Timing leveraged ETFs with moving averages

Can a moving average measured on the unleveraged index tell you when to hold its 3× cousin? Three pairs, 158,404 entry/exit combinations each, two execution assumptions, data through 2026-08-24.

Research only — no live trading, no recommendation.

1. The experiment¶

The rule under test has exactly two moving averages and no discretion. Buy the 3× ETF when the signal ETF closes above an entry MA; sell to cash when it closes below a separate exit MA.

        signal close > entry MA
   ┌──────────────────────────────────┐
   │                                  ▼
┌──────┐                        ┌───────────┐
│ CASH │                        │ LONG 3x   │
│ ^IRX │                        │ ETF       │
└──────┘                        └───────────┘
   ▲                                  │
   └──────────────────────────────────┘
        signal close < exit MA

Everything below is reported under both fill assumptions, because that gap turns out to be the most informative robustness test in the study: a rule whose profit evaporates on a one-day delay was harvesting one-day momentum, not trend.

2. What the whole search space looks like¶

Before any single rule, the shape of the parameter space. Each dot is one entry/exit combination — its drawdown against its final multiple.

Three facts fall straight out of the cloud:

• On the index pairs almost everything is profitable, and nothing is safe. The left wall sits at ~39% drawdown for TQQQ and ~40% for SPXL. No choice of moving averages gets below it, because one bad week in a 3× ETF is ~20% by itself.
• XBI→LABU is a different animal. The cloud sits low and far right: the best drawdown available anywhere in the grid is 87%.
• Buy & hold is inside the cloud, not below it. Timing buys drawdown reduction, and mostly pays for it in return.

The same information as a map of the parameter space shows where the good rules live:

The vertical structure is the story. Results vary far more with the exit window than with the entry window; the top of each map (slow exits) is bright across nearly all entries.

3. Exit speed is the dominant parameter¶

Plotting the best result reachable under a ceiling on the exit window makes the dependence explicit:

For TQQQ the curve is still climbing at window 200. Forcing the position out before the 120-day average — the "sold before SMA120" constraint — cuts the best achievable result from 661× to 136×. SPXL gives up only about a quarter of its return under the same constraint, and LABU is indifferent because it has nothing to give up.

The mechanism is simple: a 3× ETF's daily noise breaches a fast average constantly, so fast exits sell into shakeouts and re-buy higher.

4. The three pairs¶

Benchmarks over the same windows: QQQ 18.7× (−35% DD), TQQQ 335.7× (−81.7%); SPY 11.0× (−34%), SPXL 86.8× (−77%); XBI 2.13×, LABU 0.17×.

The "risk-first" rows are the ones worth attention: they are the highest-return rules whose drawdown and return survive under both fills. That constraint eliminates most of the leaderboard.

Drawdown, which is what the rules are actually buying:

5. Execution assumption is a first-class result¶

The top 12 rules per pair, each shown twice — filled at the signal close, and filled one day later:

Several leaders keep barely a tenth of their return on a one-day delay. The survivors — the slow-exit, low-turnover rules — barely move. When choosing between two rules with similar backtest returns, the one whose bars match is the one that was measuring a trend.

6. The SMA22 / SMA44 rule¶

The earlier QQQ research settled on SMA22 → SMA44. This study confirms it inside that study's search space (SMA only, windows ≤ 50): it is #1 there by Calmar, and 99.4th percentile by Calmar across the full 158k grid. Widening the space is what changes the answer.

QQQ→TQQQ, SMA22 → SMA44: 85.1× full history, 30.9%/yr, 40.4% max drawdown, 416 trades. $10k from 2018 → $209,267 same close, $59,914 next close.

Two caveats outrank the headline:

1. Its 40% drawdown is not real. Delay the fill one day and the drawdown becomes 65.5% and the return 19×. EMA156 → SMA161 holds ~150× and 48% on either fill.
2. Its shortfall is concentrated before 2016.

From 2010-02 to 2016-08 the rule made 2.8× while TQQQ made 11.8×; from 2016-08 to 2026-08 it made 29.4× against TQQQ's 27.9×. In the earlier era it sat in cash for 490 days during which TQQQ rose 3.44×: the exit fired ~97 times and 49% of those exits re-bought at a higher price, a median of 2 days later.

It is a drawdown-control rule that costs return, measured over a period whose first half had no drawdowns to control. It does not transfer: on SPY→SPXL it makes 13.7× at 59% drawdown (34th percentile, 6× worse than holding SPXL), and on XBI→LABU it loses money in every framing.

