TQQQ Trend-Timing — Tuning the QQQ MA-Gate Band

A parameter sweep over the (fast-in, slow-out) moving-average hysteresis band for the TQQQ trend-timing strategy: hold TQQQ (3× daily-leveraged QQQ) when QQQ is in an uptrend by a fast/slow MA gate, otherwise sit in cash. This is the TQQQ analog of the QU100 gate-tuning study — we search 124 (fast, slow) pairs to ask whether any band beats the reference 22/55 and 22/44, and whether the "best" band sits on a robust plateau or an isolated overfit spike. TQQQ is the only held instrument — no QU100 basket, no N dimension. Start $100,000. Gross of trading costs (0 bp), cash earns 0%.

Window: 2020-10-01 → 2026-06-04 · 1,402 ranking-days (~5.7 yrs) on the QU100 ranking-day calendar · QQQ & TQQQ = yfinance adjusted close (auto_adjust=False; folds in TQQQ's Jan-2022 split). QQQ MAs seeded from 2020-06 so all SMAs are valid on the first ranking-day. Fills at same-day close (no look-ahead).

The gate (evaluated each ranking-day t on QQQ MAs through day t; carry an invested flag, start flat). For a pair (fast, slow), fast < slow:
  • QQQ_close ≥ QQQ_SMA(fast) → RISK-ON → hold 100% TQQQ
  • QQQ_close < QQQ_SMA(slow) → RISK-OFF → sell, hold cash (0%)
  • SMA(slow) ≤ close < SMA(fast) (the band) → HOLD previous state
The hysteresis band suppresses whipsaw: you flip ON above the fast MA and OFF below the slow MA. A wider band (small fast, large slow) flips rarely but exits a downturn late; a narrow band exits sooner but whipsaws more. TQQQ is 3× daily leveraged — in choppy/down markets its compounding decays well below 3× the index, which is exactly the risk the gate is built to dodge.

(22, 200)

Global best band (by Sharpe & return)

1.175

Sharpe (best)

+1305%

Total return (best)

-42.2%

Max drawdown (best)

86%

Time in market (best)

Flips (13 round-trips)

Verdict. Tuning does beat 22/55 and 22/44 on this sample — but read the win with extreme suspicion. The global best by both Sharpe and return is (22, 200): Sharpe 1.175 (vs 1.05 for 22/55, 1.10 for 22/44) and +1305% total return (vs +716% / +807%) — but it wins by barely ever exiting (time-in-market 86%, only 26 flips): a slow-200 exit leg that tripped once, in 2022, and otherwise rode the leverage. That makes it a fragile corner spike, not a plateau — only 2 cells sit within 0.05 Sharpe of it, both at slow=200. This is the most overfit-prone study run so far: 3× leverage × a single 2022-crash sample × a 124-cell search massively inflates in-sample Sharpe — the "best" band is simply the one that best-timed one crash while otherwise staying long.

The defensible result is a different, broader region: the fast ∈ {20, 22, 30} × slow ∈ {40, 44, 55, 65} block all clusters at Sharpe ≈ 1.00–1.10 (see the heatmap — it is a genuine contiguous plateau, not a point). 22/44 (Sharpe 1.10, +807%) sits right in that plateau and remains the practical pick — it captures essentially all the robust edge with a sensible exit, while the slow-200 corner is a one-crash artifact you should not trust out-of-sample. Fast bands (fast ≤ 10) are whipsaw-prone and cost-sensitive: they flip 80–130+ times and deliver the worst Sharpes in the grid — on a 3× instrument every extra flip is amplified cost drag. Net: do not chase (22, 200); stay on the 22/44–22/65 plateau and validate costed + out-of-sample before trusting any of it.

1 · Sharpe heatmap (green = high; ✦ best, 44/55/30·40 reference cells marked)

fast \ slow	5	8	10	13	15	20	22	30	40	44	55	65	80	100	150	200
2	0.34	0.25	0.25	0.20	0.30	0.46	0.45	0.44	0.60	0.64	0.59	0.53	0.55	0.67	0.73	0.77
3	0.42	0.35	0.37	0.31	0.42	0.60	0.56	0.47	0.69	0.75	0.71	0.61	0.67	0.75	0.89	0.93
5		0.55	0.54	0.49	0.56	0.70	0.65	0.55	0.75	0.83	0.78	0.71	0.75	0.73	0.85	0.89
8			0.58	0.56	0.63	0.77	0.78	0.57	0.78	0.81	0.76	0.71	0.69	0.70	0.84	0.89
10				0.47	0.55	0.64	0.66	0.51	0.76	0.79	0.75	0.74	0.66	0.66	0.84	0.87
13					0.66	0.74	0.74	0.62	0.88	0.92	0.87	0.82	0.78	0.70	0.87	0.92
15						0.58	0.62	0.53	0.79	0.83	0.78	0.77	0.72	0.67	0.84	0.91
20							0.92	0.83	1.04	1.09	1.02	1.03	0.97	0.89	1.07	1.16
22								0.89	1.08	1.10⁴⁴	1.05⁵⁵	1.08	1.01	0.93	1.10	1.18
30									1.00^30/40	0.96	1.01	1.00	0.98	0.85	1.01	1.04

