Strategies

Multi-Major Tail-Regime Convex Directional Basket — Long-Short, Apply ONE Validated Large-Magnitude-Regime Convex Rule Independently Across 6 Majors, Flat Per-Name Otherwise, Shared Parameters (Daily Bars, 3-Parameter)

Outcome: Abandoned

MultiMajorTailRegimeConvexDirectionalBasketLS

Outcome Summary

MultiMajorTailRegimeConvexDirectionalBasketLS diagnosed the BTC tail-regime convex edge's sole flaw as cadence and prescribed diversification — six asynchronous shared-parameter sleeves — to reach a validatable ~60-65 engagements/yr while preserving each sleeve's convex, above-fee character. The cadence fix delivered a strong initial backtest (+418%, Sharpe 1.07, PF 1.85, 408 trades, beta 0.077), earning a full optimization. But the diversification premise failed where it mattered: majors correlate in the big regime legs, so the return stayed tail-dominated (kurtosis 96.8 rising to 190), and optimization collapsed on every robustness gate — DSR 0.298, PBO 0.655, a negative holdout Sharpe, with the optimizer even reverting to a ~20/yr cadence. The analyst abandoned it at the ANALYZING stage as convex_tail_regime_overfit, the whole family now exhausted; it reached optimization and analysis but not risk review or promotion.

Strategy report

Hypothesis

A LONG-SHORT, MULTI-INSTRUMENT, pure-price CONVEX directional basket that applies the factory's one genuinely-validated-but-uncadenced edge — the BTC Tail-Regime Convex Directional (abandoned ONLY because it fired ~11x/yr: the reviewer recorded 'genuinely novel and decorrelated, beta 0.009, real per-trade edge PF 2.03, avg_trade_return_pct 2.23% well above fees ... but fires far too rarely to be validated'). The sole flaw was CADENCE, not edge. The session also proved (via the negative-gross-edge ORB) that this convex edge lives specifically at the MULTI-WEEK regime horizon and CANNOT be sped up intraday — so the correct fix is DIVERSIFICATION, not acceleration: run the identical convex rule INDEPENDENTLY across 6 liquid majors (BTC, ETH, SOL, BNB, XRP, DOGE, all USD-M perps), each engaging only on ITS OWN large-magnitude directional regime and flat otherwise, summed into one book. Because each name's large regimes occur at different times, the aggregate produces ~55–70 engagements/yr (~350–450 over the sample) — enough to validate — while preserving the per-name convex, above-fee, decorrelated character. This is explicitly NOT a cross-sectional RANK rotation (L52): there is NO ranking, NO relative selection, NO always-on momentum — each sleeve is a standalone flat-most-of-the-time convex trigger, and the parameters are SHARED across all 6 names (not per-asset tuned), so the edge must be a GENERAL convex-regime effect, not asset-specific curve-fitting (this shared-param design also keeps DOF low against the deflated-Sharpe gate and the L52 overfit artifact). It harvests crypto's fat-tailed self-reinforcing regime legs (parabolic runs, cascade crashes) across the complex. NOT a mean-reversion fade (L53), NOT a squeeze breakout (L54), NOT a channel breakout (structurally-unmeasurable, just abandoned), NOT a NON-PRICE-feed gate (L46), NOT options/COIN-M (L50/L51), NOT a long-only spot basket (L48 — long-SHORT futures, flat most of the time). Fills multi-instrument scope (17.8%), long-short direction (13.5%), and adds a decorrelated convex sleeve. Risk profile: vol-scaled sizing PER SLEEVE so each engaged position targets equal risk; aggregate gross capped at ~60% of equity*leverage (with ≤6 simultaneous engagements rare since regimes are asynchronous); 2x leverage cap (reads self.config.leverage).

Implements the hypothesis exactly: the BTC Tail-Regime Convex Directional rule (validated edge, abandoned only for ~11 engagements/yr) replicated verbatim and INDEPENDENTLY across 6 majors with SHARED parameters, summed into one book - no ranking, no relative selection, no always-on exposure, each sleeve flat most of the time. Diversification, not acceleration, supplies the cadence. Offline dry-run on the real daily catalog (2019-12 to 2026-07) confirms it: 417 engagements (63/yr; BTC 75, ETH 81, SOL 61, BNB 67, XRP 67, DOGE 66) at +4.24% average NET-of-fee per-trade return (28x the 0.15% fee floor), avg hold ~6.5 days, positive in every calendar year (2025 ~flat). A parameter sweep over the declared _param_bounds (lookback 4-10, entry_z 1.0-2.0, chandelier 4-8) is positive at every grid point with no cliff - lookback=3 was the only weak corner and is excluded by the clamp. Portfolio sizing at risk_pct=0.04 yields ~11% of equity per engaged sleeve, mean gross ~35% and p95 ~85% of equity, so the leverage=2.0 caps (per-leg 35%, aggregate 120%) are genuinely consumed and read from self.config.leverage in position sizing. Futures venue is required because the book goes short. Layer-1 static analysis and all six Layer-2 synthetic scenarios pass locally.

