Strategies

QQQ Risk-On / Defensive-Rotation Trend Regime, Long-Only (QQQ.USEQ growth leg <-> GLD/TLT defensive legs, 1-DAY bars, SMA + hysteresis band, 2-parameter)

Outcome: Abandoned

QqqRiskOnDefensiveSplitRotationGldTltHysteresis

Outcome Summary

This strategy was a deliberate pivot away from the dying crypto momentum family toward equity index-ETF regime rotation, extending a surviving SPY<->TLT sibling onto higher-beta QQQ with a split GLD+TLT defensive book. Its full-history backtest passed the pre-optimization gate and the analyst approved it for optimization, but the 3-phase run exposed severe overfitting: in-sample Sharpe reversed out-of-sample, PBO hit 0.65, deflated Sharpe was 0.115, and the optimizer suppressed drawdown only by cutting exposure to ~13%. After one iteration the analyst abandoned it, noting the mechanism loses money in the stressed vol tercile and across 2022 rate-shock windows, and that four near-identical siblings had already died overfit or risk-rejected.

Strategy report

Hypothesis

A deliberate FAMILY PIVOT away from the collision. The proposed AVAX dual-timeframe momentum-confluence long-short is a re-skin of a family that keeps dying (AVAX itself died fee_edge; DOT/BNB clones died overfit), so this proposal abandons crypto directional momentum entirely and moves to the ONE family the live pipeline evidence shows is currently surviving: equity index-ETF trend/regime rotation (archetype 3 in the equity playbook). A sibling of this exact family, UsRiskOnRiskOffRegimeRotationSpyTltHysteresis, advanced through backtest_review INTO optimizing on 2026-09-06 rather than being abandoned. This is a breadth extension of that surviving structure onto a DIFFERENT, higher-beta growth index (QQQ) with a split defensive book (GLD + TLT) instead of the SPY<->TLT single-pair switch, so it is materially distinct from the in-flight sibling (different growth instrument, different defensive composition, split-weight defensive leg rather than binary). The strategy holds the growth leg (QQQ) while it is in a confirmed uptrend regime and rotates the entire book into a defensive split (gold + long Treasuries) when the growth regime breaks. It is long-only (cash account), commission-free on USEQ, trades on 1-DAY bars with DECADES of split-adjusted history, and turns over only a handful of times per year -- so fee/impact drag is structurally negligible, which is the opposite failure mode from the fee_edge graveyard that kills fast crypto perp mechanisms. Kept to 2 parameters (trend SMA length, hysteresis band width) to resist the overfit deaths that killed the higher-knob single-name momentum clones.

Implements the hypothesis's family pivot: same surviving structure as the SPY<->TLT hysteresis sibling but on the higher-beta growth index (QQQ) with a SPLIT defensive book (GLD + TLT at 50/50 of the sleeve) instead of a binary single-pair switch, and each defensive leg is separately trend-gated so a rotation never dumps the whole sleeve into a broken hedge. Exactly two searched parameters (sma_days, hysteresis_pct) plus a structural notional cap, per the hypothesis's anti-overfit constraint. All three legs are managed manually via _submit_entry_instrument/_submit_exit_position with a pending-order gate and an entry deferred one bar after the freeing exit, because CASH-account fills must settle before the next leg can be bought; the two defensive entries share one cash budget so they cannot jointly overspend, and target weights are applied to a flat-book capital snapshot (get_account_equity on a CASH account is the shrinking USD balance, not the mark-to-market book). calculate_signal returns QQQ's continuous distance from its own SMA every session; entry/exit decisions are thresholded via the hysteresis state machine, which advances for every leg on every bar so no leg's state depends on which branch queried it. Routed to USEQ (long-only, leverage 1.0) because the strategy never shorts and needs no margin; QQQ/GLD/TLT all have decades of split-adjusted daily history in the catalog.

