CrossAssetDualMomentumRotationUseq

Cross-Asset Dual-Momentum Rotation (SPY/QQQ/GLD risk sleeve, TLT defensive) on USEQ Daily Bars — Monthly Rebalance, Absolute + Relative Momentum, 12-1 Lookback, Commission-Free, Decades of OOS History

A long-only, multi-instrument, monthly-rebalanced cross-asset ROTATION strategy on liquid USEQ ETFs, using DAILY bars (decades of split-adjusted history). Deliberate pivot away from the dead crypto momentum-confluence family (two XRP variants + LTC/DOT/BNB/SOL/LINK clones all died overfit, L91): the fix is a mechanism whose OUT-OF-SAMPLE history is deep enough to defeat overfitting. Cross-asset dual momentum is a low-parameter allocation rule with 25+ years of clean daily USEQ history, near-zero turnover, and commission-free execution. Distinct from the SPY/TLT binary hysteresis run in flight (that is a two-state single-pair switch; this ranks a 3-asset risk sleeve and gates with absolute momentum) and from the dead L74 intra-equity anomalies (all single-beta bets).

Implements the hypothesis exactly: a 3-asset risk sleeve (SPY/QQQ/GLD) ranked by relative momentum, gated by absolute momentum, with TLT as the defensive leg, monthly-rebalanced on USEQ daily bars. Only three free parameters (lookback_days, skip_days, abs_threshold) are searched against ~20 years of common split-adjusted daily history (GLD lists 2004-11, so common history runs from ~2005-11 after the 13-month warm-up) — that OOS depth is the whole point of the pivot away from the short-history crypto momentum-confluence family. calculate_signal returns the leading risk asset's raw 12-1 momentum every session (a continuous, daily-varying decision variable), while the rotation decision is applied on the calendar rebalance grid. All four legs are managed manually via _submit_entry_instrument / _submit_exit_position (the base class only auto-manages the primary instrument), so should_enter returns None and should_exit False to avoid a competing SPY order; every order action is gated on no leg having a working or in-flight order, and a new entry is deferred to a later bar than the exit that frees the capital, so the book is never in two legs at once. Venue is USEQ (the only venue listing these ETFs) with leverage=1.0, which the code does not reference in sizing — consistent with the long-only, unlevered design. Distinct from the SPY/TLT hysteresis run in flight: that is a two-state single-pair SMA switch; this ranks three risk assets on 12-1 momentum and gates with an absolute-momentum hurdle.

Hard PBO gate failure (0.5516 > 0.5) makes promotion impossible and cannot be waived; it is corroborated by a decisive robustness collapse: deflated_sharpe=0.0, is_significant=false, and the candidate does not survive programme-level FDR (4 of 276) — the selected Sharpe is a best-of-N artifact over 225 trials, not a validated edge. There is no robust region to iterate toward: lookback_days is a cliff parameter and abs_threshold never binds. The strategy is structurally un-validatable by backtest (provenance flags the OOS vault as UNDERPOWERED, ~15 expected trades vs 100 needed for inference), so no parameter change can lift it over the significance gates. This is the recurring cross-asset dual-momentum ETF rotation family (SPY/QQQ/GLD/TLT), which has already been abandoned at least three times as overfit/risk_reject on near-identical mechanics (similarity 0.74-0.82). Per the overfit->abandon rule, re-tuning the same mechanism re-overfits rather than revealing a real edge; not revise_hypothesis because the mechanism itself — not a specific dead target — is what fails.

Long-only cross-asset dual-momentum rotation on USEQ daily bars. One 100%-notional sleeve is rotated monthly between a risk sleeve (SPY, QQQ, GLD) and a defensive leg (TLT). Each month, relative momentum ranks the three risk ETFs by their 12-1 total return (return over `lookback_days` = 252 sessions ending `skip_days` = 21 sessions ago, which drops the short-term reversal month) and picks the leader; absolute momentum then requires that leader's own 12-1 return to exceed `abs_threshold` (a cash hurdle, default 0). If it does, the sleeve holds the leader; if it does not, the sleeve holds TLT. The rebalance date is derived purely from the bar's own UTC (year, month) — the first session of each calendar month — so the schedule is identical across restarts, backtests and paper replay, never anchored to a process-relative bar counter. Sizing is whole shares of ~95% of equity, no leverage, no shorting. Turnover is a handful of switches per year, which matters because USEQ pays no commission but does pay the spread.

If the override is intentional (it is, given the variable lookback), leave as-is; the value is correct.

min_bars_required() overrides the base framework method (flagged by static analysis). This appears deliberate — it correctly scales the warm-up to lookback_days + skip_days + 2 so the momentum window is fully populated before the strategy decides. No functional problem observed, but confirm the base class honors the override for gating calculate_signal dispatch.

Sandbox is healthy (+72%, avg_trade_return_pct 4.34% far above the USEQ spread floor, PF 2.83), but headline Sharpe is modest (0.54, CI [-0.20, 1.30] straddling zero) with benchmark_correlation 0.45 to an equal-weight basket and information_ratio -0.13 vs that basket. Dual-momentum rotation is a low-turnover, well-trodden allocation rule; the analyst should confirm on full ~20yr history that the absolute-momentum TLT gate adds risk-off value beyond simply holding the basket, since the sandbox window (1650 days) does not include the 2004-2020 period where the crash-avoidance edge historically shows up.

Per-trade edge far above the USEQ floor: avg_trade_return_pct 4.98% (floor ~0.05%), PF 3.43, win_rate 0.66, commission 0% and impact only 0.73% of gross — not fee-fragile.