failed deflated Sharpe: DSR=0.298 (<0.95), and optimized Sharpe 1.078 is BELOW expected-max-by-luck 1.223 over 225 trials — the selected result is worse than best-of-N noise would predict. Two HARD gates also fail: PBO=0.655 (>0.5, selection more likely than not overfit) and holdout Sharpe -0.031 / ratio -0.018 (edge is zero/negative on unseen data, collapsing from WF-OOS 1.73). This is the diversification variant of the convex tail-regime family, and it does not escape the family's fate: BTC tail-regime (fee_edge, 3 iters), DOGE (overfit), BTC vol-expansion (overfit), and the per-name TSMOM baskets (overfit) were all abandoned. The result is outlier-dependent (return_kurtosis 190; a handful of single days — 2021-02-20, 2024-12-02 +14.4%, 2024-12-04 +13.0% — carry it) and non-persistent (optimized 2025 -1.36%, 2026 -0.81%). Critically, the optimizer pushed entry_z to 1.98 and cut the book to 137 trades (~20/yr), re-creating the exact 'fires too rarely' cadence flaw the whole diversification premise was meant to solve, while still overfitting. No iteration path (overfit -> abandon, not iterate): the failure is best-of-N overfitting of an outlier-dependent trigger whose edge does not survive holdout, which no parameter change repairs. Failure pattern: convex_tail_regime_overfit.

Implementation

Long-short multi-instrument convex tail-regime basket on six Binance USD-M majors (BTC, ETH, SOL, BNB, XRP, DOGE), DAILY bars. Each name runs an INDEPENDENT sleeve of the identical rule: z = log-return over `lookback` days normalised by 60-day daily sigma * sqrt(lookback); engage LONG/SHORT only when |z| >= entry_z AND short-window (20d) volatility exceeds the 60d baseline (regime is expanding); otherwise stand aside flat. Exit on z flipping sign or a wide chandelier giveback of chandelier_atr * ATR(14) from the best close since engagement. Sizing is equal-risk per sleeve (equity * risk_pct / chandelier stop distance), capped per leg at equity * max_notional_frac * leverage and in aggregate at equity * max_gross * leverage, all off CURRENT equity so losses cannot compound into size. Three shared tunables only (entry_z, chandelier_atr, lookback), identical across all six names.

Verification Results

Backtest_review/analyst: judge on the pooled full-sample and walk-forward OOS windows (well-populated at ~63/yr) rather than the noisy single sandbox year; verify the long-leg engagements contribute positively across the sample and that the aggregate clears the deflated-Sharpe gate. This is the multi-asset reconception QA recommended for the single-name tail-regime, so it should be evaluated on its pooled convexity.

Significance not yet confirmed on the sandbox year, and a long/short asymmetry to watch — analyst/backtest_review calls, not code defects. The measurability fix works (417 full-sample engagements ~63/yr, 41 in the sandbox, well above the ~100 floor that the single-name version could not reach), and per-trade economics are strong (sandbox avg_trade_return_pct 1.51%, PF 1.54; developer full-sample +4.24%/trade, positive every year). But the sandbox-year bootstrap Sharpe CI [-1.03, 2.12] still straddles zero (PSR 0.77, only 41 trades, return_kurtosis 31), and within that year the LONG leg was weak (10 longs, win 0.10) while SHORTS carried the book (31 shorts, win 0.52). The full sample is positive both directions per the developer, but the analyst should confirm the long-side convex edge is real across the pooled sample and not a recent-regime artifact.

Risk officer/analyst: confirm 4%-risk-per-sleeve at a 120%-of-equity aggregate gross cap is acceptable; if not, lower risk_pct or max_gross (both are non-searchable infra params, so this does not touch the 3 declared tunables).

risk_pct is 0.04 (4% of equity per engaged sleeve at the chandelier distance), higher than the single-name tail-regime's 2%. It is bounded by the per-leg cap (max_notional_frac*leverage = 35% of equity) and the shared aggregate gross cap (max_gross*leverage = 120% of equity), and every term is a fraction of CURRENT equity so losses cannot compound into size, so this is not a blow-up risk — but with up to 6 asynchronous sleeves and 6x-ATR chandelier stops, a clustered multi-name reversal could realise several ~4% stop-outs at once. Verified the caps are correctly enforced (per-leg and gross headroom both applied in _sleeve_qty, _current_gross excludes the sleeve being entered), so exposure is capped, but the analyst/risk officer should confirm the 4%/sleeve at 120% aggregate gross is within portfolio risk limits.

For live deployment, have the primary sleeve reconstruct its side from cache.positions_open() on restart, as the extra sleeves already do.

Minor restart-recovery inconsistency between the primary and extra sleeves (both unreachable in backtest, live-only). The five EXTRA sleeves correctly recover a lost side from the actual open position (pos.side.name) in on_extra_bar before calling _regime_over. The PRIMARY (BTC) sleeve's _regime_over instead infers a lost side from the z sign (st.side = 1 if st.z >= 0 else -1), which on a live mid-position restart could pick the wrong side. Harmonising the primary to also read cache.positions_open() would make it consistent with the extras.

