Martingale on the S&P 500 and Nasdaq 100: 2.94% vs. 10.84%

With a single stock, the risk of buying more on the way down can't be reasoned away: some stocks never come back, because the company behind them goes under. A broad index has been different so far — the S&P 500 and the Nasdaq 100 have survived every crisis they've faced; that, too, is no guarantee for the future. That thought was the starting point for this test: a buy-the-dip method (a so-called martingale strategy) wasn't applied to individual stocks that might go bankrupt, but to the S&P 500 and the Nasdaq 100 themselves — through the ETFs SPY and QQQ, and, for the very long view, through the underlying price indices reaching back to 1927.

The method was fixed in advance, before a single number was calculated: 216 combinations of follow-up size, follow-up threshold, sell target, and cap were run in full, alongside one main rule named in advance as the reference point — starting capital $100,000, with realistic trading costs and interest on the money not currently invested.

The result in four numbers

• 2.94% per year — that's what the main rule of the method returned, measured on the complete $100,000 capital, well below the 10.84% per year that buy and hold achieved over the same window (1993 to 2026). $100,000 turned into $264,294.31 under the method, and into $3,164,442.67 under buy and hold.
• 0 of 216 tested rule cells beat buy and hold on return — on both instruments (SPY and QQQ), across every combination tested.
• 86% of the main rule's total return came from interest on the cash that sat in reserve instead of being invested. Without that interest, $100,000 would have grown to only $114,952.71 over 33 years.
• 262,143 base units of reserve (a base unit is the amount a round starts with and against which every follow-up purchase is measured) would have been needed under the worst of the 36 rules tested — reached by the Nasdaq 100 in the dot-com bust from the year 2000; the S&P 500 price index hit the same figure in the crisis around 1932.

Buying the dip kept the curve calm, but over time it cost a large share of the return that simple holding would have delivered.

What martingale means here

The name is borrowed from gambling: after every loss, the stake went up, in the hope of recovering the entire loss with a single win. Applied to the stock market, that meant: a round began with the purchase of one base unit — a fixed amount that every later purchase in the same round was measured against. If the price fell by a threshold since the last purchase (5% under the main rule), the next purchase doubled the size of the previous one; if the price then rose 2% above the average entry price, the entire position was sold at once, and the next round began immediately.

The base unit under the main rule started at $392.16; every follow-up purchase doubled the size of the one before it, until at the latest the eighth purchase of a round ended it (the cap, the ceiling on purchases per round). With this doubling rule and a cap of 8, the entire possible ladder fit exactly into 255 base units — precisely the $100,000 starting capital.

How this was tested

The test ran on four data series: the ETFs SPY (S&P 500, from 1993) and QQQ (Nasdaq 100, from 1999) with trading costs and interest, plus — for the very long view on reserve needs — the corresponding price indices without dividends, reaching back to 1927 and 1985 respectively. Source: own backtest 1927–2026; interest on the waiting cash tracked the 3-month yield on US Treasury bills (T-bills, short-term US government debt) from the Federal Reserve's interest-rate statistics (FRED series DTB3).

216 rule cells were tested: three follow-up sizes (×2, ×1.5, equal tranches) times three follow-up thresholds (3%, 5%, 10%) times four sell targets (1%, 2%, 3%, 5%) times three caps (6, 8, 10 purchases per round), on two instruments. One cell — SPY, doubling, 5% threshold, 2% target, cap 8 — was the main rule, named in advance so that no accidentally good result could later be declared the intended rule. A rule counted as successful if its return per year (p. a., the annual compounding rate between starting and final capital) was at least as high as buy and hold's, and its largest drawdown (MaxDD, the deepest fall from the highest point reached so far) was no larger.

The main calculation: the S&P 500 over 33 years

Over the full window from 1993 to 2026, the main rule stood at 2.94% p. a., against 10.84% p. a. for buy and hold. Its drawdown, though, was tiny: −2.95% for the method against −55.19% for buy and hold — the method never came close to that, because on average only 0.75% of the capital was actually invested in the market at any time. This continuous full-window calculation from 1993 to 2026 necessarily includes the exceptional year 2020; all of the year-by-year evaluations and averages further down, by contrast, leave it out.

Return p.a. = the annual compounding rate that would turn the starting capital into the final capital. Largest drawdown = the deepest fall from the highest account balance reached up to that point. The fair benchmark buys one base unit once and holds it; the remaining money sits in US Treasury bills throughout. "—" means the metric does not exist for that line, because it is invested throughout and has no rounds. Source: own backtest 1927–2026.

546 complete rounds ran over this period, with a round length whose median (the typical case — half the rounds were shorter, half longer) came to 10 trading days. The mean (all rounds added up and spread evenly) was higher, at 15.4 trading days, because a handful of very long rounds — up to 137 trading days — pulled it up.

