TQQQ/SQQQ SMA-grid backtest

1. framing

The strategy is a three-state rotation driven by a single signal instrument (QQQ) against four simple-moving-average windows. On every daily bar the state machine is

CASH        → LONG_TQQQ    when QQQ > SMA(buy_T)
LONG_TQQQ   → CASH         when QQQ < SMA(sell_T)
CASH        → SHORT_SQQQ   when QQQ < SMA(buy_S)
SHORT_SQQQ  → CASH         when QQQ > SMA(sell_S)

Phase 1 of this sweep enforces sell_T ≥ buy_T and sell_S ≥ buy_S — trend-following only. Each SMA window is drawn from {1..60}, so the constrained grid has 1830 × 1830 = 3,348,900 combos.

Sweep wall-clock: 0.6 min · ~95,522 backtests/sec. Slippage: 5.0 bp per state transition; $0 commission; adjusted-close prices.

This document tells you what the search surface actually looks like — not just the single best combo. The histogram in §5 is the antidote to the table in §2.

2. top-50 winners

Sorted by final equity multiple. Each row is one DISTINCT equity curve — combos sharing a curve (e.g. dormant short-leg parameters varying freely when buy_S=1 structurally never fires) are collapsed by a deterministic equity-curve fingerprint. n_eq is the size of the equivalence class; type classifies which legs actually traded; dormant parameters render as -. Of the 3,348,900 grid combos, 2,972,310 produce distinct equity curves.

delta is the (test − train) / train ratio from the walk-forward in §7 — green = generalizes up, red = generalizes down.

3. buy-and-hold baselines

The four passive baselines a strategy must beat to be interesting:

4. heatmaps

Two 2-D slices through the 4-D grid. The winner combo is (22, 44, 1, 1).

4a. (buy_T, sell_T) with short-leg fixed at the winner

4b. (buy_S, sell_S) with long-leg fixed at the winner

5. distribution of outcomes

Where do the buy-and-hold baselines sit in the 3.35M-combo final-value distribution? Most of the mass is mediocre; the long right tail is where the headline numbers live.

6. equity curves

Top-5 strategies vs the three buy-and-hold baselines. Log scale on the y-axis so the leveraged baselines don't flatten the rest.

7. walk-forward delta

For each of the top-100 DEDUPED combos by full-sample final value (one representative per equity-curve equivalence class), we re-ran the backtest on train (2010-02 → 2018-12) and test (2019-01 → today). Points hugging the diagonal generalize; points dropping below it overfit.

Of the 100 combos that beat cash in-sample, 100 also beat cash out-of-sample (100.0%).

8. honest discussion

9. reproducibility

To reproduce: uv run rainier sma-sweep --phase 1