CrossAssetTimeSeriesMomentumManagedFuturesLS

Diversified Cross-Asset Time-Series Momentum (Managed-Futures Style), Long-Short on a USEQ Multi-Asset ETF Book (SPY/TLT/GLD/DBC/EFA, Daily Bars, Decades of History, Per-Asset Trend Sign, Inverse-Vol Weighted, Monthly Signal, 3-Parameter)

A LONG-SHORT, MULTI-INSTRUMENT, macro-horizon time-series (absolute) momentum book across five low-correlation, decades-of-history US ETFs spanning distinct macro risk factors: SPY (US equity), TLT (long Treasuries), GLD (gold), DBC (broad commodities), EFA (developed intl equity). Each instrument is INDEPENDENTLY held LONG or SHORT by the sign of its OWN 6-month (126 trading-day) trend, inverse-volatility weighted so no single asset dominates, rebalanced monthly (every ~21 daily bars). This is the classic AQR/Moskowitz 'Time Series Momentum' managed-futures construct, DELIBERATELY relocated off crypto perps (where L77 killed it purely on fee drag: the 6-perp vol-targeted TSMOM died fee_edge despite Sharpe 0.84) onto USEQ, where commissions are ZERO (~0.02-0.05% RT spread+impact on these hyper-liquid ETFs) and daily history runs DECADES — deep enough for walk-forward + holdout without curve-fitting. It is a genuinely different construct from the risk-on/off ROTATION strategies already in the pipeline (SPY/TLT/GLD hysteresis switches, three-asset cash-escape): those pick ONE winner and flee to cash; this holds every asset simultaneously long OR short by its own trend, capturing crisis-alpha from the SHORT legs (short SPY + long TLT/GLD in a 2008/2020/2022-style drawdown) that a long-only rotation cannot access.

Iteration 5 addresses the sole reported failure — avg_trade_return_pct 0.002% vs the 0.05% USEQ floor over 196 trades — as a TURNOVER defect, not a signal defect (Layers 1-3 all passed and the construct matches the hypothesis). Iteration 4 flipped a leg on any sign change of its 6-month return, so a leg oscillating around mom=0 was flattened and re-entered month after month, and every monthly inverse-vol weight wiggle above a 20% band paid spread; the multi-month trend hold that generates the per-trade return was therefore never realized in one piece. Two minimal, turnover-only changes: (1) flip hysteresis — a leg keeps its current direction until the OPPOSITE trend reaches flip_threshold = 0.30 horizon-vol units, which cuts whipsaw round trips while never suppressing a short leg or leaving a leg flat (the crisis-alpha short SPY/EFA legs remain fully available; hysteresis only delays a flip until the reversal is real); (2) the resize band widens from 20% to 50% of target notional, so routine monthly weight drift no longer trades. Everything else is byte-for-byte the previous iteration: same 126-session horizon, same 63-session inverse-vol weights, same monthly calendar-anchored clock, same 5-ETF USEQ universe, same manual multi-leg execution with deferred sign flips, same continuous inverse-vol-weighted t-stat signal. Expected effect: trade count falls several-fold and each closed position spans a full trend regime (typically months), so per-trade return should move from the spread-scale 0.002% into the whole-percent range that clears the 0.05% USEQ floor. The hysteresis state is restart-safe (read from the open position first, in-memory sign only when flat), and both new clamps are declared in _param_bounds.

Overfit with negative OOS generalization (is_overfitted, OOS -0.113, PBO 0.68 — all hard), no robust parameter region (5 cliffs, flat noisy heatmap), and a hollow near-zero holdout. Not a fee problem — iteration 5 already fixed turnover and the per-trade edge clears the USEQ floor 12x, so the developer's remaining lever has been pulled and the edge still isn't there. Structurally un-validatable at 128 trades / 20 years with an UNDERPOWERED OOS vault. Sixth USEQ cross-asset macro ETF construct on the same SPY/TLT/GLD history to die overfit/risk_reject (sim 0.73–0.76 to five prior abandons), so per the overfit→abandon rule this is not a revise_hypothesis target.