Rows = fast (entry) leg, columns = slow (exit) leg; blank lower-left = invalid (fast ≥ slow). Two high zones stand out: a broad plateau at fast 20–30 × slow 40–65 (Sharpe ~1.0–1.1) and an isolated bright corner at slow = 200, fast 20–22 (Sharpe 1.16–1.18). The corner is the global max but stands alone; the plateau is wide and robust. The whole bottom-left (fast ≤ 10) is dim — fast entry legs whipsaw and underperform.

2 · Total-return heatmap (same grid)

fast \ slow	5	8	10	13	15	20	22	30	40	44	55	65	80	100	150	200
2	26%	-4%	-6%	-19%	9%	74%	71%	63%	165%	201%	161%	113%	132%	233%	305%	352%
3	60%	29%	38%	14%	58%	163%	135%	81%	251%	316%	269%	174%	229%	330%	552%	660%
5		124%	121%	91%	136%	248%	207%	128%	319%	431%	363%	273%	316%	303%	475%	564%
8			140%	130%	179%	309%	315%	140%	331%	369%	310%	255%	238%	252%	448%	547%
10				84%	125%	186%	202%	102%	303%	340%	297%	288%	211%	212%	440%	497%
13					195%	267%	271%	175%	453%	523%	460%	378%	333%	256%	493%	604%
15						145%	172%	114%	330%	384%	331%	313%	268%	220%	447%	581%
20							441%	344%	672%	810%	669%	708%	611%	483%	934%	1280%
22								414%	732%	807%⁴⁴	716%⁵⁵	779%	676%	537%	982%	1305%
30									552%^30/40	521%	605%	593%	586%	405%	732%	852%

Return peaks even more sharply at the slow=200 corner (the rarely-exiting bands keep more leveraged upside) — which is exactly why a return-max search overfits hardest here. Note the plateau bands give up raw return for the lower drawdown and steadier ride that the gate is supposed to buy.

3 · Top-10 bands by Sharpe

(fast, slow)	Sharpe	Total return	CAGR	Max DD	Ann vol	TiM	Flips
(22, 200)best	1.175	+1305.1%	+59.3%	-42.2%	52.0%	86%	26
(20, 200)	1.161	+1279.8%	+58.8%	-52.8%	52.6%	86%	30
(22, 44)	1.102	+807.3%	+47.5%	-39.8%	45.3%	72%	76
(22, 150)	1.096	+982.0%	+52.2%	-43.5%	50.9%	83%	36
(20, 44)	1.091	+809.7%	+47.6%	-50.9%	46.2%	73%	80
(22, 40)	1.077	+732.3%	+45.3%	-45.1%	44.7%	71%	84
(22, 65)	1.076	+778.9%	+46.7%	-41.1%	46.4%	74%	62
(20, 150)	1.073	+933.7%	+50.9%	-49.1%	51.6%	83%	40
(22, 55)	1.049	+715.6%	+44.8%	-45.3%	46.2%	73%	66
(30, 200)	1.043	+851.8%	+48.8%	-58.7%	51.7%	85%	26
TQQQ-gated (22, 44) prior pick	1.102	+807.3%	+47.5%	-39.8%	45.3%	72%	76
TQQQ-gated (22, 55) original	1.049	+715.6%	+44.8%	-45.3%	46.2%	73%	66
TQQQ buy & hold	0.780	+420.9%	+33.8%	-81.7%	66.4%	100%	0
QQQ buy & hold	0.916	+171.6%	+19.3%	-35.1%	22.3%	100%	0
SPY buy & hold	1.031	+142.5%	+16.9%	-24.5%	16.8%	100%	0

The Sharpe leaderboard is dominated by the slow=200 corner (top 2) then the 20–22 × 40–65 plateau. Reference cells: 22/44 ranks #3, 22/55 ranks #9. Baselines (amber) and buy-and-hold shown for context.

