ATLAS is a LETF strategy designed for client deployability: LETF-scale CAGR with a 60/40-compatible drawdown envelope. Backtest full-sample CAGR is 14.9% at Sharpe 0.93 over 21 years (2005-2026, including the 2008-09 GFC, the 2020 COVID crash, and the 2022 dual stock-bond bear) with max drawdown −26.8%. 0% of 2-year rolling windows end with a drawdown worse than −30% — matching 60/40 on path discipline while delivering materially higher CAGR.
Two layers. (1) A time-series momentum core, Moskowitz-Ooi-Pedersen style, taking the sign of the trailing 3-month return on SPY/QQQ/TLT/GLD and expressing the bet through the 3× LETFs UPRO, TQQQ, TMF, UGL, vol-targeted to 15% annualised with monthly rebalance. (2) A drawdown-throttle overlay that scales exposure based on how far the live strategy sits below its own 252-day peak, floor at 25%. Residual capacity sits in BIL.
Why the overlay is the invention. Every LETF strategy we tested ran a −40 to −80% drawdown at some point — that is the user-experience problem, not Sharpe. The overlay cuts the full-sample max drawdown from the base TSMOM’s −44% (on 2011-26 live data) to −27% without harming Sharpe, AND holds the cap even through the 2008-09 GFC stress test on simulated-LETF data (see below). A 500-permutation null test (21-day block shuffle of the multiplier) confirms the MDD reduction is real (p<0.001).
Pre-registered holdout. Parameters were fixed on 2011-01-01 → 2023-01-01 discovery data. On the 2023-01-01 → 2026-04-02 holdout, with zero re-tuning, ATLAS delivered Sharpe 1.07, CAGR 18.1%, MDD −20.0% — better than discovery on all three metrics, consistent with a real signal rather than an overfit.
Extended history via simulated LETFs. Real LETF history begins 2008-2010. To stress the strategy over the 2008-09 GFC we simulate pre-inception LETF daily returns via r_letf = L × r_underlying − (L−1) × fed_funds/252 − expense/252, calibrated so the first 3 years of overlap with REAL LETF returns match on mean (correlations 0.995–0.998). The simulated period (Jan 2005 → LETF inception) is therefore a conservative proxy. Headline numbers on this page blend real LETF history 2008+ with simulated history 2005-2008.
What it is not. Not a Sharpe booster — the overlay’s Sharpe lift is within sampling noise (p=0.15). Its value shows up in worst-case path, not long-run CAGR/vol. Not a tail hedge — it reacts AFTER drawdown has started, it does not pre-empt. Not tested over 100 years — 21 years contains three serious equity bears (2008-09, 2020, 2022) and a multi-decade bond-bull tail, but no high-inflation 1970s regime and no 1987-style single-day crash.
| Signal day (T−1) | At close T−1, compute 63-day log return on each underlying (SPY, QQQ, TLT, GLD). If return > 0, target the corresponding LETF (UPRO/TQQQ/TMF/UGL); if ≤ 0, target BIL for that slot. |
| Vol target | Compute 63-day realised vol of the raw target portfolio. Scale weights so ex-ante annualised vol equals 15%, capped at 3×. This is the TSMOM base. |
| DD-throttle | Track the 252-day running max of the TSMOM NAV. Current DD = NAV / peak − 1. Apply a leverage multiplier: 0 to −5% → 100%; −5% to −10% → linear 100→50%; −10% to −20% → linear 50→25%; below −20% → floor 25%. 5-day smooth. |
| Execution | Weights and multiplier set at close T−1 — execute at next-day open T. Monthly rebalance for the core (21 trading days); overlay adjusts daily. 15 bps transaction cost on |Δw|. |
| Residual | Any capacity not deployed by TSMOM or × overlay multiplier is parked in BIL (1-3 month T-bills). Gross exposure never exceeds the multiplier’s current cap. |
| Signal lag | Two-day lag throughout: signal computed at close T−1 using data through T−2. No look-ahead. Execution assumed at open T. |