| date | mode | day |
|---|---|---|
| 2026-07-17 | leg | +5.93% |
| 2026-07-20 | leg | +4.53% |
| 2026-07-21 | leg | -2.26% |
| 2026-07-22 | leg | -4.54% |
| 2026-07-23 | leg | +0.38% |
| 2026-07-24 | leg | +3.27% |
| 2026-07-27 | leg | -2.59% |
| 2026-07-28 | leg | +0.01% |
| 2026-07-29 | leg | +5.91% |
| 2026-07-30 | leg | +3.76% |
What it does. Signal ON (QQQ MA36 > MA140, closes through yesterday): hold TQQQ overnight — resting limit at QQQ prevLow−2% intraday, else buy the 16:00 close; sell next open. Signal OFF: equal-weight defensive basket (AZO ORLY ROST SHW TJX WM YUM) at vol-targeted leverage (30% ann. target, 4× cap). ~72% of days in the leg.
Honest numbers, not the headline. Fully cost-loaded conservative floor (v1.9 hold): OOS 2018+ 32.3%/yr, Sharpe 0.78. Live v2.1 dip-entry adds ~+9%/yr gross of execution costs. Regime-blended forward expectation ≈ 15–25%/yr; max drawdown ≈ −69%, worst modeled day ≈ −26 to −32%. This is not an alpha machine — it is a survivable way to run ~3× exposure. The binding constraint is behavioral: sitting through −70% marks.
Why each parameter survives (IS 2000–17 fit, OOS 2018+ verdict). MA 36/140: middle of a flat cluster — alternatives (incl. 50/200) are neutral-to-worse OOS. Leverage 3×: interior optimum — Sharpe falls monotonically with leverage and CAGR peaks at 3×. Dip 2%: plateau member (0.5–2% all similar). No TP/SL: 0 of 95 realizable bracket configs improved in-sample under both intraday fill orderings — on daily data every attractive bracket is an ordering illusion; the MA filter is the stop. Vol-targeted basket: removes the −95% single-day ruin paths of fixed 4×.
Execution gate. The dip-entry edge over the hold floor survives only if all-in slippage ≤ ~1.7bp/side (~504 auction executions/yr). Measured slippage ≥2bp ⇒ run the hold mechanic instead. Slippage measurement precedes arming.
Under protocol review (not wired): deep-dip entry (−8.5% vs the 10-day-ago low, MOC fallback) — OOS Sharpe 1.20 vs 1.08 and maxDD −48% vs −69% at equal cost load, but the edge sits in ~70 crash-day fills where daily bars are least trustworthy. Backtests are not forecasts.