Research note · James Bowmar

Leveraged ETF "Decay" Is Not What You Think It Is

Six leveraged pairs, 16–18 years each, separating the compounding effect from the financing cost — because the usual argument conflates them, which is why it never resolves.

Both sides of this argument are wrong in the same way.

The first camp says 3x ETFs bleed value over time — volatility decay eats them, so they're day-trading instruments only. The second camp points at TQQQ, up more than 36,000% since inception, and calls decay a myth.

Both are treating a path-dependent compounding effect as if it were a constant fee. It isn't. It's a function of realized volatility over your holding period, it can run in either direction, and it is entirely separate from the cost that does behave like a fee — which almost nobody quotes.

Here's the data.

TL;DR — the three findings, for the reader who won't sit through the tables:

• The funds track their daily multiple almost perfectly — realized beta 2.96–2.99 against a 3.0 target. Whatever happens over long horizons is compounding arithmetic, not mismanagement.
• The compounding effect people call "decay" is not a fee. Its sign is set by realized volatility: below roughly 14% underlying vol the 3× funds beat their naive multiple, above it the effect turns sharply destructive. TQQQ returned 6.7× its naive prediction; TNA — same leverage, same era — returned less than half of its.
• The real fee is the financing: 6–8% a year for the 3× funds against a ~0.9% stated expense ratio. Constant, unavoidable, and missing from almost every discussion of these products.

Method

Six leveraged/underlying pairs, daily adjusted closes from each fund's inception through August 2026 (16–18 years each), split into three quantities:

The two gaps answer two different questions:

• naive → mechanism is the compounding/path effect. This is the thing people call "decay." It cuts both ways.
• mechanism → actual is fees, financing, and tracking error. This only ever costs you.

Separating them is the whole point. Conflating them is why the argument never resolves.

Finding 1: daily tracking is nearly perfect

Regressing daily leveraged returns on daily underlying returns:

These funds do exactly what their prospectus says, every single day, to within a few basis points. Whatever happens over longer horizons is not a tracking failure. The fund is not broken. The arithmetic of compounding is just not what people assume.

Finding 2: compounding is not a tax — it's a coin that lands both ways

TQQQ returned 6.7× what the naive 3x calculation predicts. If decay were a fee, that's impossible.

Now look four rows down. TNA — same 3x leverage, same 17-year window, same broad US equity market — returned less than half the naive 3x figure. 920% against a predicted 1,979%.

Same structure, same era, opposite sign. Any theory that says "3x ETFs decay" has to explain TQQQ, and any theory that says "decay is a myth" has to explain TNA.

The annualized view makes TNA's failure vivid:

TNA took 3x the leverage and an 88% drawdown to earn 14.1% a year against IWM's 12.2%. Essentially the same return as just holding the index, with a catastrophically worse path. That is the actual risk of these products — not that they go to zero, but that they hand you index-like returns for triple-digit-volatility pain.

Finding 3: realized volatility is the switch

Here's the mechanism. The theoretical annual drag of L-times leverage relative to L × the underlying's geometric return is:

drag ≈ ½ · L · (L−1) · σ²

For 3x, that's 3σ². At 15% underlying vol it's 6.8%/yr. At 30% vol it's 27%/yr. It scales with the square of volatility, which is why the effect feels like it appears out of nowhere — it's negligible in calm markets and overwhelming in violent ones.

Sorting rolling 252-day windows into volatility quintiles and measuring the edge (actual leveraged return minus naive Lx):

TQQQ

SOXL

UPRO

Monotonic in every fund tested. In the lowest volatility quintile, leverage beats the naive multiple most of the time — compounding works in your favor when the path is smooth and upward. In the top quintile it loses by 19 to 52 points, and does so 83–98% of the time.

There is no fixed decay rate. There's a volatility threshold, somewhere around 14% realized for the 3x funds, below which the effect is positive and above which it turns sharply negative.

Finding 4: the cost nobody quotes is the financing

Now separate out the part that is a fee — the gap between the frictionless mechanism and what the fund actually delivered:

Roughly 6–8% per year for the 3x funds, 3.5–4% for the 2x. That is an order of magnitude larger than the ~0.9% expense ratio these funds advertise, and it's the piece that never appears in the marketing.

It isn't mysterious. A 3x fund holds 3x notional against 1x of your cash, so it borrows 2x — and pays the going rate on it. Expense ratio plus financing on double your capital plus swap spreads lands almost exactly where the data says: 6–8%.

Unlike volatility drag, this one is constant, unavoidable, and charged whether the market is calm or violent. The steady cost of leveraged ETFs is the borrowing, not the decay. Everyone argues about the wrong one.

Finding 5: most holding periods lose

Fraction of rolling windows where the leveraged fund beat L × the underlying:

The typical outcome, at every horizon, in every fund, is underperforming the naive multiple. TQQQ's spectacular headline number comes from a minority of windows — the right tail is enormous but it is a tail. Median 252-day edge is negative for all six funds.

