Strategies

US Risk-On/Risk-Off Regime Rotation — SPY vs Defensive Sleeve (TLT/Cash), Single Long-SMA Trend Filter with Hysteresis Band, Decades of Daily History (2-Parameter)

Outcome: Abandoned

UsRiskOnRiskOffRegimeRotationSpyTltHysteresis

Outcome Summary

A slow, months-long regime-rotation mechanism generates too few trades to ever be statistically validated in-sample, and reframing the recurring cross-asset macro-rotation family onto adjacent ETFs re-overfits every time.

Strategy report

Hypothesis

A LONG-ONLY (always-long-ONE-asset or flat) cross-asset REGIME-ROTATION strategy on the USEQ venue, rotating a single 100%-notional sleeve between the equity index ETF SPY (risk-on) and a defensive government-bond ETF TLT (risk-off), stepping to CASH when neither is trending. This is a deliberate FIVE-AXIS pivot away from everything currently saturating the portfolio: (1) VENUE — USEQ is 0.5% of 2922 experiments vs a BINANCE that is 70.8% (target <=40%); this is the single most under-represented venue with real data behind it. (2) MECHANISM — not momentum-confluence, not funding/OI/liquidation gating, not VWAP reversion, not cross-sectional rank rotation, not options VRP — all documented dead classes. It is the risk-on/risk-off defensive-rotation archetype explicitly endorsed as 'the most robustly backtestable strategies in the whole factory' because of decades of daily history. (3) DATA — pure daily OHLCV, ZERO supplementary feeds (the highest-survival tag family at 3%), so no data_unavailable/data_wall risk. (4) HORIZON — long_term with regime holds measured in MONTHS, so turnover is ~2-4 switches/year/sleeve and fee/impact drag is negligible on a commission-free venue. (5) OVERFIT DEFENSE — only TWO parameters and a multi-decade daily backtest window, the direct antidote to the #2 killer (371 overfit deaths, all short-history/high-param curve fits). The core insight the meta-learner has NOT killed: equities carry a persistent equity-risk-premium in trending regimes and a flight-to-quality bid into bonds during sustained equity downtrends, and a slow trend filter harvests the ERP while side-stepping the deep bear leg. The known 2022 failure mode (stocks AND bonds fell together) is defended structurally: the defensive sleeve is only held when TLT is ITSELF in an uptrend, otherwise the strategy sits in cash rather than blindly buying bonds.

Implements the hypothesis exactly: five-axis pivot (USEQ venue, defensive-rotation mechanism, pure daily OHLCV, months-long holds, 2 parameters). The equity-risk-premium is harvested while SPY trends above its long SMA, and the deep bear leg is side-stepped by rotating out. The 2022 stocks-and-bonds-fall-together failure mode is defended structurally: TLT is only bought when TLT ITSELF passes the identical trend filter, otherwise the sleeve sits in cash. Hysteresis is the second parameter and the reason a single slow SMA is tradeable rather than whipsaw-prone. Venue is USEQ with leverage 1.0 and long-only sizing in whole shares (size precision 0), sized off get_account_equity() * notional_pct. Both legs are driven manually via _submit_entry_instrument / _submit_exit_position with a pending-order gate and a deferred entry, because separate USEQ instruments do not net and an unsettled fill is invisible to the cache; should_enter/should_exit therefore stay inert so the base class never submits a competing SPY order. calculate_signal returns SPY's continuous distance-from-SMA every session, and the regime state is a pure function of price history (no uptime-anchored counters, no calendar rebalance grid).

Overfit with negative out-of-sample generalization on a strategy that is structurally too slow to validate. Four HARD gates fail: walk-forward is_overfitted=true (IS 1.28 -> OOS -0.016, window-2 IS 2.95 -> OOS -1.02), OOS Sharpe -0.016 <= 0, PBO 0.754 > 0.5, and holdout carries only 3 trades (<10). Quantitatively confirmed by failed deflated Sharpe: DSR=0.0, expected-max 1.37 over 225 trials, is_significant=false, does not survive programme FDR (4 of 275). The 18-35 trades over 24 years leave no in-history slice that can both train and validate this cadence (OOS vault flagged UNDERPOWERED). This is not a fee or tuning problem - the sensitivity surface is smooth with no robust region to iterate toward, the edge simply does not appear out-of-sample. It is the recurring cross-asset macro regime-rotation family on USEQ (SPY/TLT, QQQ/GLD/TLT, SPY/TLT/GLD siblings all abandoned overfit/risk_reject/fee_edge); reframing the same defensive-rotation mechanism onto adjacent macro ETFs has re-overfit every time, so revise_hypothesis is not warranted. Per the decision framework, overfit -> abandon.

Implementation

Long-only US risk-on/risk-off regime rotation on the USEQ venue, daily bars. A single 100%-notional sleeve is rotated between SPY (risk-on), TLT (risk-off) and CASH. Each ETF carries the SAME trend filter: close vs its own sma_days SMA with a symmetric hysteresis_pct band -- the state flips to 'up' only when price exceeds the SMA by the band and back to 'down' only when it falls the same distance below, so oscillation around the line does not whipsaw the book. SPY up -> hold SPY; SPY not up but TLT up -> hold TLT; neither -> flat. Two searchable parameters (sma_days, hysteresis_pct); zero supplementary data; ~2-4 switches per year, which keeps spread+impact drag negligible on a commission-free venue.

Verification Results

Remove notional_pct from _param_bounds (leave sma_days and hysteresis_pct only), keeping notional_pct a fixed structural constant as described in the code comment.