Overfit with negative out-of-sample generalization: walk-forward IS 0.374 -> OOS -0.178 (is_overfitted=true, 2 of 3 windows non-positive), OOS Sharpe -0.178 <= 0, PBO 0.65 > 0.5, holdout only 3 trades. Confirmed quantitatively by failed deflated Sharpe: DSR=0.115 vs expected-max 0.798 over 150 trials (47 effective), is_significant=false, does not survive programme FDR (candidate_p 0.88, keeps 4 of 280). The optimizer collapsed notional_pct from 0.95 to 0.134 — it found no edge and simply minimized exposure, so the flattering optimized PF 11.6 / DD 4.2% is an artifact of a near-flat book, not a captured edge; the strategy also loses money in the stressed vol tercile and across the 2022 rate-shock windows, i.e. the defensive rotation fails to hedge when needed. This is the USEQ index/defensive regime-rotation family that has repeatedly died overfit — four near-identical siblings (similarity 0.63-0.77) already abandoned overfit/risk_reject, including the SPY<->TLT hysteresis structure this extends. Overfit -> abandon rather than revise: pointing the same absolute-trend rotation mechanism at another growth index re-overfits rather than fixing a premise flaw, and there is no robust parameter region to iterate toward (the sensitivity grid is uniformly ~0.5 in-sample but the WF-OOS is negative).

Implementation

Long-only US ETF regime rotation on 1-DAY USEQ bars. One ~95%-notional sleeve is held in QQQ while QQQ is in a confirmed uptrend (close above its sma_days SMA by more than hysteresis_pct, staying 'up' until it falls the same distance below the SMA). When the growth regime breaks, the book rotates into an equal-weight defensive split of GLD and TLT, each leg admitted only while it passes the identical SMA+hysteresis filter; if neither defensive leg trends up (the 2022 stocks-and-bonds-together case) the strategy sits in cash. Whole-share sizing, no leverage, commission-free venue, a handful of round trips per year.

Verification Results

Leave as-is if the base template calls min_bars_required(); otherwise the extra history is only guarded by the internal len(closes) < sma_days checks (which are present), so behaviour is still safe.

min_bars_required() overrides a framework method (Layer-1 base_shadow warning). This appears deliberate — the override returns sma_days+2 so the base warm-up gate scales with the trend length. No functional problem, but confirm the base class actually consults this override for the primary-instrument warm-up rather than its own default.

Sandbox looks strong (Sharpe 1.00, total_return 143%, PF 96, avg_trade_return_pct 12.75%) and there is no profitability doubt, but the smoke window is only ~8 trades over 1650 bars, so the edge estimate is low-N; the real full-history backtest (decades of daily QQQ/GLD/TLT) is where regime behaviour and 2022-style stocks+bonds-down periods should be judged. The stress_windows in the report are crypto-labelled (covid/luna/ftx) and irrelevant to a USEQ strategy — report-template artifact, not a code issue.

Backtest Review

Sharpe: 0.60
Total return: 672.12%
Max drawdown: 26.90%
Trades: 61
Win rate: 52.5%
Profit factor: 5.34

USEQ equity ETF regime-rotation — the ONE family currently surviving the live pipeline (sibling UsRiskOnRiskOffRegimeRotationSpyTltHysteresis and several USEQ rotation strategies advanced to optimizing this week). This is a genuine breadth extension (QQQ growth leg, split GLD+TLT defensive book), not a re-skin of a dead crypto momentum clone.

avg_trade_return_pct 4.25% is ~85x the USEQ ~0.05% fee floor; commission $0 and impact only 0.37% of gross — cost drag is structurally negligible exactly as the slow-turnover (61 trades / 27 years) design intends. This is the opposite failure mode from the fee_edge graveyard.

profit_factor 5.34, MTM-basis Sharpe 0.596 with CI [0.23, 1.01] that does NOT straddle zero, PSR 0.9959, positive alpha +2.35%/yr (beta 0.60), max_drawdown 26.9% (<60%).

