Binance 3-Major Confluence Trend Basket, Long-Short (Proven 4H+1D Momentum-Confluence Signal Applied INDEPENDENTLY per Name to BTC/ETH/BNB USD-M Perps, Equal-Risk Legs, Net Exposure Floats, MULTI-YEAR History, 3 Shared Parameters)

A LONG-SHORT, MULTI-INSTRUMENT, pure-OHLCV trend basket that runs this factory's ONLY optimization survivor — the 4H-primary + 1D-confirm momentum-confluence signal (BTC single-name version at Sharpe ~1.99, in paper) — INDEPENDENTLY on three decorrelated Binance USD-M majors (BTCUSDT, ETHUSDT, BNBUSDT) and holds the equal-risk portfolio average. The design is dictated by this session's failure analysis. The confluence signal itself works, but every attempt to prove NEW single-name variants keeps failing the multiple-testing gates (PBO > 0.5, DSR insignificant, holdout ratio < 0.70) because a single instrument's Sharpe is too noisy for a 225-trial optimizer to distinguish from best-of-N luck, and the Hyperliquid basket that would have diversified this died on HL's ~2.5-year history. This hypothesis fixes BOTH at once: (1) DIVERSIFICATION suppresses overfit — averaging three independent confluence legs cuts the idiosyncratic variance of the portfolio return, which directly lowers PBO and raises deflated Sharpe (a more stable objective is far harder to noise-select), the textbook antidote to the exact gate failures that killed the single names; (2) BINANCE MULTI-YEAR HISTORY (4+ years on all three names) delivers 300+ combined trades and long multi-regime walk-forward windows — clearing the ~100-trade measurability floor and giving the OOS enough data to hold up, the specific ingredient the HL basket and HL single-names lacked. It uses only 3 parameters SHARED identically across all three legs (no per-asset tuning), so parameter dimensionality is no higher than a single-name strategy while the edge is averaged over three assets. Asset choice maximizes decorrelation: BNB (exchange-token, burns, launchpad cycles) is the lowest-BTC-beta major, so its confluence timing differs from BTC/ETH, improving the aggregate Sharpe. It is materially different from every pipeline item: the single-name confluence ports (BTC/ETH/BNB) are correlated one-legged bets; the HyperliquidFiveMajorTrendFollowingBasket uses a dual-MA-state signal on HL's short/hostile history — THIS is the PROVEN confluence signal on Binance's multi-year history at the portfolio level. Fills long_short direction (13.6% vs target) and multi-instrument scope. Same venue, no cross-venue schema landmine. Pure OHLCV, no supplementary feed (avoids the short-history L/S-ratio/liquidations data_unavailable trap). I acknowledge it adds to the over-represented BINANCE venue; multi-year history — the proven survival ingredient — exists only there.

I reused the validated confluence signal verbatim per leg and shared all three parameters across names, so parameter dimensionality matches a single-name strategy while the return stream averages three assets. The hypothesis's central claim is testable, so I tested it on real aligned data (14,164 bars, 2020-02-10 onward, 0.10% fees) rather than asserting it: single legs score Sharpe +0.31 (BTC), +0.52 (ETH), +0.52 (BNB), averaging +0.45, while the basket scores +0.55 on 2,359 combined trades. Crucially the leg-RETURN correlations — BTC-ETH 0.607, BTC-BNB 0.459, ETH-BNB 0.504 — sit well below the underlying asset correlations (0.84 and 0.72), because the confluence rule times each name independently, and the realized Sharpe gain matches the theoretical sqrt(3/(1+2*rho)) prediction at rho≈0.52 almost exactly. So the diversification mechanism is real, quantified, and behaves as the theory says it should, which is the property that should lower PBO. That measurement also caught a sizing trap I then fixed: at the single-name risk setting the three correlated legs stack to ~0.87x gross and their drawdowns compound, giving the basket a 42.4% drawdown against 19-25% for any single leg — near the abandon line and exactly the 'diversification made me feel safe, so I sized up' error. I cut per-leg risk from 1% to 0.45%, which leaves Sharpe unchanged (scaling is Sharpe-neutral) while bringing expected drawdown to roughly 25%, and I set the per-leg cap at 0.30 so it acts as a rare tail guard rather than an always-binding clamp that would render the risk sizing inert. On structure: all three legs share venue and timeframe, so the base class's alignment barrier defers the basket until every leg reports — which is what a portfolio decision needs — and it is safe because the three series share 14,164 timestamps with zero gaps; all legs including the primary are managed in one per-leg loop with the base hooks inert, so expect Layer 3's 'should_enter returned a side 0 times' diagnostic. Two caveats for the analyst. The trailing year is weak (+1.9%, Sharpe +0.24), consistent with all three single-name ports being flat to negative recently, so the +0.55 full-sample Sharpe is not evenly distributed. And +0.55 is modest in absolute terms: diversification here buys stability and a harder-to-noise-select objective, not magnitude — it makes a thin edge more robust, it does not make it large. One factual correction to the hypothesis's framing: BNB is described as the lowest-BTC-beta major, and by price correlation it is not (DOGE 0.551, XRP 0.620, SOL 0.651 all rank below BNB's 0.715), though BNB does happen to be the best diversifier of these three by leg correlation, so its inclusion is justified on the measurement that actually matters here even if the stated reason is wrong.

No promotable edge, and the hypothesis's central diversification-as-robustness claim is empirically falsified by its own backtest. Over the largest possible sample (2148 trades, 6.6 years) the basket Sharpe is 0.319 — LOWER than the single-name ETH-Binance leg (0.43) already abandoned this session — because the three legs remained positively correlated (0.46-0.61) so drawdowns compounded (max_drawdown 27%, CI high 83.4%) instead of diversifying away. profit_factor 1.116 is < 1.2 (direct L9 abandon signal for OHLCV momentum, zero survivors), Sharpe CI [-0.29, 0.99] straddles zero, and information_ratio -0.539 shows negative risk-adjusted active value vs the benchmark. On top of the thin edge there is a fatal capacity blocker: impact_cost_pct 22.58% with capacity_usd only $1.96M — the edge is real only at toy scale, with market impact consuming ~a quarter of gross PnL ($41k on $100k). avg_trade_return_pct 0.459% clears the fee floor, so this is not a fee death, but the well-measured true edge (~0.32 Sharpe / 1.12 PF) is far below the 1.5 promotion floor and cannot be lifted by 3-parameter tuning without overfitting — any sub-window showing 1.5 would be a best-of-N selection that PBO/DSR reject, the exact death every sibling in this dual-TF-confluence family suffered. The mechanism is correctly coded and two-sided (no bug to iterate on). Abandon rather than spend 2 hours repeating the optimize→overfit-reject cycle on a diversification thesis the data already disproves.