The hypothesis explicitly sells a TWO-parameter overfit defence (sma_days, hysteresis_pct), but config.parameters._param_bounds includes notional_pct [0.1, 1.0] as a third searchable bound. notional_pct only affects sizing (not the signal), and the code clamps it, so this is benign for correctness, but exposing it to the optimizer mildly weakens the '2-parameter, decades of history' overfit-defence narrative. Consider dropping notional_pct from _param_bounds so the search space matches the stated 2 parameters.

No change required.

Static analysis flags min_bars_required() as shadowing a base method. This is a FALSE POSITIVE: the base template calls self.min_bars_required() (base_template.py:478) to gate warmup, so overriding it to require sma_days+2 bars is the intended, supported extension point. No action needed; noted so the warning is not mistaken for a defect.

The hypothesis's central claim is a structural defence against the 2022 stocks-and-bonds-fall-together regime, yet the sandbox's rate_shock_2022 stress window still shows -15.5% (Sharpe -2.48, DD 16.9%) and the 'normal' vol tercile is the weakest (Sharpe 0.41). The cash-when-neither-trends rule reduces but does not eliminate the 2022 leg, and TLT's slow SMA lags the flight-to-quality turn. Headline sandbox is nonetheless positive (+69.8%, Sharpe 0.82, PF 5.46, PSR 0.97) over ~4.5y. Worth watching that 2022 behaviour on the full multi-decade backtest, but this is a performance question for the analyst, not a correctness blocker.

Backtest Review

Sharpe: 0.51
Total return: 435.39%
Max drawdown: 25.86%
Trades: 35
Win rate: 45.7%
Profit factor: 4.26

Trades implement the hypothesis: 35 long-only rotations, 0 shorts, ~1.5 switches/yr over 24y of TLT-available history — the months-long regime cadence the premise predicts.

Trend-following signature is coherent: win_rate 0.457 with profit_factor 4.26 (small losers, large winners), recovery_factor 29.6.

Economically viable on USEQ: avg_trade_return_pct 5.45% is ~100x the ~0.05% USEQ floor; commission-free, impact only 0.56% of gross — fee drag is a non-issue.

Strong overfit defence for the 3-phase run: only TWO free parameters (sma_days, hysteresis_pct) over decades of daily history, and an under-represented venue (USEQ) diversifying a Binance-heavy corpus.

Reliable, plausible result: metrics_reliable=true, return is real MTM with only ~14% unrealized at end, no artifact signature.

Modest risk-adjusted edge: Sharpe 0.51, and information_ratio is negative (-0.14) with thin alpha (~1.46%/yr) vs the SPY+TLT basket benchmark — partly explained by excluded SPY dividends (~1.5-2%/yr), but the active edge is small.

2022 rate-shock window still lost -16% (Sharpe -2.39): the cash-when-neither-trends rule mitigates but does not eliminate the stocks-and-bonds-fall-together leg (QA's concern — a post-optimization performance question).

35 trades is a low sample for parameter fitting; acceptable only because the search space is 2-dimensional and the mechanism is slow/structural.

Analysis

Trades implement the hypothesis: 18-35 long-only rotations, zero shorts, months-long regime holds, drawdown 18.3% inside the pre-registered 25% cap, and per-trade return well above the USEQ fee floor (fees are a non-issue here).

Sensitivity grid is smooth (0 cliffs, full-period Sharpe ~0.53-0.66 across the sma_days x hysteresis surface) - the parameters are stable, the problem is not parameter fragility.

Decisively OVERFIT and non-generalizing: walk-forward IS 1.28 -> OOS -0.016 (is_overfitted=true), window-2 IS 2.95 -> OOS -1.02, PBO 0.754 > 0.5. The full-period Sharpe is fitted data; on unseen windows the edge disappears or reverses.

Statistically indistinguishable from noise: DSR=0.0, expected-max Sharpe 1.37 over 225 trials (29 effective), is_significant=false, does not survive programme-level FDR (keeps 4 of 275).

Structurally un-validatable: only 18-35 trades over 24 years. Holdout carries just 3 trades (<10), and the OOS vault is explicitly flagged UNDERPOWERED (~10 expected trades, need ~100). The strategy is too slow for any in-history slice to both train and validate it.

Matches a repeatedly-dead family: SPY/TLT and QQQ/GLD/TLT and SPY/TLT/GLD macro regime-rotation siblings on USEQ have all been abandoned (overfit / risk_reject / fee_edge) - reframing the same defensive-rotation mechanism onto adjacent macro ETFs has re-overfit every time.

Pre-registered predictions missed on the two that matter: declared min_trades 60 (actual 18) and min_oos_sharpe 0.8 (actual -0.016).

Benjamini-Hochberg at q=0.10 over 275 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

The analyst abandoned it at the ANALYZING stage after the initial backtest passed BACKTEST_REVIEW (verdict: optimize): four HARD gates failed — is_overfitted=true, OOS Sharpe -0.016 (≤0), PBO 0.754 (>0.5), and only 3 holdout trades (<10) — and the OOS vault was flagged underpowered.

A long-only, two-parameter regime-rotation strategy on the USEQ venue that rotates one 100%-notional sleeve between SPY (risk-on), TLT (risk-off, only when TLT itself trends up), and cash, using a single long-SMA trend filter with a hysteresis band over decades of daily history.

The optimized full-period backtest looked strong on paper — total return 561%, CAGR 10.3%, Sharpe 0.57, profit factor 15.0, max drawdown 18.3%, and avg_trade_return_pct 12.5% far above the USEQ floor — but over just 18 trades in 24 years. Walk-forward collapsed out-of-sample (avg IS Sharpe 1.28 vs OOS -0.016, window-2 IS 2.95 vs OOS -1.02), with PBO 0.754 and deflated Sharpe 0.0.

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