7. Hitting a 20% or 25% drawdown target¶

No fully-invested rule can do it — the left wall of the cloud is at 39–40%. Reaching 20–25% is a sizing problem: hold a fraction of the account in the 3× ETF and the rest in T-bills.

Both hold under next-close fills (11.5× / 19.5× / 8.0× / 11.5×) — that filter is what eliminated most candidates. The honest benchmark is not buy & hold but sized buy & hold at the same drawdown: 4.4× (QQQ) and 3.4× (SPY) at the 20% cap, so the timing rule is worth roughly 2.5× that.

Worth stating plainly: 35% of an account in a 3× ETF is 1.05× effective exposure, with 65% in T-bills. At a 20% drawdown budget you are running slightly-levered beta, and the leveraged ETF is merely the capital-efficient way to hold it.

8. Risk overlays¶

Six overlays on top of the drawdown-controlled rule for each index pair:

Why the stop-loss fails: these drawdowns are slow grinds that the MA exit is already handling. A 20% stop fires mid-shakeout and the re-entry signal then arrives at a higher price, converting drawdown into whipsaw instead of removing it. Sizing down is strictly better — a 35% position has a 20% drawdown by construction, with no timing skill required.

9. Constraining the exit to ≤ SMA120¶

Requiring the position to be out before the 120-day average caps the exit window at 120 and removes the entire slow-exit family:

Under the same constraint, the drawdown-capped frontier (holding under both fills):

That is 25–35% of the return given up at the same risk. Note that the constrained QQQ winners pair a fast entry (SMA6–7) with the ~EMA119 exit: once the exit is forced fast, entering early is how the lost upside is recovered.

10. Bullish MA alignment as an entry filter¶

An additional entry condition — fast MA > mid MA > slow MA on the signal ETF, e.g. SMA5 > SMA20 > SMA50:

• XBI→LABU: it is what makes the strategy work. The median across all 158k combinations rises from 0.33× to 1.32× (11% → 69% profitable) and, critically, it survives next-close fills (1.92× vs 0.57×). Drawdowns stay near 50%.
• QQQ→TQQQ: it hurts (median 22.6× filtered vs 35.1× unfiltered). It sits out too much of a market that mostly went up.

11. Out of sample¶

Every leaderboard above is in-sample. Re-picking the parameters each January from only the data available at the time is the honest number:

• QQQ→TQQQ: 20.4× against QQQ's 11.96× — but nothing beats holding TQQQ (131×).
• SPY→SPXL: 7.4× against SPY's 7.7× and SPXL's 48.5×. The specific windows are what was worth money, not the method of finding them.
• XBI→LABU with the alignment filter fixed: 2.19× against XBI's 1.81× — the only out-of-sample win in the study.

12. Conclusions¶

1. The edge is real but small, and it is a risk-control edge, not a return edge. On both index pairs the best drawdown-constrained rules match or beat buy & hold at ~30 points less drawdown; none reliably beats holding the 3× ETF on raw return out of sample.
2. Exit speed dominates everything else. Slow exits (150–200 days) beat fast ones by an order of magnitude on a 3× vehicle.
3. The fill assumption is a result, not a footnote. Rules that lose most of their return on a one-day delay — SMA22/44 among them — should be read as fitted.
4. Drawdown targets below ~40% are a sizing problem, not a signal problem. Stop-losses do not deliver them; position size does, mechanically.
5. XBI→LABU is not risk-manageable this way. The best drawdown available anywhere in its grid is 87%; LABU decays too fast between XBI's trends. Only the alignment-filtered version is worth anything, and it still carries ~50% drawdowns.

13. Reproducing¶

Every generated data point lives in a local SQLite database (research.db): prices, three 158k-row grids under both fills, alignment searches, walk-forward folds, equity curves, yearly returns, and the sizing / vol-target / overlay frontiers — ~2.9M rows across 29 tables, keyed by pair. Key tables: ma_grid, ma_grid_lag1, ma_top20, ma_risk_frontier, ma_sized, ma_voltgt, ma_overlay, ma_exit_cap, ma_sized_exitcap, ma_voltgt_exitcap. Figures on this page are generated from those tables.

14. Disclaimer¶

Backtests on a single realised price history, with in-sample parameter selection in most sections. Nothing here is a recommendation to trade.