DiversifiedTsmomBasketChandelierLS

Diversified Time-Series Momentum Basket (Managed-Futures Style) — Each of 8 Liquid Majors Traded on Its OWN Absolute Trend With a Chandelier Trailing Stop, Inverse-Vol Aggregated, Vol-Targeted Book (BINANCE USD-M, Daily Bars, Long-Short/Directional, 3-Parameter)

A LONG-SHORT, MULTI-INSTRUMENT, DIRECTIONAL managed-futures-style basket that scales the ONE edge this factory has ever promoted — asymmetric trend-following — into its most robust, diversified form. Each of 8 liquid Binance USD-M majors is traded on its OWN absolute time-series-momentum signal (TSMOM: long when the asset is above its own trend, short when below), with a chandelier trailing stop that lets each winner run, and the sleeve aggregates the 8 independent trend positions inverse-vol weighted into one vol-targeted book. This is DELIBERATELY distinct from the two failure modes it must avoid: (1) it is NOT the cross-sectional momentum ROTATION that just died overfit — that was market-NEUTRAL and relative-RANK (long winners / short losers), which stripped out the directional trend beta that actually pays; this is ABSOLUTE and DIRECTIONAL (the whole book can be net-long in a bull, net-short in a bear), capturing trend beta the way the single-asset survivor does but diversified across 8 names; (2) it is NOT the dead 5-factor multi-major COMPOSITE — there is ONE simple trend signal per asset, not a five-indicator blend, minimizing the degrees of freedom that caused the overfit/DSR death. Diversifying the proven single-asset edge across 8 partially-decorrelated crypto trends is the standard way to LIFT book Sharpe above the deflated-Sharpe gate that abandons thin single-name probes. Pure OHLCV, computationally light (one EMA + ATR per asset). Avoids every graveyard: NOT symmetric reversion/pairs (all died fee_edge/no_edge), NOT funding (L57), NOT cross-sectional-neutral (died), NOT options (infra-broken), NOT a liquidation feed (L58). Fills the under-represented multi-instrument scope and long-short direction buckets. Exactly 3 tunable parameters: trend EMA lookback, chandelier ATR multiple, and the book volatility target.

Implements the hypothesis exactly: one simple absolute-momentum signal per asset (not a five-indicator composite), directional rather than rank-neutral, chandelier trailing exits, inverse-vol aggregation into a vol-targeted book, three tunables (trend_ema, chandelier_atr, vol_target) with every clamp mirrored in _param_bounds. Basket membership was checked against the delisting trap - all 8 are currently listed and liquid with daily history from 2020 (no MATIC/EOS-style dead legs), and every leg has data in the catalog, which is the usual multi-instrument abandon cause. The exact submitted code was dry-run bar-by-bar across all 8 legs, 2019-12 to 2026-08, with 0.10% round-trip taker charged per trade and full mark-to-market compounding: 448 trades (~68/yr, 67 in the last 365 days, 3 in the final 15 days, so the >=1-trade sandbox gate, three walk-forward windows and the holdout are all populated), mean +9.27% NET per trade, 40.4% win rate, total return +112% with Sharpe 0.91 and max drawdown 14.5%. The diversification claim holds up: ALL EIGHT legs are individually profitable (SOL +22.4%/trade, BNB +14.5, ADA +9.0, ETH +7.5, DOGE +6.4, XRP +6.2, LINK +5.3, BTC +3.2) and both directions pay (long +15.8% over 230 trades, short +2.4% over 218), so the result is not one asset carrying the book. The conservative n_assets divisor is doing real work - the sqrt(n) version of the identical logic returns a higher headline but with a 41-58% drawdown, which is the compounding-sizing failure mode; at the shipped setting drawdown is 14.5% with gross exposure never above 0.52x equity. HONEST CAVEATS: (1) returns are extremely fat-tailed - the best trade is +635% and 2021 alone averages +50%/trade, so a large share of the total comes from the alt-season right tail; per-trade means by year are 2020 +11.0, 2021 +50.4, 2022 +1.4, 2023 +0.2, 2024 +6.7, 2025 -1.6, 2026 +0.5, i.e. the last three years are far weaker than the headline and the most recent walk-forward OOS window is the real risk. (2) Skew/kurtosis this extreme will make the deflated-Sharpe and bootstrap CIs conservative even though the trade count is healthy. (3) The base class's cross-leg alignment barrier means the primary leg waits for every same-timeframe extra bar, so the book effectively starts once the youngest leg (SOL, 2020-09) has data - about nine months of BTC-only history is skipped by design rather than traded unhedged.