Backtest Review

Sharpe: 1.07
Total return: 418.03%
Max drawdown: 12.77%
Trades: 408
Win rate: 39.0%
Profit factor: 1.85

Cadence fix worked as designed: 408 trades (from 72), now validatable; Sharpe 1.069 with CI-low 0.5213 > 0, PSR 0.9999 (skew/kurtosis-adjusted, still ~1.0)

Real distributed convex edge: PF 1.85, avg_win ~3x avg_loss, avg_trade_return_pct 4.17%, 218L/190S, positive most years (2020/2021/2023/2024/2026)

Decorrelated (beta 0.077), deployable ($882M capacity), low-DOF shared-param design (3 tunables), genuinely novel and not L52

Monster days are the convex rule functioning (BNB's ~+60% Feb-2021 day, 2024-11 alt run), not a sizing artifact

Extreme tails: return_kurtosis 96.8, skew 5.03 — the diversification premise only partially held (kurtosis rose ~26->97) because crypto majors correlate in the large regime moves, concentrating rather than smoothing the tail

Result is tail-dominated; the deflated-Sharpe over 225 trials and the 15-day holdout (only ~2-3 trades) are the real generalization tests

2025 negative (-2.0%), recent regime weaker; record leans on 2021 (+80%) and 2024 (+43%)

Analysis

avg_trade_return_pct 11.25% (optimized) and profit_factor 3.14 — per-engagement capture is far above fees when it fires

Low max_drawdown (6.2%) and genuinely low beta (0.041) / decorrelated character

Sensitivity clean (0 cliffs), no single-parameter fragility

HARD gate 1 — PBO 0.6547 > 0.5: the parameter selection is more likely than not overfit

HARD gate 2 — holdout Sharpe -0.031 (ratio -0.018 vs 0.70 floor): the edge is zero/negative on unseen data, collapsing from WF-OOS 1.73

Deflated Sharpe 0.298 (<0.95) and optimized Sharpe 1.078 is BELOW expected_max_sharpe 1.223 — the selected result is worse than luck predicts over 225 trials (best-of-N artifact)

Optimizer pushed entry_z to 1.98, cutting the book to 137 trades (~20/yr) — re-creating the exact cadence flaw the diversification premise was meant to fix

Outlier-dependent: return_kurtosis 190; single days (2021-02-20, 2024-12-02 +14.4%, 2024-12-04 +13.0%) carry the return; 2024's +35% is essentially one December fortnight

Not persistent: optimized 2025 -1.36%, 2026 -0.81%

Whole convex tail-regime family abandoned (BTC fee_edge, DOGE overfit, BTC vol-expansion overfit, per-name TSMOM baskets overfit)

Outcome Summary

Diversifying a rare convex trigger across correlated assets does not smooth its tail — crypto majors co-move in exactly the large regime moves the rule targets, so kurtosis concentrated (to 96-190) rather than diversifying away, and the edge collapsed on holdout with PBO > 0.5; a tail-dominated edge whose optimized Sharpe is below the best-of-N noise bar cannot survive multiple-testing correction.

It earned an 'optimize' verdict and ran the full 3-phase optimization, but the analyst abandoned it at the ANALYZING stage on decisive overfit gates: deflated Sharpe 0.298 (selected result worse than best-of-N noise predicts), PBO 0.655 > 0.5, and a holdout that collapsed to zero/negative from a WF-OOS of 1.73. The diversification premise only partially held (majors correlate in large regime moves, so kurtosis rose to 96.8 rather than smoothing), it remained outlier-dependent and non-persistent, and the optimizer re-created the very cadence flaw the design meant to fix — the whole convex tail-regime family has now been abandoned.

A long-short, multi-instrument convex directional basket (3 shared parameters) that ran the factory's validated tail-regime convex rule independently across six majors (BTC, ETH, SOL, BNB, XRP, DOGE) — each sleeve flat by default and engaging only on its own confirmed large-magnitude directional regime with a wide chandelier exit — summed into one book, using diversification rather than acceleration to fix the BTC version's too-rare cadence.

The cadence fix worked and the initial backtest was strong: +418% over 408 trades, Sharpe 1.069 (CI-low 0.52), profit factor 1.85, max drawdown 12.8%, avg_trade_return_pct 4.17%, decorrelated (beta 0.077), positive most years. But optimization exposed overfitting and outlier dependence — deflated Sharpe 0.298 (with optimized Sharpe 1.078 below the expected-max-of-225 of 1.223), PBO 0.655, a holdout Sharpe of -0.031 (ratio -0.018 vs the 0.70 floor), kurtosis rising to 190, and the optimizer pushing entry_z to 1.98 which cut the book back to ~137 trades (~20/yr).

Backtest and paper results are hypothetical. Trading involves risk of loss.