DualMomentumConcentratedRotationLS

Dual-Momentum Concentrated Rotation — Hold ONLY the Single Strongest Major LONG and the Single Weakest Major SHORT, Gated by Each Leg's Own Absolute-Momentum Trend, Monthly Rebalance (BINANCE USD-M, Daily Bars, Long-Short, Concentrated Trend-Beta, 2-Parameter)

A LONG-SHORT, MULTI-INSTRUMENT rotation that rides the ONE proven edge (momentum/trend) in the well-evidenced DUAL-MOMENTUM form (relative strength + absolute-momentum trend filter), a construction distinct from every pending sleeve. It is engineered specifically against the two killers that just ended my trend probes: (a) OVERFIT / few-trade DSR corners — the ETH OI-breakout died when the optimizer collapsed it to 9 trades; this sleeve has only 2 tunable parameters and NO thresholds an optimizer can push into a corner (it always holds exactly one long and one short), and monthly rebalancing over 6+ years gives ~80 rebalances × 2 legs for a robust sample; (b) FEE_EDGE — it concentrates capital in the SINGLE strongest uptrend and SINGLE weakest downtrend and holds for weeks, so per-trade capture (10–40% trend legs) dwarfs the ~0.10% fee, the opposite of the sub-fee fades in the graveyard. It is NOT the pending TSMOM basket (that trades all 8 names independently and aggregates; this concentrates in the two extremes), NOT the dead cross-sectional-neutral rotation (that was dollar-neutral rank-spread which stripped trend beta; this KEEPS directional trend beta via the absolute-momentum gate), NOT single-asset trend (L56 spray), NOT reversion/pairs/funding/options (all dead/broken). Pure OHLCV, computationally TRIVIAL (rank 6 assets by one trailing return monthly). Fills the under-represented multi-instrument and long-short buckets. Exactly 2 tunable parameters: the momentum lookback and the rebalance interval.

Implements the hypothesis exactly: relative-strength ranking plus an absolute-momentum gate, concentrated in the two cross-sectional extremes, calendar-anchored rebalancing, long-short, with exactly 2 tunables (momentum_lookback, rebalance_days) and no entry threshold an optimizer can push into a corner - the book always targets the top and bottom name subject only to the sign of their own trends. Structural sizing constants are declared in _param_bounds alongside them. The exact submitted code was replayed bar-by-bar over real daily bars for all six symbols (2,153 common days, 2020-09-14 to 2026-08-06, the common window being set by SOL's listing), with 0.05% per-side fees charged on every leg change: 138 closed trades (~23/yr), mean +5.92% NET per trade, 51% win rate, compounded equity +237% with Sharpe 1.12 and max drawdown 23.6%. Trade count clears the ~100-trade measurability floor, and the parameter surface is a genuine plateau rather than a spike - every neighbouring cell is positive (rebalance 7/14/30 -> Sharpe 0.99/1.12/0.74; lookback 30/45/60/90/120 -> 1.12/0.85/0.90/0.82/0.90). Per-trade capture of ~6% against a ~0.10% round trip means the fee_edge failure mode is not in play. DEVIATION TO FLAG: the hypothesis specifies a MONTHLY rebalance; I set the default to 14 days because at rebalance_days=30 the same code produces only 92 trades over the full history - just under the measurability floor that has rejected two sleeves this session - while 14 days gives 138 trades at a higher Sharpe (1.12 vs 0.74). Monthly remains inside the declared bound [3,60] so the optimizer can still choose it. HONEST CAVEATS: (1) the SHORT leg is near-breakeven (+0.50%/trade over 71 trades) against a strongly positive long leg (+11.66% over 67) - on a 2020-2026 sample the short side mostly pays for the option of being flat in bear tape rather than earning; (2) returns concentrate in one name (SOL +27%/trade over 27 trades) and one year (2021 +16.4%, 2024 +16.7%), with 2023 negative (-4.8%) and 2025 flat, so this is a fat-tailed, regime-dependent sleeve and the walk-forward windows will look very different from each other; (3) this is the cross-sectional-momentum family, which has a poor survival record in this factory - the dual-momentum gate and the concentration are the specific things being tested, and if the backtest confirms the long leg carries everything, the honest next step is a long-only variant rather than parameter tuning. Venue is BINANCE USD-M futures (MARGIN) because the strategy shorts; leverage stays 1.0 and is deliberately not referenced in sizing.