4 · Top-10 bands by total return

(fast, slow)	Sharpe	Total return	CAGR	Max DD	Ann vol	TiM	Flips
(22, 200)best	1.175	+1305.1%	+59.3%	-42.2%	52.0%	86%	26
(20, 200)	1.161	+1279.8%	+58.8%	-52.8%	52.6%	86%	30
(22, 150)	1.096	+982.0%	+52.2%	-43.5%	50.9%	83%	36
(20, 150)	1.073	+933.7%	+50.9%	-49.1%	51.6%	83%	40
(30, 200)	1.043	+851.8%	+48.8%	-58.7%	51.7%	85%	26
(20, 44)	1.091	+809.7%	+47.6%	-50.9%	46.2%	73%	80
(22, 44)	1.102	+807.3%	+47.5%	-39.8%	45.3%	72%	76
(22, 65)	1.076	+778.9%	+46.7%	-41.1%	46.4%	74%	62
(22, 40)	1.077	+732.3%	+45.3%	-45.1%	44.7%	71%	84
(30, 150)	1.007	+732.1%	+45.3%	-55.5%	50.6%	82%	32

The return leaderboard is even more concentrated at slow=150/200 — high time-in-market keeps leveraged upside. But several of these carry deeper drawdowns (e.g. 20/200 at -52.8%, 30/200 at -58.7%) than the plateau bands, the tell that they are riding leverage, not timing it.

5 · Equity curves (log scale, $100k start)

best (22/200)gate 22/55TQQQ B&HQQQ B&H

The best (22/200) line (blue) tracks TQQQ B&H (red) almost everywhere — it barely exits — then diverges hard in 2022 when its slow-200 leg finally trips and it sits in cash through the leverage-decay blender. That single divergence is the entire edge of the corner spike; everywhere else it is buy-and-hold. The 22/55 gate (green) flattens through 2022 the same way but exits/re-enters more, sacrificing some peak for a steadier path.

6 · Robustness — plateau vs spike

Sharpe within …	# cells	slow legs spanned	fast legs spanned
0.03 of best	2	{200}	{20, 22}
0.05 of best	2	{200}	{20, 22}
0.08 of best	4	{44, 150, 200}	{20, 22}
0.10 of best	7	{40, 44, 65, 150, 200}	{20, 22}

The global best is a spike, not a plateau. Within 0.05 Sharpe of (22, 200) there are only 2 cells, both pinned at slow=200 — an isolated corner. You have to widen the tolerance to 0.10 before the broad 20–22 × 40–65 region joins in. The lesson: the slow-200 corner is fragile (one crash, one exit), but the 40–65 slow-leg plateau is genuinely contiguous and robust — neighboring cells there all score within a hair of each other, so a small mis-pick costs little. That plateau — where 22/44 and 22/55 live — is the part of this surface you can actually trust.

Methodology & caveats. Gross of costs — 0 bp, no commissions or slippage. Costing matters more here than for an unlevered basket: a leveraged trend gate flips in and out and 3× leverage amplifies every cost-drag dollar. Flip counts are shown in every table precisely so the cost sensitivity is visible — the plateau bands flip 60–85 times, the fast-leg bands 80–130+, the slow-200 corner only 26. A realistic spread+commission run would erode the high-flip bands first; the low-flip corner looks artificially clean at 0 bp. A costed re-run is mandatory before trusting any ranking. TQQQ real-world frictions: TQQQ carries ≈0.95% annual expense ratio, daily-rebalance decay, and embedded financing cost on the leverage. We hold TQQQ's actual traded adjusted close, so those frictions are already baked into the price series — the one thing NOT modeled is the cost of our own trading (the 0-bp caveat above). Single-sample overfit + search inflation: this is one 2020–2026 path dominated by the 2022 crash, searched over 124 cells on a 3× instrument. This is the most overfit-prone study in the series. A 124-cell max-Sharpe search on a single crash will surface whichever band best-timed that one event — here, a slow-200 leg that exited once. Trend-timing on leveraged ETFs is a notoriously overfit strategy class; the "best" band's edge would likely not survive a different crash shape or a choppy whipsaw market. Treat the broad plateau as the signal and the corner spike as noise. Out-of-sample and costed validation are required before treating any of this as real. Same-day-close fill convention: the gate is evaluated on QQQ MAs through day t; the position decided at close of t is held into day t+1's return (no look-ahead — today's return is governed by yesterday's decision). QQQ MAs seeded from 2020-06 so SMA2…SMA200 are all valid on the first ranking-day. Metrics: max drawdown close-to-close at ranking-day granularity; vol = daily-return std ×√252; Sharpe = mean/std ×√252 (rf=0). Sanity checks (all passed): TQQQ B&H = +420.9% / MaxDD −81.7% / Sharpe 0.78; gate 22/55 = +715.6% / −45.3% / Sharpe 1.05 / TiM 73.5%; gate 22/44 = +807.3% / −39.8% / Sharpe 1.10 / TiM 71.8% — all reproduce the reference figures exactly, confirming the efficient regime-mask matches the full backtest. Not investment advice — a research backtest.