Long/short cross-asset time-series (absolute) momentum managed-futures book on five USEQ ETFs (SPY, TLT, GLD, DBC, EFA), daily bars. Each leg is held long or short by the sign of its own 126-session (6-month) total return, inverse-volatility weighted off a 63-session realized-vol estimate, with the book's signs re-evaluated monthly on a calendar-anchored clock (first evaluated session of each new month). Legs are never flat: a leg's direction flips only once the opposite trend reaches flip_threshold horizon-vol units, and a leg already carrying the correct direction is only re-sized when its notional has drifted more than rebalance_band (50%) from target — so each closed position spans one whole multi-month trend leg rather than being chopped into monthly round trips. Gross notional = equity x gross_exposure x leverage (2x Reg-T), whole shares, zero commissions on USEQ with ~0.02-0.05% round-trip spread+impact.

Confirm the base honors the override for warm-up gating; no change needed if it does.

min_bars_required(self) overrides the framework method (static-analysis base_shadow warning). This appears deliberate here (the strategy needs max(lookback_days, vol_lookback_days+1)+2 primary bars before signals are valid) and returns a sane value, so it is informational only.

Sandbox (43 trades, ~6.5y window) is healthy: avg_trade_return_pct 1.928% (well above the ~0.05% USEQ spread/impact floor), PF 1.52, Sharpe 0.76, all three vol terciles positive with the stressed tercile strongest (+32.6%) — consistent with the crisis-alpha short-leg thesis. No profitability doubt to flag; the full-history backtest and optimizer will decide edge.

Trades MATCH the hypothesis and the construct is a genuinely distinct class: 65 long / 63 short balanced TSMOM book across 5 low-correlation macro ETFs (SPY/TLT/GLD/DBC/EFA), inverse-vol weighted, monthly rebalanced by each asset's own 6-month trend sign. NOT a single-name OHLCV momentum clone (L82/L93), NOT a crypto dollar-neutral perp basket (L45).

avg_trade_return_pct 0.6313% (pct_notional) is comfortably above the USEQ ~0.05% floor and above the 0.10% preference. The negative-avg MEASUREMENT ARTIFACT that dominated iterations 1/2/4 is fully resolved and now consistent with PF 1.147, expectancy +$147/trade, avg_win $4,470 vs avg_loss $2,118 at a 34% win rate — a legitimate positive-expectancy trend book, not fee-fragility.

Reliable 128-trade sample over ~20 years of real daily history (all five legs present from 2006), deep enough for walk-forward + holdout. impact_cost_pct only 4.27%, capacity ~$55M — costs are not eating the edge.

Real tunable parameter space exists (lookback_days, vol_lookback_days, gross_exposure, flip_threshold, rebalance_band), so optimization has something to tune toward rather than curve-fit noise.

Sharpe 0.187 with CI [-0.162, 0.538] straddling zero, and PF 1.147 sits below L75's 1.2 skepticism line — the edge is thin and may not survive the 3-phase deflation/holdout.

Crisis-alpha thesis is NOT confirmed on full history: the stressed vol tercile is the WEAKEST regime (return -1.48%, Sharpe 0.03) and 2020 covid was -8.25%; the book earns in normal (+32%) and calm (+13%) regimes, contradicting the QA sandbox where stressed was strongest. The short legs are not delivering the claimed crisis alpha out-of-sample.

end_unrealized_pct 23.2% of the 48.4% headline — realized track record is still positive (~25%) but a meaningful slice is open-position MTM; watch this at the promotion stage.

max_drawdown 37% is high; recovery_factor 0.91.

Implementation faithfully matches the hypothesis: 65 long / 63 short legs across SPY/TLT/GLD/DBC/EFA, multi-month holds (238d, 268d in sampled trades), inverse-vol sized. Iteration 5's flip hysteresis DID fix the turnover artifact — avg_trade_return_pct is now +0.63%, ~12x the USEQ ~0.05% floor. This is not a fee-edge failure.

Payoff structure is the correct TSMOM signature: win rate 0.344 with 2.11 payoff ratio (avg_win $4,470 vs avg_loss $2,118) — the low hit rate is not a defect.

Crisis-alpha from short legs is genuinely present (2008 +13.4%, 2022 +12.2%, rate_shock_2022 +10.9% at Sharpe 1.22).

Capacity is not a constraint ($54.9m, impact 4.3% of gross, zero commissions).

THREE unwaivable HARD gates fail: is_overfitted=true (IS 0.442 → OOS -0.113), avg OOS Sharpe -0.113 ≤ 0, PBO 0.6786 > 0.5. The one window with in-sample edge (IS 1.326) went to OOS -0.839 — the edge reverses, it does not merely decay.