Why the calm ride wasn't buying-the-dip's own doing

A drawdown of just −2.95% sounds like a clever method. But the comparison with a third, fair benchmark showed otherwise: it came mostly from the fact that almost no money ever sat in the market — not from buying the dip itself. This benchmark bought exactly one base unit and simply held it, while the rest sat earning interest on the idle cash in reserve; it reached 2.67% p. a. ($241,852.20) with a drawdown of only −0.80%. The method beat this fair benchmark by only 0.27 percentage points — while it fell 7.90 percentage points short of buy and hold. The calm curve, in other words, came from most of the money never being invested at all, not from any advantage of buying the dip itself.

Interest was the real return

Because almost all the capital sat in reserve most of the time, the result depended heavily on whether that money earned interest — a counter-calculation with no interest on the waiting cash made that clear.

The waiting cash sits in US Treasury bills with three months to maturity (FRED series DTB3). The share of interest in the return is derived: 1 − (return without interest ÷ return with interest) = 85.8%, rounded to 86% in the text. Return p.a. = the annual compounding rate that would turn the starting capital into the final capital. Largest drawdown = the deepest fall from the highest account balance reached up to that point. Source: own backtest 1927–2026.

Without interest, the return fell from 2.94% p. a. to 0.42% p. a., and the final capital from $264,294.31 to $114,952.71. That means 86% of the annual return came from interest alone, not from the trading itself.

All 216 rules in the sweep

So that the result didn't hinge on the choice of a single rule, all 216 combinations were evaluated.

The cap is the ceiling on purchases per round. Each of the 216 cells is its own rule combination, all computed at the main cost level. Median = the typical case in the middle: half the values were above it, half below. Not a single cell returned more than buy and hold, and every single one had a smaller drawdown. The best cell in each group never uses ×2. Source: own backtest 1927–2026.

Not one of the 216 cells beat buy and hold on return; none fully met the success criterion fixed in advance. Every one of the 216 cells, though, had a smaller drawdown than buy and hold. The return p. a. stood at a median (the typical value — half the cells were above it, half below) of 4.45%. The best cell of the whole sweep (Nasdaq 100, follow-up size 1.5, threshold 3%, target 1%, cap 6) reached 8.83% p. a., but it also failed to meet the criterion, because its own buy-and-hold comparison, at 10.72% p. a., was still higher.

Follow-up size: aggressive vs. gentle

72 cells per group. Median = the typical case in the middle: half the values were above it, half below. Mean (average) = everything added up and spread evenly; single outliers pull it strongly up or down. For ×2 the median and the mean drift apart because a few very good cells lift the mean; for ×1 the two are practically identical, as those outliers are absent. The finding in one sentence: the more aggressively the method bought the dip, the lower the return came out in the test and the calmer the curve ran. Source: own backtest 1927–2026.

With true doubling (×2), the return p. a. stood at a median of 3.29% and a mean of 3.60% — the two figures drifted apart because a handful of especially good cells pulled the mean up. At ×1.5, median and mean stood at 4.14% and 4.38%; at equal-sized tranches (×1, essentially a grid bot), both figures came to 5.58% — here there were practically no outliers. The drawdown rose sharply with gentler buying: from −7.75% (median) at ×2, through −16.95% at ×1.5, to −49.39% at ×1. The finding in one sentence: the more aggressively the method bought the dip, the lower the return came out in the test and the calmer the curve ran.

The counter-check on the Nasdaq 100

So that the result didn't hinge on one particular choice of stocks, the same main rule was applied unchanged to the Nasdaq 100 (QQQ).

Return p.a. = the annual compounding rate that would turn the starting capital into the final capital. Largest drawdown = the deepest fall from the highest account balance reached up to that point. The fair benchmark buys one base unit once and holds it; the remaining money sits in US Treasury bills throughout. "—" means the metric does not exist for that line, because it is invested throughout and has no rounds. Source: own backtest 1927–2026.

The picture stayed the same here too: 3.73% p. a. ($273,550.60) against 10.72% p. a. for buy and hold ($1,638,884.13), with a −2.39% drawdown against −82.98%; the fair benchmark reached 2.15% p. a. ($179,549.12). Here too, the success criterion went unmet.

How much reserve would have been needed — and 1929 to 1932

A cap of 8 purchases limits a round because the money ran out — not because the price couldn't have fallen further. How much reserve the method would actually have needed is shown by a calculation with no cap at all, in multiples of the base unit.

Computed without a cap: the method keeps buying as long as the market falls. Base unit = the amount a round starts with and against which every follow-up purchase is measured. The main rule is ×2 / −5% / +2%. The highest requirement across all series and rules was 262,143 base units — that much money would have had to be on hand to see a single round through. Source: own backtest 1927–2026.