Trade cadence matches the hypothesis: 56 long-only trades (0 shorts) over ~21yr for a monthly rotation that mostly re-holds — the expected ~2.6 switches/yr, not over-trading or a frozen signal.

Deep, clean OOS history (2005-2026 spanning multiple crash regimes) — the deliberate remedy to the dead single-name crypto-momentum overfitting family; sample is more than adequate for optimization.

Viable risk-adjusted profile: Sharpe 0.487 with CI [0.136, 0.824] (positive, not straddling zero), PSR 0.988, CAGR 11.16%, max_drawdown 27.7%, positive alpha +1.88%/yr.

Information ratio vs the equal-weight basket is ~0 (-0.009) and benchmark_correlation 0.55 — the absolute-momentum TLT gate may add little risk-adjusted value beyond simply holding the basket; must be confirmed on the full walk-forward.

Headline Sharpe (0.49) is well below the promotion tier; the edge is real but modest and will be tested hard by deflation over ~225 trials on only 3 free parameters.

Large accumulated per-period prints (e.g. 2026-06 +54%) are multi-month-hold accumulations sampled at trade-close, not real single-month moves — plausible but worth verifying the equity-curve shape during analysis.

Per-trade edge far above the USEQ floor (avg_trade_return_pct 6.38%, PF 4.82, 0% commission, impact only 0.59% of gross) and trade cadence matches the monthly-rotation hypothesis (50 long-only trades, 0 shorts over ~21yr).

Walk-forward is NOT overfit by the ratio test (is_overfitted=false, avg OOS Sharpe 0.672 > 0.5), holdout is nominally positive (Sharpe 0.826 on 11 trades, ratio 1.229, consistent_with_oos=true).

Deep, clean split-adjusted history and a low-parameter allocation rule — the intended fix for the short-OOS crypto momentum family.

HARD gate failure: PBO 0.5516 > 0.5 — the parameter selection is more likely than not overfitting. This can NEVER be waived, so promotion is impossible.

Robustness collapses under multiple-testing correction: deflated_sharpe 0.0, is_significant=false, does not survive programme-level FDR (Benjamini-Hochberg keeps 4 of 276). The selected Sharpe is indistinguishable from best-of-N noise over 225 trials.

lookback_days is a CLIFF parameter (Sharpe 1.22 at base but 0.28 at 328, 0.67 at 268) — no robust region to iterate toward. abs_threshold sits pinned at 0.0 across every variation (the absolute-momentum hurdle never binds).

Structurally un-validatable: the provenance note flags the OOS vault as UNDERPOWERED (~15 expected trades vs 100 needed) — no history slice is both short enough to leave training data and long enough to validate a rule this slow.

max_drawdown 30.3% doubles the 15% equities floor; worst stress window (rate_shock_2022) −30.4% return / −33% DD.

Near-identical siblings have died at least three times: SPY/TLT/GLD relative-momentum (overfit), QQQ/GLD/TLT dual-momentum (risk_reject), SPY/TLT/GLD macro rotation (overfit). The cross-asset dual-momentum rotation mechanism does not survive this pipeline.

Benjamini-Hochberg at q=0.10 over 276 programme candidates keeps 4. A candidate that does not survive here is not distinguishable from the programme's own noise, however good its individual statistics look.

CrossAssetDualMomentumRotationUseq-80a8b797f6

This strategy was a deliberate pivot away from the dead single-name crypto momentum family: a long-only cross-asset dual-momentum rotation across SPY/QQQ/GLD with a TLT defensive leg, chosen because its monthly, low-parameter allocation rule carries decades of clean USEQ history. The backtest review approved optimization on the strength of a per-trade edge far above the fee floor (avg_trade_return_pct ~5-6%, PF 3.4-4.8) and a deep multi-regime sample, and the optimized run returned 854.7% (CAGR 11.04%) at Sharpe 0.58. But the analyst abandoned it after optimization: PBO 0.5516 breached the unwaivable 0.5 hard gate, deflated Sharpe was 0.0 with the candidate failing programme-level FDR, lookback_days was a cliff parameter and abs_threshold never bound, and a 30.3% drawdown doubled the equities floor. With the OOS vault flagged underpowered and near-identical siblings already dead three times, the mechanism itself was judged un-validatable, ending the run at the ANALYZING stage.

A slow, low-turnover rule can be structurally un-validatable by backtest (the OOS vault was flagged UNDERPOWERED at ~15 expected trades vs ~100 needed), so re-tuning the same cross-asset rotation mechanism — which has now died 3+ times on near-identical mechanics — re-overfits rather than revealing a real edge.

The analyst abandoned it after optimization on a hard PBO gate failure (0.5516 > 0.5, unwaivable), corroborated by a robustness collapse: deflated_sharpe 0.0, not significant, and failure to survive programme-level FDR (4 of 276 kept) — the selected Sharpe was a best-of-N artifact over 225 trials.

A long-only, monthly-rebalanced cross-asset dual-momentum rotation on USEQ daily bars, ranking a SPY/QQQ/GLD risk sleeve by 12-1 momentum and gating with absolute momentum into a defensive TLT leg, deliberately pivoting to a low-parameter rule with decades of clean OOS history to escape the dead crypto-momentum family.

The optimized backtest over ~21 years returned 854.7% total (CAGR 11.04%) with Sharpe 0.58, profit factor 4.82, win rate 0.80, avg_trade_return_pct 6.38% and 50 long-only trades, but max drawdown was 30.3% and the deflated Sharpe was 0.0 with is_significant=false. The walk-forward was not flagged overfit (avg OOS Sharpe 0.672) and the holdout passed (Sharpe 0.826 on 11 trades), yet PBO came in at 0.5516.