Degenerate walk-forward: ~75 of 150 trials returned Sharpe exactly 0.0, window 1 produced nothing, n_trials_effective 24 — the 'optimum' is a best-of-N pick from effectively one window.

5 of 7 parameters are cliffs including both model parameters; the lookback × flip_threshold heatmap is a uniformly noisy 0.09–0.28 field with no plateau — nothing to tune toward.

Holdout 'passed' is hollow: Sharpe 0.0862 on 27 trades with sharpe_se 0.446 — indistinguishable from zero.

Even fitted, full-period economics are weak: Sharpe 0.187 with CI [-0.162, 0.538] straddling zero, PF 1.147, 37.2% DD vs the pre-registered 20% cap, Calmar 0.042, ~14yr DD duration, and end_unrealized_pct 23.24 means ~half the 48% headline is un-exited mark-to-market.

Stressed-vol tercile returns -1.5% — the crisis-alpha premise fails exactly where it should pay.

0 of 5 pre-registered criteria met (evidence, not a floor).

Structurally underpowered: 128 trades over 20 years of usable 5-asset history; provenance flags the OOS vault UNDERPOWERED (~35 expected vs 100 needed trades).

Per-trade edge below costs: avg_trade_return_pct 0.002% vs the 0.05% fee floor over 196 trades. Optimization cannot fix this; change the mechanism, holding period or timeframe so the per-trade return clears the floor, or recommend abandoning.

CrossAssetTimeSeriesMomentumManagedF-810876e419

The strategy took the AQR/Moskowitz time-series momentum construct that had died on crypto perp fee drag and moved it to zero-commission US ETFs, holding SPY, TLT, GLD, DBC and EFA each long or short on its own 6-month trend with inverse-vol weights and a monthly rebalance. After five iterations — the last of which added flip hysteresis and a wide resize band to stop the monthly clock from chopping multi-month trend holds into spread-paying round trips — the backtest cleared the pre-optimization gate at avg_trade_return_pct 0.63% and earned an 'optimize' verdict despite a thin Sharpe of 0.187. Optimization then broke it: the one window with in-sample edge (IS Sharpe 1.326) reversed to OOS -0.839, PBO hit 0.679, every model parameter registered as a cliff, and the 'passing' holdout was a 0.086 Sharpe on 27 trades that was statistically indistinguishable from zero. The analyst abandoned it as overfit and structurally un-validatable — not a fee failure, but an edge that was never there, and the sixth such USEQ macro-ETF construct to end the same way.

Fixing the cost/turnover problem does not create an edge — iteration 5's flip hysteresis lifted per-trade return to ~12x the USEQ floor, yet 128 trades over 20 years remain structurally too few to validate a slow monthly construct, so future macro-ETF trend ideas need either far more instruments or a faster clock to generate a testable sample.

The analyst abandoned it after optimization on three unwaivable hard gates — is_overfitted true, average OOS Sharpe at or below zero, and PBO above 0.5 — compounded by 5 of 7 parameters flagged as cliffs with no plateau in the lookback x flip_threshold heatmap and a degenerate walk-forward (n_trials_effective 24, ~75 of 150 trials at exactly 0.0 Sharpe). It was the sixth USEQ cross-asset macro ETF construct on the same SPY/TLT/GLD history to die overfit or risk-rejected, so the overfit-to-abandon rule blocked a hypothesis revision.

A classic managed-futures time-series momentum book that held each of five low-correlation USEQ ETFs (SPY/TLT/GLD/DBC/EFA) independently long or short by the sign of its own ~6-month trend, inverse-vol weighted and rebalanced monthly, relocated off crypto perps to a zero-commission venue with decades of daily history.

The full-history backtest (1993-2026, 8,458 bars) returned 48.4% total with Sharpe 0.187 (CI -0.162 to 0.538), profit factor 1.147 and 37.2% max drawdown across 128 trades (65 long / 63 short), with a correct TSMOM payoff signature of 34.4% win rate against a 2.11 payoff ratio and avg_trade_return_pct of 0.63%. Optimization was far worse: average walk-forward OOS Sharpe -0.113 against IS 0.442, PBO 0.679, and a holdout Sharpe of 0.086 on 27 trades with standard error 0.446.