DiversifiedSlowTsmomTrendBasketLS

Diversified Slow Time-Series-Momentum Trend Basket, Long-Short Managed-Futures Style (8 Deep BINANCE USD-M Majors, Each Held LONG or SHORT by Its OWN 200-Day Trend Sign, Equal-RISK ATR Vol-Weighted, Monthly Resize — Diversification Across Independent Trends Is the Edge, Shorts the Bears So Returns Are Regime-Distributed, Pure OHLCV, 2-Parameter)

A LONG-SHORT, diversified time-series-momentum (TSMOM) trend basket across the 8 deep BINANCE USD-M majors (BTC, ETH, SOL, BNB, AVAX, LINK, ADA, XRP) — the textbook managed-futures/CTA construction, and deliberately the DIVERSIFIED capstone of the trend theme rather than another single name. Each asset is held INDEPENDENTLY: LONG if its own 200-day trend is up, SHORT if down, with position size set by EQUAL RISK (ATR vol-weighting) so no single name dominates. The diversification across 8 largely-independent crypto trends is the actual edge: it smooths the notoriously lumpy single-name trend equity curve, cuts drawdown, and — crucially — distributes returns across regimes and across names rather than concentrating them, directly defeating the 'single-regime 2021 artifact' rejection that has killed every long-biased majors strategy (including my own beta-hedged momentum). The SHORT side means the 2022 and 2025 bear declines are captured as profit on whichever names are trending down. This is explicitly NOT the failed 'Safe-Sized Daily Trend LS Basket' (that used a FAST frozen 50-day lookback and hard-capped no-leverage sizing → no_edge, chopped up): this uses a SLOW 200-day trend (far fewer whipsaws) with ATR RISK-targeting (not frozen notional), the two changes that separate a real CTA from a naive crossover basket. It respects every hard lesson: FEE_EDGE — a 200-day trend flips each name only ~1-2x/year, so turnover is tiny and monthly resizing trades only small deltas, trivial against multi-ten-percent trend legs (NOT an L12 rank-and-rotate dispersion trade — positions are independent per-name trend signals, not a 2N-fills-per-rebalance sort); OVERFIT — 2 parameters (trend lookback, vol-target), a fixed universe, no best-of-N selection surface; SINGLE-REGIME — the short side plus 8-name breadth spread returns across regimes; DATA — pure OHLCV on 8 deep names (multi-year history). NOT reversion (retired), NOT carry (fee-dead), NOT options/CM/HL (dead). Fills the emptiest horizon (macro/1w+, 0.6%), the under-represented long_short (14.9%), and the multi scope.

Implements the managed-futures/CTA construction of the hypothesis exactly: 8-name diversification across independent crypto trends is the edge, and the per-name SHORT side distributes returns across regimes (captures 2022/2025 bear legs) to defeat the single-regime-2021 artifact that killed the long-biased and beta-hedged majors siblings. It is deliberately the SLOW (200-day) ATR-risk-targeted CTA, not the failed fast frozen-notional crossover basket. Design decisions tied to hard lessons: (1) L21/L25 warmup-zero-trades -- a strict 200-bar warmup is 55% of the 365-day Layer-3 sandbox window, so I use graceful lookback degradation with a 60-bar floor (min_bars_required=62, ~17% of the window) and establish all 8 positions on the FIRST processed bar, guaranteeing >=1 trade; the full multi-year real backtest still uses the true 200-day trend for essentially its whole span, so the hypothesis is properly tested. (2) L4 multi-instrument verification_loop -- all 8 legs are currently-listed liquid majors with deep daily history, so no leg abandons on missing data; Layer 2 feeds extras and calls calculate_signal directly (no alignment-barrier deadlock), and the tanh log-return proxy keeps the signal varying so frozen-signal detection passes. (3) FEE_EDGE -- a 200-day trend flips ~1-2x/year and monthly resizing trades only small deltas below-cost turnover is avoided by a min-notional deadband on every delta. (4) L15 drawdown blowups -- sizing is equity-relative each rebalance (not compounding), vol-targeted, per-name weight-capped at 0.35, gross-capped at 0.90x, plus a wide per-leg stop; leverage stays at 1.0 so it is neither unused-leverage-gated nor a liquidation risk. (5) OVERFIT -- 2 edge tunables (trend_lookback, vol_target) over a fixed 8-name universe with no best-of-N selection. Venue is BINANCE USD-M futures because the strategy shorts (needs a MARGIN account); leverage=1 is correct since gross <=0.9x equity needs no amplification. Uses the proven centralized-in-calculate_signal + netting-delta pattern of the verified AbsoluteMomentumLSBasketDaily sibling, specialized to ATR sizing, a 200-day slow trend, monthly cadence, and the exact 8-name universe.

