gate decisively (DSR 0.0249, optimized Sharpe far below the luck bar, is_significant=FALSE), with PBO 0.549, is_overfitted=TRUE and a negative average OOS, a core-parameter cliff, and a spurious holdout 'pass' (ratio 0 against a negative WF-OOS), the edge being entirely carried by two 2021 outlier days.

A long-short, single-instrument directional strategy on BNBUSDT.BINANCE USD-M daily perpetual futures that took a position ONLY when the price trend and the funding rate AGREED in direction (long on up-trend + positive funding, short on down-trend + negative funding) and stayed FLAT on disagreement — a transfer of the factory's one promoted survivor (ETH Daily Funding-Confirmed Trend, Sharpe 2.74) to BNB for diversification via its idiosyncratic exchange-token driver, with funding as a directional signal only.

The base backtest was marginal and outlier-influenced: Sharpe 0.42 with CI straddling zero (-0.73 to +1.18), profit factor 1.20, total return +41.6% over 226 two-sided trades, max drawdown only 10.3%, $24M capacity, but negative in 4 of 7 years with return kurtosis 165.8 (two 2021 days, +19.7% and +13.0%, accounting for ~32 of the 41.6% total); optimization made it worse — optimized Sharpe 0.49 vs a 1.35 luck bar, walk-forward is_overfitted=TRUE (avg IS 0.945 collapsing to avg OOS -1.00, windows -3.10/+1.21/-1.12), and a sensitivity cliff on the core trend_lookback.