This is the distribution people are actually buying, and it is heavily right-skewed: usually a bit worse than expected, occasionally far better.

Finding 6: the drawdowns are survivable only on paper

Every fund spent roughly 3,700–4,200 sessions — 15 to 17 years of trading days, the large majority of its life — below a previous high-water mark. TQQQ's 43%/yr return required sitting through an 81.7% drawdown and a single day down 34.5%.

The backtest holds through that. Almost no human does, and anyone using margin or facing a liquidity need is forced out at the bottom regardless of conviction.

Finding 7: the result survives a common window

The pairs have different inception dates, so the headline table compares funds over different windows. Worth checking — but the imbalance runs opposite to the obvious worry. TNA (Nov 2008) is fifteen months older than TQQQ (Feb 2010), and its extra window contains the March 2009 bottom, the single best stretch for leveraged long equity in the sample.

Re-running every pair from the latest inception in the set — SOXL, March 2010 — so all six share one identical window:

Stripping TNA's head start makes it worse, not better — 0.36× against 0.46× on the full sample. The unequal windows were flattering it.

The common window also exposes the driver cleanly. Sorted by the underlying's return-to-volatility ratio, TNA is the only fund below 0.80 — it sits at 0.50 — and the only one where compounding hurt. Within a matched underlying, more leverage amplifies more: TQQQ 5.75× against QLD's 2.86× on the same QQQ, UPRO 3.00× against SSO's 1.87× on the same SPY.

What this actually implies

The decision rule falls out of the arithmetic. Leverage is additive to long-run return when

L · μ  >  ½ · L · (L−1) · σ²  +  financing

Run it on the two extremes in this sample:

• QQQ: μ = 19.6%, σ = 20.7% → drag 12.8% + ~7% financing ≈ 19.8% against 58.8% of levered expected return. Comfortably positive. TQQQ worked.
• IWM: μ = 12.2%, σ = 23.1% → drag 16.0% + ~6% financing ≈ 22% against 36.6%. Barely clears, and the margin is inside the error bars. TNA delivered index-like returns for triple the pain.

The relevant variable is the underlying's return-to-volatility ratio, not its return. Leverage amplifies Sharpe, roughly — apply it to a high-Sharpe underlying and it compounds beautifully; apply it to a mediocre one and the drag eats the entire premium. QQQ at 0.95 return/vol worked. IWM at 0.53 did not.

Three concrete takeaways:

1. "Don't hold leveraged ETFs overnight" is wrong, and so is "TQQQ is a free lunch." The horizon isn't the variable. The underlying's Sharpe and the realized volatility regime are.
2. The 6–8% annual financing cost is the real, permanent drag and it should be in every discussion of these products. It isn't.
3. 2x is not "less 3x." Drag scales as L(L−1): going from 2x to 3x triples the volatility penalty while only adding 50% more exposure. SSO's cost gap is half of UPRO's for two-thirds the leverage. The risk-adjusted sweet spot in this sample is 2x, not 3x — which is the opposite of where retail flow goes.

Limitations

Stated plainly, because a backtest that doesn't list these isn't worth reading:

• Survivorship bias. These are six funds that survived. Multiple 3x products launched in this period closed after catastrophic losses (notably in energy and volatility), and excluding them biases every number here optimistically. This is the single largest caveat.
• One regime. The sample runs from the 2009 low through August 2026 — predominantly a bull market in US equities with an extended zero-rate period. The financing cost estimate in particular is understated relative to a normal-rate environment, since much of the sample had near-zero short rates.
• Overlapping windows. Rolling windows are sampled every 5 sessions, so observations are highly autocorrelated. The window counts (n ≈ 700–880) dramatically overstate the independent sample size; treat the percentages as descriptive, not as inputs to a significance test.
• Adjusted closes assume distributions are reinvested with no tax or slippage. Real after-tax returns in a taxable account are meaningfully lower.
• No entry/exit rules. This measures the products' mechanical behavior, not a tradeable strategy. The vol-quintile result is measured with volatility known during the window — using it prospectively requires forecasting realized vol, which is a different and much harder problem.

Reproducing this

Full source: backtest_levetf.R — R, 288 lines, pulls adjusted daily bars from Yahoo, no API key or paid data required. Every figure on this page is inline SVG generated from the script's CSV output — no charting library, no JavaScript. Runs in under a minute. Available on request; email james@adjustedclose.com and I'll send it.

A companion piece, the regime monitor, turns Finding 3 into a standing question: given where realized volatility actually sits today, is leverage additive or destructive in each of these pairs right now?

I build data tooling for people who trade — backtests, portfolio analytics, and automated reporting in R and Python, through Adjusted Close. If you have a recurring chart pack, a strategy that needs documenting honestly, or a reporting job someone does by hand every week, that's the work: james@adjustedclose.com.

Nothing here is investment advice. It is a description of how these products have behaved historically, not a recommendation to buy or sell any of them.