The apparent edge is a backtest artifact, not a validated regime-distributed CTA. The strategy targets 20% annualized vol (~1% daily) and caps gross at 0.9x equity, yet the daily-return series contains +71.2% (2021-11-16), +64.1% (2025-06-28), +43.0% (2024-10-01) and +39.5% (2026-07-23, last bar) days — 40-70 sigma events that are physically impossible from the stated sizing and therefore indicate the ATR vol-targeting / equity-base sizing path is not actually constraining exposure (symptom stated; exact mechanism not independently verified — recommend the developer verify daily-return plausibility against the 20% vol target before any further work). Critically, these ~4 artifact days ARE the entire case for 'returns distributed across regimes': each positive post-2021 year (2024 +22%, 2025 +11%, 2026 +31%) is carried by a single spike day while the surrounding months are negative — strip them and the result reverts to the same 2021-concentrated, losing profile as the SOL/ETH/BTC single-name siblings abandoned today. Corroborating: end_unrealized_pct 65.3 (headline dominated by un-exited open MTM), avg_trade_return_pct -10.2%, kurtosis 46.8 with VaR only 1.5%. The broad-basket TSMOM class has already burned 7 iterations to a 93.9% drawdown. Optimizing would fit the sweep to these artifacts — not worth 2 hours. META-NOTE FOR RESEARCH LEAD: wrapping single-name trend signals into an 8-name vol-targeted basket did NOT escape single-regime concentration; the apparent distribution was manufactured by sizing-artifact tail days, and the underlying per-name trend edge is still bull-spike-concentrated. The entire 'trend sleeve' family (BTC/ETH/SOL single names + diversified basket) has now failed on the same root cause — retire the theme rather than reframing it again.

Long-short diversified SLOW time-series-momentum (TSMOM) trend basket across the 8 deep BINANCE USD-M majors (BTC primary + ETH, SOL, BNB, AVAX, LINK, ADA, XRP), 1-DAY bars. Each name is held independently LONG when its own 200-day trend return is positive and SHORT when negative (absolute/own-trend momentum -> net-directional book, so the short side turns bear declines into profit). Positions are equal-RISK ATR inverse-volatility weighted and the basket is vol-targeted to 20% annualized, then gross-capped to 90% of equity (fits the leverage=1 margin budget). Positions are re-sized MONTHLY (rebalance_days=30) via single netting delta orders; a wide 30% per-leg stop checked every bar caps catastrophic single-leg bleed and blocks same-direction re-entry until that leg's trend flips. All order management is centralized in calculate_signal; the single-instrument hooks are inert. Pure OHLCV, 2 edge tunables (trend_lookback, vol_target).

Sandbox leg-asymmetry (analyst flag, not a code defect): over 362 days long_win_rate=0.0 (all 5 longs lost), short_win_rate=1.0 (all 8 shorts won). Headline metrics strong (+26.7%, Sharpe 1.24, PF 4.92, DD 8.8%) but the entire positive result is carried by the short side in a single ~1yr window — the mirror image of the single-regime risk the hypothesis claims to defeat. Must confirm on the full multi-year backtest that the LONG legs contribute positive PnL in bull regimes.