For the main rule alone, SPY would have tied up 255 base units at the worst point of its history (the 2007–2009 financial crisis), across 8 purchases in a row; QQQ and the S&P 500 price index needed eight times as much in their own worst crisis, at 2,047 base units. Across all 36 rules and the whole available history, the highest requirement ever reached was 262,143 base units — reached by the Nasdaq 100 (×2, 3% threshold, 2% target) and by the S&P 500 price index at a 3% target, whose deepest round began on March 8, 1932.

Computed with the main rule and the cap of 8 purchases per round. "Empty reserve" means the method wanted to buy more but had no money left — during the Great Depression on 103 trading days; total capital was at times 29.15% below its high-water mark. Largest drawdown = the deepest fall from the highest account balance reached up to that point. The 2020 Covid crash is deliberately left out; house rules keep it out of every evaluation. Source: own backtest 1927–2026.

How tight things actually got from 1929 to 1932 is shown by the capped main rule on the S&P 500 price index (without dividends, because this long series has no dividend history): with the market down −86.19%, the reserve sat completely empty on 103 trading days — total capital fell by as much as 29.15% along the way. For comparison: during the 2007–2009 financial crisis, the ETF's reserve ran empty on only a single day, and the deepest drawdown stayed at −2.66%. Over the full 1927–2026 window, the capped S&P 500 price index returned 0.55% p. a. at a −29.15% drawdown, against 6.35% for buy and hold at −86.19%; the Nasdaq 100 price index (1985–2026) showed the same pattern (0.77% against 14.59%). Both series lack dividends — the price paid for the long history.

Year by year

Even on a year-by-year basis — excluding 2020, whose extreme price swings would have distorted every yearly average — the picture stayed the same.

Calendar years excluding 2020 — house rules keep the exceptional year out of every annual evaluation. Median = the typical case in the middle: half the values were above it, half below. The method closed every single evaluated year in the black and still trailed buy and hold in most years. Source: own backtest 1927–2026.

For SPY, the method's annual return stood at a median of 2.81% and a mean of 2.97% — the two figures sat close together here, because there were no extreme outlier years. Buy and hold reached a median of 15.06% but a mean of only 11.99%; a handful of weak years for the market pulled the mean down, while the typical year (the median) came in noticeably better. For QQQ, the pattern was similar (method median 2.59%, mean 3.90%; buy and hold median 18.11%, mean 13.69%). In every single one of the 33 (SPY) or 27 (QQQ) years, the method finished in the black — and still trailed buy and hold in 26 of 33 and 20 of 27 years respectively; it was actually ahead in only 7 years each.

What this study can't show

• No taxes. The calculation showed pre-tax, gross returns; with 546 rounds over 33 years, the tax bill from all those realized gains would likely have come out higher than for buy and hold, which sells only once at the end.
• Trading costs are an assumption. The calculation used a spread of 0.02% per side plus a minimum commission — realistic for liquid ETFs, not necessarily for every conceivable real-world execution.
• Crash days were executed in a simplified way. Every order was treated as filled at the most favorable price within the day's range; on real crash days with extreme demand, actual prices can turn out worse than assumed here.
• The base unit scales with the capital. Every figure here is for $100,000 in starting capital; with a larger amount, the follow-up purchases would be larger too — whether the market could absorb orders of that size without moving the price wasn't tested.

Close

Over 33 years on the S&P 500 and 27 years on the Nasdaq 100, buying the dip in its main rule, named in advance, returned 2.94% and 3.73% per year respectively — well below buy and hold's 10.84% and 10.72%. Of 216 rule combinations, none met the criterion fixed in advance, even though practically every one of them cut the drawdown sharply compared with buy and hold. Most of the return came not from the trading itself, but from interest on the capital that sat unused in reserve most of the time — and in the worst crisis tested, the Great Depression of 1929 to 1932, that reserve would have sat completely empty on 103 trading days.

Source: own backtest of price data, 1927 through 2026. Four series: S&P 500 ETF from 1993 (8,449 trading days), Nasdaq 100 ETF from 1999 (6,907), S&P 500 price index from 1927 (24,779) and Nasdaq 100 price index from 1985 (10,308). Interest on the idle reserve follows the FRED series DTB3 (3-month U.S. Treasury bill). Figures as of August 26, 2026. Design preregistered before the first number was computed.

This article is a historical analysis of publicly available price data and is not investment advice. It contains no buy or sell recommendation, no forecast, and no statement about any individual company listed today. Past results — simulated or real — are not a reliable indicator of future returns. Anyone making investment decisions should assess their own situation and risks, and seek professional advice where in doubt.