holy-grail-tqqq

Companion code for The Holy Grail of Investing — a 27-year backtest of a simple moving-average rotation between QQQ and TQQQ.

The strategy

One rule, two ETFs:

• If EMA(5) of QQQ > EMA(200) of QQQ → hold TQQQ (3× leveraged Nasdaq)
• Else → hold QQQ (unleveraged Nasdaq)

Evaluated at each daily close; trade executes the next close. No discretion, ~3–4 rotations per year.

Headline results (1999-03-10 → 2026-04-18)

Strategy	CAGR	Max DD	$10K → final	Sharpe
Rotation (this code)	15.88%	−95.5%	$543,387	0.67
QQQ buy-and-hold	10.52%	−83.0%	$150,523	0.52
S&P 500 buy-and-hold	8.4%	−55.2%	$89,000	0.49
TQQQ buy-and-hold	1.37%	−99.98%	$14,439	0.22

Rotation beats QQQ B&H in 79% of rolling 3-year windows by an average of 16 percentage points.

Pre-2010 synthetic TQQQ

TQQQ launched in 2010-02-11. For the 1999–2010 window we reconstruct synthetic TQQQ daily returns from QQQ:

synth_tqqq_ret = 3 × QQQ_ret − expense_daily − financing_daily − slippage_daily

with year-by-year Fed-funds rates for financing, 0.84% annual expense ratio, and 30 bps annual rebalancing slippage. Calibrated against real TQQQ 2010–2026 for drift correction. See scripts/synth_tqqq_v2.py.

Stress tests

The file scripts/peer_review_fixes.py runs seven separate robustness checks after adversarial peer review of the initial result:

1. Gayed 2016 baseline — price > SMA(200) → leveraged, else T-bills
2. Parameter robustness grid — {SMA, EMA} × {5, 10, 20, 50} × {150, 200, 250}
3. Deflated Sharpe under N = 10 / 100 / 1000 effective tests
4. Start-date sensitivity — 1999 / 2003 / 2010
5. Volatility-regime-conditional synthetic error — how wrong is the pre-2010 reconstruction in high-vol periods?
6. Peak-to-signal lag analysis
7. Tax / slippage stress scenarios

The file scripts/fixed_params_cpcv.py runs combinatorial purged cross-validation (45 alternative paths with 21-day embargo) on the fixed EMA(5,200) rotation. Result: the rotation's CPCV distribution sits to the right of QQQ B&H's in every percentile, and is ~3.5× more likely to deliver >15% annualized.

Repo layout

scripts/                    # All backtest code
├── synth_tqqq.py           # 1999–2010 synthetic TQQQ reconstruction (v1)
├── synth_tqqq_v2.py        # Same, with corrected cost model + calibration
├── multi_source_rotation.py # QQQ/TQQQ rotation core logic
├── rotation_vs_cash.py     # Bear-leg variants: QQQ vs cash vs T-bills
├── compare_vs_bh.py        # Buy-and-hold comparisons
├── forward_wf_proper.py    # Causal walk-forward validation
├── fixed_params_cpcv.py    # CPCV on fixed EMA(5,200)
├── peer_review_fixes.py    # 7 robustness checks post-review
├── full_adaptive_wf.py     # Full walk-forward with adaptive params (pre-review)
├── clean_adaptive.py       # Cleaner rewrite of the adaptive walk-forward
├── retail_ma_candidates.py # Candidate scan over retail-implementable MA rules
├── low_dd_strategies.py    # Low-drawdown variants
├── low_dd_with_tqqq.py     # Low-DD with TQQQ exposure
├── qqq_multi_signal.py     # Multi-signal QQQ filter experiments
├── pure_tqqq_pyramid.py    # Pyramid scaling on pure TQQQ
├── sqqq_pyramid.py         # Short-side pyramid via SQQQ
├── wf_pyramid_correct.py   # Corrected pyramid walk-forward
├── verify_data.py          # Data-quality checks on yfinance downloads
├── verify_logic.py         # Logic sanity checks on the rotation rule
├── run.py                  # First-pass runner
└── run2.py                 # Second-pass runner
results/                    # Pickled backtest outputs
├── fixed_cpcv_results.pkl  # CPCV paths (used for fig2_cpcv_distribution.png)
├── forward_wf_results.pkl  # Walk-forward outputs
├── retail_candidates.pkl   # Retail MA candidate scan results
└── cpcv_vs_bh.pkl          # Rotation vs QQQ B&H CPCV comparison
figures/                    # Rendered figures used in the article
├── fig1_equity_curves.png  # 27-year equity curves + 4 crisis zoom panels
├── fig_rolling_dd.png      # Rolling outperformance and drawdown
├── fig2_cpcv_distribution.png # CPCV outcome distribution: rotation vs QQQ B&H
└── fig_interactive.html    # Hover-to-inspect equity curves (embedded in blog)

Running

pip install -r requirements.txt

# Main result: CPCV on the fixed EMA(5,200) rotation
python scripts/fixed_params_cpcv.py

# Headline backtest + buy-and-hold comparison
python scripts/compare_vs_bh.py

# Full robustness suite from peer review
python scripts/peer_review_fixes.py

All scripts download QQQ / TQQQ / BIL data on demand from Yahoo Finance — no API keys needed. Expect a few minutes per script; CPCV takes longest.

Notes

• Transaction costs modeled at 2.5 bps fee + 5 bps slippage per rotation.
• Pre-2010 synthetic TQQQ is calibrated on 2010–2026 overlap; residual drift is distributed evenly across the synthetic window.
• Drawdown figures use the full 27-year path. The −95.5% drawdown occurred during 2000–2002; the rotation survived it and recovered.
• This is research code, not a production trading system. Past performance does not predict future results, and leveraged ETFs carry real risk.

License

MIT — see LICENSE.