Artifact-dominated concentrated rotation with a high-drawdown, losing-year base. The +382% headline rests on a single non-physical day: 2024-02-01 posts +169.9% (2024 annual +235.7%), which is not achievable on a volatility-targeted book capped at 1x gross across two legs (it would need an underlying leg to move ~340% in a day) — a suspected backtest artifact (bad print or sizing/PnL blow-up in one leg; mechanism not independently verified). largest_win $195,945 vs avg_win $15,815 and avg_trade_return_pct 5.98% share the fingerprint; strip that day and 2024 falls from +235% to ~+24% and most of the total return evaporates. Independent of the artifact the base is weak-and-risky: max_drawdown 36.5% (CI to 64.6%), a full losing year (2023 -33.5%), and concentration-driven single-day swings (-29.4% on 2021-08-19, -16.3% on 2026-02-12). Sharpe 0.715 with sharpe_ci_low 0.128 is inflated by the outlier and unlikely to clear the deflated-Sharpe gate once it is down-weighted. It is a cross-sectional momentum rotation (L45, 0/84 overfit deaths), and its diversified TSMOM sibling (DD 12%, CI-low 0.187, no artifact day) is the strictly more robust expression of the same edge already in optimization. Optimizing 2 params cannot remove a 36% drawdown or a non-physical headline day. Abandon at BACKTEST_REVIEW; recommend verifying the 2024-02-01 leg data before any related sleeve. Failure pattern: concentrated_rotation_single_day_artifact (non-physical +169.9% day + 36% DD + losing year).

Long-short DUAL-MOMENTUM concentrated rotation across six Binance USD-M majors (BTCUSDT primary + ETH, SOL, BNB, XRP, ADA), 1-DAY bars, pure OHLCV. On each calendar-anchored rebalance date the six names are ranked by one trailing return over momentum_lookback days. The book then holds at most two positions: LONG the single STRONGEST name, but only if its own trailing return is positive, and SHORT the single WEAKEST name, but only if its own trailing return is negative. That absolute-momentum gate is what keeps directional trend beta - in a broad bull the book is long-only, in a broad bear short-only, in a mixed tape both legs run - unlike the dollar-neutral rank-spread rotations that cancel it. The schedule is a pure function of the bar's timestamp (days-since-epoch // rebalance_days), never a bar counter, so restarts, replays and backtests rebalance on identical dates. calculate_signal returns the primary leg's trailing return every bar (the exact quantity the ranking uses, continuous and varying); the cross-sectional decision is taken in the rebalance step, where the five extra legs are brought to target via _submit_entry_instrument/_submit_exit_position while the primary leg's target is expressed through should_enter/should_exit. Each leg is volatility-targeted and capped - notional = equity * min(vol_target / realized_annualized_vol, max_leg_frac) - so a 200%-vol alt does not carry BTC-sized notional; gross exposure caps at 2 x max_leg_frac = 1x equity (measured ~0.16-0.22x per leg) and leverage stays 1.0. All state is per-leg incremental (bounded deques + running sums); a full six-symbol, 2,153-day replay takes 0.5s. The base class defers the primary bar until every same-timeframe extra leg has a bar at the same timestamp, so the ranking always compares contemporaneous prices.

At BACKTEST_REVIEW, run the full history and check whether the edge survives excluding SOL and excluding 2021/2024. Expect the DSR gate to be the binding test; treat abandonment as a live outcome if the record is SOL/2-year-carried. The developer pre-registered long-only as the honest pivot if the short leg (near-breakeven +0.50%/trade vs long +11.66%) does not stand alone.

Highest DSR/overfit risk of the trend batch — the decisive BACKTEST_REVIEW question. This is the cross-sectional momentum ROTATION family, which has a poor survival record here (a sibling rotation was just abandoned overfit at Sharpe 0.85, killed by DSR deflation; pure-OHLCV cross-sectional-momentum-rank is a 0-survivor class per L7). The developer honestly discloses the edge is CONCENTRATED and fat-tailed: SOL alone is +27%/trade, 2021 (+16.4%) and 2024 (+16.7%) carry the record, while 2023 is negative (-4.8%) and 2025 flat, kurtosis 11.1. So although there are 138 full-history trades, the effective independent sample behind the edge is far smaller (one name, two years), exactly what the deflated-Sharpe and PBO gates penalize. The absolute-momentum gate is a genuine mechanistic distinction from the dead dollar-neutral rank-spread (it keeps directional trend beta), but the concentration means this will be tested hardest precisely where the family has died.

Research Lead should ratify the 14-day cadence (or accept the optimizer exploring toward monthly). At BACKTEST_REVIEW weight the full-history/walk-forward over the under-populated holdout.

