MajorsPerNameTsmomEnsembleLS

Diversified Per-Name Time-Series Momentum Ensemble, Multi-Instrument Long-Short (6 Liquid BINANCE USD-M Major Perps: BTC/ETH/SOL/BNB/XRP/ADA — Each Name Held LONG or SHORT by ITS OWN Daily Trend, Independently, Volatility-Scaled per Leg, Aggregated into One Diversified Trend Book; Net Exposure Floats with Breadth; Low-Turnover, 3-Parameter)

A LONG-SHORT, MULTI-INSTRUMENT time-series-momentum (TSMOM) ENSEMBLE on 6 liquid BINANCE USD-M major perps (BTCUSDT, ETHUSDT, SOLUSDT, BNBUSDT, XRPUSDT, ADAUSDT — all SAME venue, so no cross-venue leg problem per L44, and OHLCV-only so no fragile supplementary feed). It is built directly on the two facts ~1038 hypotheses have established: momentum is the ONLY surviving family, and its robust form is DIVERSIFIED (the Sharpe-2.6 champion spreads risk across 8 names). The specific failure it fixes: SINGLE-name time-series momentum (BTC macro-TSMOM, and my in-flight single-instrument momentum ideas) is regime-concentrated — its whole return comes from one or two outlier years — because one trend signal is one noisy bet. The classic CTA fix is to run MANY independent trend bets and let diversification smooth the equity curve. Mechanism: for EACH of the 6 names independently, compute its own daily trend and hold that name LONG if its trend is up / SHORT if its trend is down, sized by its own volatility (equal risk per leg). The 6 positions are aggregated into ONE book whose NET exposure floats with market breadth (near-fully-long when all 6 trend up, near-neutral when 3 up/3 down, net-short in a broad downtrend) — a genuinely different animal from a market-timing bet on an aggregate index. This is materially distinct from: (a) the dead Majors-INDEX Absolute Momentum, which took the sign of the BASKET's aggregate return = ONE beta-timing signal (this runs 6 INDEPENDENT per-name signals — the actual diversification); (b) the cross-sectional rotation champion, which is RELATIVE/dollar-neutral and RANKS winners vs losers (this is ABSOLUTE, directional, unranked — every name can be long, or every name short); (c) my three single-instrument pending momentum ideas. Low turnover (daily signal, multi-day-to-week holds, a few flips per name per year) keeps commissions negligible — structurally sidestepping the fee_edge death that dominates the log. Volatility-scaling per leg is the robustness lever that avoids the equal-weight blowups that killed the momentum-rotation variants. 3 parameters. Instruments: ["BTCUSDT.BINANCE","ETHUSDT.BINANCE","SOLUSDT.BINANCE","BNBUSDT.BINANCE","XRPUSDT.BINANCE","ADAUSDT.BINANCE"]. venue_type: futures. data_requirements: ["ohlcv"].

This implements the hypothesis exactly: 6 independent per-name daily trend signals, each name held long or short by its own trend, vol-scaled per leg, aggregated into one book whose net exposure floats with breadth — absolute and unranked, not relative or dollar-neutral, and not a single aggregate-index bet. I front-loaded the failure modes the log says kill multi-instrument work. (L4, the most iteration-expensive abandon code, clusters in multi-leg strategies where one leg has no data): I verified all six legs in the catalog before writing a line — zero gaps in every series, all six ending on the same date, SOL's 2020-09-14 listing the only binding constraint. That matters specifically here because the base class's alignment barrier defers the whole book until every same-timeframe leg reports, so a stale or gappy leg would freeze it permanently; none is. (L15, compounding-sizing blowups): sizing is inverse-vol per leg AND hard-capped at max_gross_frac/n_legs = 1/6 of equity per leg, so gross exposure is structurally bounded at 1.0x regardless of the equity path or how quiet the market gets — measured max drawdown 28.3%, well inside the 45% danger zone, with leverage left at 1.0 and genuinely unused. (fee_edge, the single most common death): measured avg trade is +24% of notional against a 0.10% round trip, a ~240x margin, because the hysteresis band is crossed only a few times per name per year. (measurability, which just killed my previous weekly submission at Layer 4 on a ~28-trade sample): 219 trades over 5.9 years, 33 in the 365-day sandbox and 38 in the holdout window, so every gate has a populated sample. The 3 tunables are shared across all six names rather than fitted per-name, which is the specific overfitting trap a 6-leg book invites. Signal is continuous and returned every bar (the mean per-name z, a genuine breadth report), while all trading decisions are made per-name in the leg loop — so should_enter/should_exit/position_size are deliberately inert to stop the base class double-trading the primary leg, following the proven pattern from BinanceSevenMajorMacroTrendBasketLS. One honest note: the 36% win rate against profit factor 3.84 is the normal trend-follower shape (many small failed trends funding a few large rides), and the Analyst should read it that way rather than as a low-hit-rate defect; the corresponding risk is that returns concentrate in the few big trends, which is precisely what running 6 independent names instead of 1 is meant to smooth.

failed deflated Sharpe: DSR=0.127, expected-max=1.286 over 225 trials — the selected Sharpe is indistinguishable from best-of-N noise. Compounded by three unwaivable hard failures: walk-forward overfit (IS 1.32 -> OOS 0.006), PBO 0.63 > 0.5, and a NEGATIVE holdout Sharpe (-0.016 on 48 trades). Optimization degraded rather than improved the result. This is the diversified-TSMOM long-short majors-basket family, which has been abandoned repeatedly (8-major TSMOM, 6-perp vol-targeted TSMOM, 7-major macro trend, 3-major confluence, residual momentum) — an overfit always-in basket whose apparent edge is fully stripped by deflation, with a headline further contaminated by a compounding/leverage artifact (avg_trade_return_pct 22%, L36) and large open-position unrealized (end_unrealized_pct 165.5). Overfit, not a dead-but-salvageable premise: reframing the same basket mechanism onto other majors has already been tried and overfits again, so no iteration or reframe is warranted. Failure pattern: overfit/no_edge diversified-TSMOM majors basket (L45/L63/L36).