Trades implement the hypothesis: 61 long trades, 0 shorts (long-only cash account), and the 2022 rate-shock window (QA's specific stocks+bonds-down concern) shows defensive rotation rather than a blow-up — QA's low-N concern is not borne out on full history.

information_ratio ~0.004 — the rotation barely beats a passive equal-weight QQQ/GLD/TLT basket; the actual active edge over the naive book is thin.

Loses in the stressed vol tercile (return -9.7%, Sharpe -0.01) and max_drawdown_duration ~1407 days (~3.8 yr) is long.

Sparse close-date daily_returns are lumpy realized-PnL attributions of long-held positions (e.g. 2022-01-24 +63.6% on one close date); the honest risk-adjusted number is the modest MTM Sharpe 0.596, not the eye-catching annual/daily prints.

Modest headline Sharpe means the deflated-Sharpe / holdout gates in the 3-phase run will be the decisive test.

Analysis

USEQ commission-free daily rotation — fee/impact drag is structurally negligible (impact 0.18% of gross), so this is not a fee_edge death.

Sensitivity is clean (0 cliffs) and per-trade return clears the venue floor by a wide margin.

Low nominal drawdown on the optimized config (4.2%) — but see weaknesses, it is an artifact of collapsed exposure.

OVERFIT: walk-forward IS 0.374 -> OOS -0.178 (is_overfitted=true), 2 of 3 windows non-positive. The edge does not generalize out-of-sample — it reverses.

OOS Sharpe -0.178 <= 0 and PBO 0.65 > 0.5 (HARD gates) — promotion is impossible and the parameter selection is more likely than not overfitting.

Optimizer collapsed notional_pct 0.95 -> 0.134: it found no edge and minimized market exposure to suppress drawdown. The headline PF 11.6 / DD 4.2% is an artifact of near-flat exposure, not a captured edge.

Deflated Sharpe 0.115 (<< 0.95), is_significant=false, fails programme-level FDR (candidate_p 0.88, keeps 4 of 280) — the selected Sharpe is indistinguishable from best-of-N noise over 150 trials.

Holdout carries only 3 trades — statistically empty; the nominal 0.71 holdout Sharpe is meaningless at that sample.

Loses money in the stressed vol tercile (Sharpe -0.17) and in every 2022 rate-shock window — the defensive rotation does not actually hedge when it matters.

This is the recurring USEQ index/defensive regime-rotation family: four near-identical siblings (similarity 0.63-0.77) all abandoned overfit/risk_reject, including the very SPY<->TLT hysteresis structure this claims to extend.

Benjamini-Hochberg at q=0.10 over 280 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.

Outcome Summary

Pointing the same absolute-trend regime-rotation mechanism at a new growth index re-overfits rather than fixing a premise flaw — this USEQ index/defensive rotation family repeatedly dies overfit, and a strategy this slow simply lacks the trade count for statistical validation.

The post-optimization analyst issued an abandon verdict: the edge failed to generalize out-of-sample (OOS Sharpe <= 0, PBO > 0.5, fails programme FDR), and the optimizer collapsed notional_pct from 0.95 to 0.134 — minimizing exposure rather than capturing an edge — so the flattering optimized PF 11.6 / 4.2% drawdown was an artifact of a near-flat book.

A long-only, commission-free USEQ daily rotation that holds QQQ while it is in a confirmed SMA-plus-hysteresis uptrend and rotates into an equal-weight split defensive book of GLD and TLT (or cash) when the growth regime breaks, using just two searchable parameters (sma_days, hysteresis_pct).

The initial backtest over ~27 years looked strong (Sharpe 0.596, profit factor 5.34, avg_trade_return_pct 4.25%, +2.35%/yr alpha, 26.9% max drawdown, 61 trades), but the 3-phase optimization broke down: walk-forward IS Sharpe 0.374 flipped to OOS -0.178 (is_overfitted=true), PBO 0.65, deflated Sharpe 0.115, and the holdout carried only 3 trades.

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