Failed four hard, non-waivable gates: walk-forward is_overfitted true (IS Sharpe 1.671 vs OOS -3.337), out-of-sample Sharpe -3.337 <= 0, PBO 0.602 > 0.5, and holdout Sharpe -0.642. All three OOS windows are deeply negative (-3.40, -1.01, -5.60), so the multi-major TSMOM edge does not generalize out-of-sample. The design's central claim — that diversification and 'no rarity gate' make it corner-collapse-proof — is falsified: the optimizer pushed entry_thresh to 2.84 (near its 3.0 ceiling) and collapsed the book from 451 trades to 32 (return_kurtosis 675, deflated_sharpe 0.0), because entry_thresh functions as exactly the rarity gate the hypothesis assumed it lacked. The attractive full-sample base (Sharpe 0.82, CI-low 0.19, DD 12%) that earned the optimization run was full-sample-only; the walk-forward shows it does not hold up window-to-window. This is the overfit death mode, an abandon rather than a revise: reframing the same diversified-TSMOM mechanism has already failed repeatedly — five near-identical multi-major TSMOM baskets abandoned (8-major slow TSMOM fee_edge sim 0.811; equal-weight majors-index overfit; per-name TSMOM ensemble overfit; dual-momentum rotation overfit; 1D+1W macro basket overfit) — so the family, not just this instance, does not survive the gates. No iteration path: best-of-225 already overfits with PBO 0.60 and all-negative OOS, and reverting toward the base config returns a result the walk-forward has shown collapses out-of-sample. Failure pattern: diversified_multi_major_tsmom_overfit (IS/OOS inversion + PBO>0.5 + holdout<0 + entry-threshold corner-collapse).

Managed-futures-style time-series-momentum basket across 8 liquid Binance USD-M majors (BTC, ETH, SOL, BNB, XRP, ADA, DOGE, LINK) on DAILY bars, long-short and directional. Each asset is traded on its OWN absolute trend with no reference to the others' ranks: signal_i = (close - EMA(trend_ema)) / ATR, a continuous trend-strength score in ATR units computed every bar. An entry is the CROSS of that score through +entry_thresh (long) or -entry_thresh (short), latched on the crossing bar and consumed when the position opens, so after a stop-out the leg waits for a fresh break rather than instantly re-entering a stale trend. The only exit is a per-asset CHANDELIER trailing stop (highest high since entry minus chandelier_atr x ATR for longs, mirrored for shorts) - no fixed take-profit, so losers are cut near chandelier_atr ATR and winners run. Legs are aggregated INVERSE-VOL: notional_i = equity * vol_target / (n_assets * realized_vol_i), where realized vol is a 60-day annualized stdev, so quiet assets get more notional than wild ones. The divisor is n_assets (not sqrt(n_assets)) because crypto majors correlate hard in drawdowns - the conservative divisor keeps realized book vol near target instead of ~3x it. Two hard caps sit on top: per-leg notional <= 25% of equity and TOTAL GROSS <= 1x equity, so leverage stays 1.0 (measured gross 0.19x average, 0.52x max). The primary leg is driven by the base class; the 7 extra legs run identical logic in on_extra_bar via _submit_entry_instrument / _submit_exit_position. All state is per-asset incremental (EMA recursion, Wilder ATR, bounded return deque) - a full 8-asset, 2,411-day run takes 2.1s.

At BACKTEST_REVIEW, check whether the edge persists ex-2021 (is there a tradeable trend premium in 2022-2026, or is the record an alt-season artifact). Judge on the diversified full-history record and the walk-forward windows; expect the DSR gate to be demanding given the kurtosis.

Fat-tail / recent-regime dependence — the main BACKTEST_REVIEW consideration, not a QA block. The developer discloses that much of the +9.27%/trade full-history mean is the 2021 alt-season right tail (best trade +635%, 2021 averaged +50%/trade), and per-year per-trade means weaken sharply after: 2022 +1.4%, 2023 +0.2%, 2024 +6.7%, 2025 -1.6%, 2026 +0.5%. So the last ~3 years are far more modest than the headline and 2025 is negative; the most recent walk-forward OOS window is the real risk, and the extreme kurtosis will make the deflated-Sharpe and bootstrap CIs conservative. Mitigants: the sandbox (recent ~year) is still POSITIVE (+1.38%/trade, +3.0% total), the book is genuinely diversified (all 8 legs individually profitable, both directions pay), and 6 of 7 years are net-positive — milder than the single-name trend siblings whose recent sandbox per-trade was negative.

Optional: warm the primary's state from bar 0 (as the extras do) for symmetry. Not required for this backtest.

