Study based on data up to October 2026. The simulator and the asset pages are updated on the 1st of every month.
Leverage only pays when the trend is strong
The optimal (Kelly) leverage, L* = (μ − rf) / σ², changes a lot from one era to the next: 4.1 for the Nasdaq-100 in 1985-99, 0.1 in 2000-12 and 4.4 in 2013-26. For the S&P 500 it ranges from 0.1 (1970-81) to 8.7 (1950-69). A daily-reset 2x only beats 1x when L* is above roughly 1.5, so whether leverage works depends on an era nobody knows in advance.
Out of sample, 2x does not win in general
Across the 53 assets in the universe (28 indices and 25 sectors, each with its own total return and local short-term rate, against a 1x with 0.20% in fees), 2x beat 1x in 26 (median −0.1 pp a year; 13 of 28 indices and 13 of 25 sectors) and 3x in only 10 (median −3.8 pp a year).
- Where 2x won: the US (S&P 500, Nasdaq-100, Dow, total market), Switzerland, Stoxx 600 and MSCI World since 2012.
- Where it lost: Japan, Hong Kong, Korea, China, India, Brazil, Mexico, DAX, CAC 40, Russell 2000, energy, financials and semiconductors.
- Sensitivity: with dividends one point lower, the average gap for indices goes from −1.0 to −2.1 pp a year; with interest rates halved, to +1.5.
Drawdowns are the price
Over each index’s full history (Nasdaq-100 1985-2026, S&P 500 1928-2026). The longest stretch is the longest period without recovering the previous high.
| Index | Leverage | CAGR | Max drawdown | Days >50% below peak | Days >80% below peak | Longest stretch |
|---|---|---|---|---|---|---|
| Nasdaq-100 | 1x | 15.2% | −82.9% | 24.9% | 0.4% | 14.9 years |
| Nasdaq-100 | 2x | 18.3% | −98.9% | 50.1% | 38.7% | 20.4 years |
| Nasdaq-100 | 3x | 14.4% | −99.98% | 72.8% | 62.0% | 26.5 years (ongoing) |
| S&P 500 | 1x | 10.1% | −83.7% | 4.6% | 0.2% | 15.2 years |
| S&P 500 | 2x | 11.2% | −98.5% | 32.0% | 14.2% | 22.1 years |
| S&P 500 | 3x | 9.0% | −99.91% | 58.7% | 32.6% | 28.9 years |
On the Nasdaq-100, 2x earned more than 1x over the whole history but spent half of its days more than 50% below its high; 3x earned less than 1x.
The entry date decides
Starting at the end of December 1999 on the Nasdaq-100, 10,000 became 8,433 after 20 years with 2x and 26,786 with 1x. 2x beat 1x in 27% of 20-year entry dates, and it is a step, not a spread: 0% for entries from October 1985 to January 2001 and 100% from February 2001 to October 2006. That is about two independent stretches. On the S&P 500, 2x won in 58% of entry dates (median ×1.22).
Gaussian simulations understate the extreme days
Days beyond three standard deviations are 0.27% in a normal distribution and 1.67% in the real S&P 500 since 1928. A Markov-switching multifractal model (MSM) reproduces that frequency (1.69%) but falls short on the largest moves (kurtosis 5.8 against 17.3). With the MSM, a standalone 2x loses more than 90% from a peak within 20 years in 20% of cases. In historical 20-year windows that figure is about 4.7% for the S&P 500 and about 74% for the Nasdaq-100, but there are so few independent stretches that history can neither validate nor refute the model.
A small leveraged slice in a 50/50 portfolio
Reference case: a 50% SPY / 50% QQQ portfolio over 20 years with 10,000 €, simulated with the MSM, adding a leveraged slice (QLD, TQQQ or SSO) of different sizes and rebalancing rules: 108 combinations in total.
- None of the 108 combinations lowers the probability of a portfolio fall of more than 50% or 70% compared with the 50/50 alone.
- None improves the median and the unfavourable case (p10) at the same time.
- With the Nasdaq’s historical drift, 10% in QLD lifts the median from 98,108 € to 110,927 €, and the probability of a fall above 70% from 13.4% to 19.8%.
- With the S&P’s drift, the median stays put or drops (62,531 € alone; 61,895 € with 10% in QLD; 58,197 € with 30%) and the p10 falls from 15,340 € to 12,215 € (10% in QLD).
- A 2x on the S&P 500 (SSO) is the most stable option: 10% in SSO gives a median of 66,540 € (p10 14,653 €).
- With a 4% risk-free rate instead of the historical 3.25%, the medians of portfolios with a leveraged slice fall by 0.5% to 6.5%. The conclusions do not change.
Real leveraged ETFs
- The simulated replica tracks the leveraged ETFs you can actually buy: across the 34 of 46 with reliable data (US, Europe, Japan and Hong Kong), the median annual gap with the real fund is −0.04 pp.
- What they really cost on top of financing: about 1% a year in the US (QLD 1.2%, TQQQ 1.4%, SSO 1.0%) and about 1.2% in Europe, a bit more than the published expense ratio.
- For a euro investor, currency handling matters: LQQ and DBPG leverage the index in dollars and leave the currency at 1x; CL2 leverages the currency too.
- Two model biases found with real ETFs: leveraged FTSE 100 funds return 2-2.5 pp a year less than their replica, and the 2x Taiex about 4 pp more. Conclusions on those two indices are less reliable.
What is fragile
- The edge of 2x on the Nasdaq-100 depends on the Nasdaq repeating its drift since 1985: choosing it in 1985 means choosing a known winner (selection bias).
- Overlapping historical windows give few independent observations: about 2 windows of 20 years on the Nasdaq-100 and about 5 on the S&P 500.
- Calibration: fitting the costs only with data up to 2018 overshoots out of sample (2019-2026) by 0.5 to 1.5 pp a year.
Limits
No taxes or rebalancing costs, which can change the result, especially with frequent rebalancing. The MSM understates kurtosis and long memory and uses a constant risk-free rate. The same financing cost is used for every asset and currency. Simulated series are not spliced with real ETF prices. Historical and simulated analysis for educational purposes; not financial advice.