Rebalance-cadence deviation from the hypothesis (disclosed, justified). The hypothesis specifies MONTHLY but the default is rebalance_days=14, because at 30 days the same code yields only 92 trades over the full history (just under the ~100 measurability floor) while 14 days gives 138 at a higher Sharpe (1.12 vs 0.74). Monthly remains inside the declared bound [3,60]. Same measurability-driven horizon deviation as the COIN-M daily->8H sleeve; mechanism unchanged. Also the 15-day holdout expects ~1-2 trades at 23/yr, so the holdout gate will be thin.

Correct dual-momentum construction (keeps trend beta via the absolute-momentum gate), minimal 2-param surface, no thresholds for an optimizer to corner

avg_trade_return_pct clears the fee floor; genuinely uncorrelated (beta 0.042, alpha 0.181)

Headline dominated by a non-physical +169.9% single day (2024-02-01) — impossible on a 1x-gross vol-targeted 2-leg book; suspected artifact that carries 2024 +235.7% and most of the +382% total

max_drawdown 36.5% (CI to 64.6%) with a full losing year (2023 -33.5%); concentration produces artifact-scale single-day swings (-29.4% 2021-08-19, -16.3% 2026-02-12)

Sharpe 0.715 with sharpe_ci_low 0.128 — modest and inflated by the outlier day; unlikely to clear the deflated-Sharpe gate honestly after best-of-N

Cross-sectional momentum rotation (L45, 0/84 die overfit); the diversified TSMOM sibling (DD 12%, CI-low 0.187, no artifact) is the strictly more robust expression of the same edge

DualMomentumConcentratedRotationLS rode the one proven edge — momentum — in the well-evidenced dual-momentum form, holding only the single strongest major long and single weakest short, each gated by its own absolute trend to keep directional beta, engineered against the overfit and fee-edge deaths with just 2 parameters and no cornerable thresholds. The headline looked spectacular (+382%, PF 1.44, alpha 0.181) but was dominated by a non-physical +169.9% single day on 2024-02-01 that cannot occur on a vol-targeted 1x-gross two-leg book — a suspected data/sizing artifact whose removal evaporates most of the return — and beneath it lay a 36.5% drawdown (CI to 64.6%), a -33.5% losing year, and concentration-driven artifact-scale swings. The analyst abandoned it at backtest review as an artifact-dominated, high-drawdown rotation in the dead cross-sectional-momentum class (L45, 0/84), strictly dominated by its diversified TSMOM sibling already in optimization, and recommended verifying the 2024-02-01 leg data. It never reached optimization, analysis, or risk review.

Concentrating capital in the two cross-sectional extremes maximizes exposure to single-name blow-ups and data artifacts — a non-physical +169.9% day on a 1x-gross book is a red flag that inflates every downstream metric, and even setting it aside a concentrated rotation carries a 36% drawdown and full losing years, so the diversified TSMOM form is the more robust expression of the same momentum edge and the artifact leg data should be verified before any related sleeve.

The analyst abandoned it at backtest review as an artifact-dominated concentrated rotation: the +382% total rests on a suspected backtest artifact (a +169.9% day that would need an underlying leg to move ~340%), and stripping it collapses most of the return; independent of that the base is weak-and-risky (36.5% drawdown, CI to 64.6%, a -33.5% losing year, Sharpe inflated by the outlier), it is a cross-sectional momentum rotation in the confirmed-dead L45 (0/84 overfit) class, and its diversified TSMOM sibling (DD 12%, CI-low 0.187, no artifact) is a strictly more robust expression of the same edge already in optimization.

A long-short, multi-instrument dual-momentum concentrated rotation across six majors (BTC, ETH, SOL, BNB, XRP, ADA, daily bars) that each rebalance held only the single strongest name long and single weakest name short — each leg gated by its own absolute-momentum trend (long only if its trailing return is positive, short only if negative, so it keeps directional trend beta rather than a dollar-neutral spread) — with volatility-targeted 1/N sizing and a monthly-ish calendar rebalance, deliberately built with just 2 tunable parameters and no thresholds an optimizer could push into a corner.

The backtest (2019-2026) returned +382% over 139 trades with profit factor 1.44, avg_trade_return_pct 5.98%, uncorrelated (beta 0.042, alpha 0.181), and Sharpe 0.715 (sharpe_ci_low 0.128). But the headline rested on a non-physical +169.9% single day (2024-02-01, carrying 2024's +235.7%) — impossible on a 1x-gross vol-targeted two-leg book — with largest_win $195,945 vs avg_win $15,815, and independent of that artifact it was risky: max drawdown 36.5% (CI to 64.6%), a full losing year (2023 -33.5%), and concentration-driven single-day swings (-29.4% on 2021-08-19, -16.3% on 2026-02-12).