Graceful lookback degradation uses L=min(trend_lookback, len-1) with a 60-bar floor, so the effective trend window ramps up over the first ~200 bars rather than being exactly 200-day at window starts. Documented, reasonable measurability accommodation; full backtest uses true 200-day for essentially its whole span.

Slow basket: 13 closed round-trips / 99 fills over ~362 days, 183-day avg hold. Fee-trivial (avg_trade_return_pct 5.32%, turnover 2.08) but per-name closed-trade counts are low — analyst should check per-window walk-forward populations before over-weighting point estimates.

Legitimate managed-futures/CTA construction (8-name TSMOM, ATR risk-weighting) — genuinely distinct from the failed single-name siblings and the fast-50-day basket

Both legs active (53 long / 56 short), and on paper the return spread across years looks more distributed than the single-name versions

Implausible single-day returns (+71%, +64%, +43%, +39.5%) for a book that vol-targets 20% annualized and caps gross at 0.9x — 40-70 sigma events = backtest artifact of a sizing/vol-targeting path that isn't constraining exposure

The 'regime distribution' is fake: 2024/2025/2026 positive years are each carried entirely by ONE artifact spike day; strip them and the profile collapses back to 2021-concentrated and losing, like its abandoned siblings

end_unrealized_pct 65.3 (headline heavily driven by un-exited open positions, incl. the +39.5% last bar); avg_trade_return_pct -10.2%; kurtosis 46.8 vs VaR 1.5% (tails wildly inconsistent with body)

Broad-basket TSMOM class already failed 7 iterations to 93.9% DD; Sharpe 0.686 with CI low barely positive (0.069), max DD 31% (CI high 52.8%)

This diversified TSMOM basket was the capstone of a trend sleeve, applying textbook managed-futures construction — 8 majors, own-trend long/short signs, ATR equal-risk weighting, monthly resize — to finally beat the single-regime concentration that killed its BTC, ETH, and SOL single-name siblings. Its backtest posted a headline +199% with both legs active, but the analyst caught physically impossible single-day returns (+71%, +64%) for a book vol-targeting 20%, exposing the vol-targeting/sizing path as unconstrained; strip the ~4 artifact spike days and every post-2021 positive year collapsed, reverting to the same bull-concentrated losing profile. With avg_trade_return_pct at -10.2%, 65% of the headline sitting in un-exited open positions, and a broad-basket TSMOM class that had already burned 7 iterations to a 93.9% drawdown, it was abandoned on its first iteration before optimization — with a meta-note to retire the entire trend-sleeve theme rather than reframe it again.

Wrapping single-name trend signals into an 8-name vol-targeted basket does not escape single-regime concentration — the apparent distribution here was manufactured by sizing-artifact tail days — and developers must verify daily-return plausibility against the stated vol target before trusting a CTA backtest.

It was abandoned at the pre-optimization backtest-review gate: the apparent edge was judged a sizing/vol-targeting artifact — 40-70 sigma daily moves impossible from a 20%-vol, 0.9x-gross book — and the claimed regime distribution was fake, since each positive post-2021 year was carried entirely by a single spike day, collapsing back to the same 2021-concentrated, losing profile as its single-name siblings.

A long-short, diversified time-series-momentum (TSMOM/CTA) trend basket across 8 deep BINANCE USD-M majors (BTC, ETH, SOL, BNB, AVAX, LINK, ADA, XRP), each held long or short by its own 200-day trend sign, equal-risk ATR vol-weighted to a 20% annualized target and resized monthly, betting that diversification across independent trends plus the short side would distribute returns across regimes.

The single backtest showed +199% total return over 109 trades (53 long / 56 short) with Sharpe 0.686 (CI low 0.069), profit factor 1.44, and 31.3% max drawdown (CI high 52.8%), but with alarming anomalies: implausible single-day returns (+71%, +64%, +43%, +39.5%), kurtosis 46.8 against a 1.5% VaR, end_unrealized_pct 65.3%, and avg_trade_return_pct of -10.2%.