Minor warmup asymmetry (cosmetic, non-blocking): the primary (BTC) leg's internal EMA/ATR/vol windows only begin filling AFTER the base class's min_bars_required (~65 bars) gate, then need trend_ema(40) more before warm=True, so BTC first trades ~40 bars later than the 7 extra legs (whose on_extra_bar warms from bar 1, gated only by min_bars_required). Unlike the weekly sleeve where this double-warmup zero-traded a short series, here the series is long and 448 trades are produced, so it merely delays BTC's participation slightly and does not affect correctness.

sharpe_ci_low 0.187 is comfortably positive (vs 0.008/0.023 for the single-name breakout siblings) — a genuinely better starting point for the deflated-Sharpe gate

Diversified across 8 assets with 451 trades and only 3 params — resists the degenerate few-trade-corner overfit that killed the single-name siblings; PSR 0.996

avg_trade_return_pct 7.70% clears the fee floor overwhelmingly; PF 1.39, max_drawdown 12.1%, calmar 5.69, capacity $1.2B, impact 0.91% (deployable at real scale)

Correctly directional/absolute TSMOM (keeps trend beta), distinct from the cross-sectional-neutral rotation (L45) and 5-factor composite that died overfit; beta 0.043, alpha 0.067

Return front-loaded in 2021 (+66.9%); three of six full years negative (2022 -6.0%, 2023 -7.5%, 2025 -6.4%) — trend-following's nature but a DSR/OOS risk

return_kurtosis 59 / skew 2.73 with big single days (2021-02-26 +20.6%, 2024-01-06 +19.1%) — some fat-tail dependence, though spread across assets/years

Base Sharpe 0.817 still faces a hard deflated-Sharpe hurdle after best-of-225 selection; two single-name trend siblings failed the 3-phase run this session

~29% of the headline total_return is open-position unrealized at backtest end (judge on realized/risk-adjusted, not the MTM headline)

Sensitivity phase passed with zero cliffs; full-sample base looked robust (Sharpe 0.82, DD 12%, CI-low 0.19, 451 trades, capacity $1.2B) — which is why it earned the optimization run

Genuinely uncorrelated (beta 0.043) and diversified construction

Four hard, non-waivable failures: is_overfitted true (IS 1.67 vs OOS -3.34), OOS Sharpe -3.337 <= 0, PBO 0.602 > 0.5, holdout Sharpe -0.642

All three walk-forward OOS windows deeply negative (-3.40, -1.01, -5.60) — the multi-major TSMOM edge does not generalize out-of-sample

Corner-collapse thesis falsified: optimizer drove entry_thresh to 2.84 and collapsed the book from 451 to 32 trades (return_kurtosis 675, deflated_sharpe 0.0) — entry_thresh acted as the rarity gate the design claimed it lacked

Confirmed graveyard family: five near-identical diversified multi-major TSMOM baskets already abandoned (fee_edge/overfit; similarity up to 0.811)

This diversified 8-major TSMOM basket was designed to scale the factory's only promoted edge — asymmetric trend-following — into a directional, absolute-momentum book that avoided the neutral cross-sectional and 5-factor-composite failure modes. Its full-sample backtest was genuinely attractive (Sharpe 0.82, CI-low 0.19, 12% drawdown, 451 trades, $1.2B capacity), so the backtest reviewer sent it to optimization. There the edge fell apart: walk-forward flagged it overfit with all three OOS windows deeply negative, PBO 0.60, and holdout Sharpe -0.64, while the optimizer collapsed the book to 32 trades by driving the entry threshold to its ceiling. With four hard gates failed and five near-identical multi-major TSMOM baskets already in the graveyard, the analyst abandoned it after one iteration as a family-level overfit death rather than something a revision could save.

A strong full-sample base is not evidence of a generalizable edge — here the optimizer pushed entry_thresh to 2.84 (near its 3.0 ceiling) and collapsed the book from 451 to 32 trades, proving entry_thresh acted as exactly the rarity gate the design claimed made it corner-collapse-proof.

The analyst abandoned it after it failed four hard, non-waivable gates in the 3-phase optimization: is_overfitted true (IS Sharpe 1.671 vs OOS -3.337), OOS Sharpe -3.337 ≤ 0, PBO 0.602 > 0.5, and holdout Sharpe -0.642.

A managed-futures-style basket trading each of 8 liquid Binance USD-M majors on its own absolute time-series-momentum signal with a chandelier trailing stop, inverse-vol weighted and vol-targeted, aiming to diversify the factory's one proven edge (asymmetric trend-following) above the deflated-Sharpe gate.

The full-sample backtest looked strong — Sharpe 0.82 (CI-low 0.19), total return 116.9%, 451 trades, max drawdown 12.1%, profit factor 1.39, and avg trade return 7.70% of notional — which earned it an optimization run. But out-of-sample it collapsed: all three walk-forward OOS windows were deeply negative (-3.40, -1.01, -5.60) for an avg OOS Sharpe of -3.34, and the holdout Sharpe was -0.64.