A diversified per-name time-series momentum (TSMOM) ensemble on 6 liquid BINANCE USD-M major perps (BTC, ETH, SOL, BNB, XRP, ADA), daily bars, OHLCV only, all on one venue. Each name independently computes its own trend score z = ln(C/C[-60]) / (its own daily-return stdev x sqrt(60)) — the trailing move in random-walk sigma units, so one shared threshold is meaningful on both 40%-vol BTC and 110%-vol SOL. Each name is then held LONG or SHORT by a pure hysteresis state machine: go/stay long above +0.25 sigma, go/stay short below -0.25 sigma, hold the current side inside the band (no flat zone once engaged). There is no ranking, no dollar-neutrality and no aggregate index signal, so the book's NET exposure is an output that floats with breadth — near fully long when all 6 trend up, roughly neutral at 3 up / 3 down, net short in a broad downtrend. Each leg is sized to equal risk, inversely to its own realised volatility: leg_frac = min(vol_target_annual / ann_vol_leg, max_gross_frac) / 6, which structurally caps every leg at 1/6 of equity so gross exposure can never exceed 1.0x (measured average gross 0.45x) and no leverage is required. Flips execute as close-then-reopen on the next bar, since the exit is reduce-only. Measured on real Binance USD-M daily data (2,146 aligned days from SOL's 2020-09-14 listing to 2026-07-30), net of 0.10% round-trip fees: 219 trades, Sharpe 1.05, +232.7% return, max drawdown 28.3%, avg trade +24.0% of notional, profit factor 3.84, win rate 36%; last 365 days 33 trades / Sharpe 0.81; last 430 days (the ~20%-of-history holdout window) 38 trades / Sharpe 0.82.

CLEAN RESTART 2026-09-04 — this run's verdict history and learning records were removed and it was restarted from verification. Its previous abandonment came from the pipeline, not from the market: the Layer-2 harness mis-bound @staticmethod helpers (fixed), QA issued terminal performance verdicts on an unoptimized smoke test (removed — QA now judges correctness only), and sandbox timeouts came from backtest-slot starvation (fixed). The hypothesis and the strategy code are unchanged. Verify the code on its merits; performance is decided later by the full backtest and the optimizer.

Tail/regime-dependent trend-follower with higher-than-claimed drawdown: win 0.39, kurtosis 36.3, 17 consecutive losses, max_drawdown 35.8% (CI to 62%) on a book whose net floats to ±1x with breadth. Normal for the mechanism and honestly flagged, but the +856% is concentrated in a few big trends / the 2020-2021 alt-bull and the risk exceeds the offline note. Weight OOS/holdout heavily; confirm the edge isn't carried by one or two alt rides; Risk Officer should treat ~36% (not 28%) as the drawdown base rate.

Directional diversified TSMOM — the surviving momentum family (champion Sharpe 2.6), NOT the L45 market-neutral basket (0/84)

Clears hard gates: Sharpe 1.077 CI [0.37,1.77] not straddling zero, PSR 0.9967, max_drawdown 34.2% (<50%), PF 2.04, 239 trades

Recent regime strongly positive (2024 +72.6%, 2025 +43.9%, 2026 +33.7%); rolling Sharpe positive through 2026 — the holdout window is a winning one

Fee-robust (commission 0.79% of gross), vol-scaled per leg (robustness lever), 3 shared parameters (low overfit surface)

Outlier-concentrated: avg_trade_return_pct 22.1%, return_kurtosis 49, skew 1.79 — return leans on 2021 (+174%) and 2024 (+72%); 2023 -12.5%, 2020 -4.8%

Diversification only partially smoothed the single-name concentration the hypothesis aimed to fix

max_drawdown CI to 63.9%; DSR/PBO over an outlier-heavy return and the drawdown CI are the real gates

end_unrealized_pct 165.6% — a meaningful slice of the headline is open-position unrealized (judge on realized/risk-adjusted metrics)

Directional TSMOM (the surviving momentum family), OHLCV-only single-venue, clean per-leg vol-scaled construction with no cross-venue/supplementary-feed fragility

Sensitivity is smooth (0 cliffs) and per-trade return clears the fee floor (avg_trade_return_pct 22%), so it is neither cliff-fragile nor fee-dead

Overfit: walk-forward IS Sharpe 1.32 collapses to OOS 0.006 and holdout Sharpe is NEGATIVE (-0.016) on 48 trades — the edge does not generalize to unseen data

Fails the multiple-testing gates decisively: deflated_sharpe 0.127 (<0.95), PBO 0.63 (>0.5), is_significant=false, does not survive programme FDR (keeps 3/223)

Optimization degraded the result (holdout negative, max_drawdown 37.5%, DSR unchanged) rather than finding a robust region

Headline is a compounding/leverage + unrealized artifact: avg_trade_return_pct 22% (L36) and end_unrealized_pct 165.5 mean the +855% is not a realized, risk-adjusted edge

Class prior: the diversified-TSMOM long-short majors-basket family is 0-for-many (8-major, 6-perp vol-targeted, 7-major macro, 3-major confluence, residual momentum all abandoned overfit/fee_edge)

Benjamini-Hochberg at q=0.10 over 223 programme candidates keeps 3. A candidate that does not survive here is not distinguishable from the programme's own noise, however good its individual statistics look.