Research notes
Methodology, backtests, durable findings. Informal write-ups with full data tables, not polished essays. Every number is reproducible from the open-source sma200-bt harness.
Informational and educational use only. Specific portfolio compositions discussed here are research outputs, not allocation recommendations. Full disclaimer →
Durable findings
Cross-study conclusions that hold across multiple tests and regimes. Each links to its supporting evidence.
-
01
The SMA200 trend filter is real on unleveraged equity.
Over 86 years of S&P 500 data the filter adds +0.18 Sharpe over plain buy-and-hold. The benefit holds in 8 of 9 decades, with the only underperformance in the 2010s QE bull. Most durable single-result on the entire research sheet.
-
02
The SMA200 filter is a portfolio construction tool, not a single-asset rescue.
Sharpe lift from the filter scales with how much leveraged-equity vol-decay sits in the portfolio: +0.238 on a 75/25 UPRO/UGL split, +0.019 on a pure 3x equity position with no defensive. Don't expect the filter to "fix" 100% TQQQ; do expect it to make heavy-equity portfolios survivable.
Evidence: Borrow-Cost Correction: Synthetic LETF Methodology + Revi..., Portfolio Archetypes Search: Finding the Sweet Spot
-
03
Gold helps as the defensive bucket; long bonds hurt.
Across 21 years on a leveraged-equity/defensive strategy, 100% gold gives Sharpe +0.032 over plain T-bills. 100% TLT (long Treasuries, ZROZ-proxy) finishes worst on Sharpe AND worst on max drawdown. The 2022 rate shock broke the bond/equity inverse correlation that pre-2022 portfolio theory assumed.
Evidence: Defensive Bucket Comparison: Cash vs Gold vs TLT vs Combinations, Borrow-Cost Correction: Synthetic LETF Methodology + Revi...
-
04
Managed futures show promise as a third defensive option.
DBMF (only 2019+ data so far) gives +0.076 Sharpe over plain T-bills and the best single-asset max drawdown protection (-47%) in a 7y test window. Regime-dependent until longer data confirms (the window captured 2022 which was uniquely good for trend-following CTAs).
Evidence: Borrow-Cost Correction: Synthetic LETF Methodology + Revi...
-
05
Synthetic LETF backtests must model borrow cost to be trusted.
The simple L*daily − ER/252 formula used in most retail synthetic LETF backtests overstates real TQQQ by 62% over 2015-2024. Adding (L-1)*borrow_rate/252 (Testfolio's approach, matching real swap-financing mechanics) tracks real TQQQ to within 5%. This is a methodology requirement, not a refinement.
Evidence: Borrow-Cost Correction: Synthetic LETF Methodology + Revi...
-
06
Volatility decay destroys 3x leverage above ~25-30% underlying vol.
SPY at 19% vol handles 3x leverage. QQQ at 27% handles it poorly. ^SOX at 37% is destroyed by it. For high-vol underlyings, 2x with a trend filter dominates 3x with a trend filter on multi-decade windows.
Evidence: Leverage vs Volatility per Underlying: 1x vs 3x, B&H vs SMA200, Full Leverage Spectrum: 1x/2x/3x × B&H/SMA200 across 5 Un...
-
07
Heavy-gold beats heavy-equity in Sharpe terms over 25 years.
The Sharpe-optimal UPRO/UGL ratio over 2000-2026 is roughly 33% UPRO / 67% UGL (Sharpe 0.745). Common "75/25 leveraged equity/gold" intuition gives Sharpe 0.536 on the same window. Gold's risk-adjusted return over a window with two major gold bulls materially exceeded leveraged equity's.
Evidence: Portfolio Archetypes Search: Finding the Sweet Spot, Borrow-Cost Correction: Synthetic LETF Methodology + Revi...
-
08
Walk-forward parameter tuning on leveraged ETFs is statistical noise.
Train→test Sharpe correlation across ~60 parameter combinations on TQQQ/SOXL/UPRO: +0.19, -0.09, +0.09. Real predictive correlations would be 0.4-0.7. The "optimal" parameters from one decade have no predictive power for the next decade. Beware of in-sample optimization marketed as backtested edge.
Evidence: LRS Family: Replication + Walk-Forward Parameter Sweep
-
09
SMA200 filter helps equities; it hurts metals.
Applying the same filter to gold reduces Sharpe from 0.71 to 0.59 over 25 years; on silver from 0.49 to 0.34. Metals have slower bull/bear cycles and choppy ranges that produce too many false signals. Apply trend filtering to equity sleeves, not metals sleeves.
Evidence: Full Leverage Spectrum: 1x/2x/3x × B&H/SMA200 across 5 Un...
-
10
Rebalance frequency matters less than asset selection.
Daily and yearly rebalancing tie for best Sharpe on a tested 9-asset leveraged portfolio (~0.945); middle frequencies cluster at 0.91-0.93. For tax efficiency in taxable accounts, yearly wins. Don't over-optimize this dial.
Evidence: Rebalance Frequency Study on a 9-Asset Leveraged Portfolio
All studies
-
Borrow-Cost Correction: Synthetic LETF Methodology + Revised Numbers
The simple L*daily − ER/252 synthetic LETF formula overstates real TQQQ by 62%. Adding the (L-1)*borrow_rate/252 term tracks real TQQQ within 5%. Walk through the bug, the fix, and revised numbers for synthetic TQQQ 1999-2026 + defensive bucket comparison.
-
LRS Family: Replication + Walk-Forward Parameter Sweep
Replicates the Testfolio LRS-style trend strategies on TQQQ, SOXL, UPRO with walk-forward parameter sweeps. The brutal finding: train-test Sharpe correlation is near zero, meaning 'optimal' parameters don't predict next-decade performance.
Reproducibility
All synthetic LETF return series in these studies come from sma200-bt, an open-source Python module calibrated against real TQQQ to within 5%. Test the numbers yourself:
pip install sma200-btReal-fund data comes from yfinance. Underlying indices and macroeconomic series are public; no proprietary data is used anywhere in these studies.