Averaging down and pyramiding: which way to add

Averaging down means buying more of a holding as it falls to lower the average cost. Pyramiding means buying more as it rises to grow a winning position. Both add the same amount; only the direction differs. Korean investors call them mul-tagi and bul-tagi. This article writes both as the same rule and runs them over every five-year window of QQQ and TQQQ, comparing the money actually put in, the end balance, the value per dollar put in, and the largest amount that could have been lost. Each rule puts in a different amount, so comparing balances alone makes the rule that put in more look better. Figures are as of September 30, 2026.

Rules

• Budget: $40,000 is set aside for one holding, and money not used is counted as cash with no interest.
• First buy: $10,000 at the open of the window's first trading day.
• Averaging-down ladder: each time the close is 10%, 20% or 30% below the first buy price, another $10,000 is bought at the next open. Each step fires once, so at most three times.
• Pyramiding ladder: each time the close is 10%, 20% or 30% above the first buy price, another $10,000 is bought the same way.
• Baselines: buying only the first $10,000 and keeping the rest in cash, and buying all $40,000 on the first day.
• Prices and costs: signals use split-adjusted closes; fills and valuation use prices with dividends reinvested. Every trade costs 0.15%. There is no selling rule.

Buying the same amount four times on the way down puts the average cost 16.5% below the first buy price. Someone who bought once needs +42.9% from a point 30% below the first buy just to break even; someone who bought four times needs +19.3%. That arithmetic is why averaging down is tempting. In exchange, the money at risk grew from $10,000 to $40,000, and if the price keeps falling the whole larger sum falls with it. Pyramiding is the reverse. The average cost rises by 13.9%, so the four-buy position is at −12.2% as soon as the price merely returns to the first buy price.

Every five-year window of QQQ and TQQQ

The four approaches were run on every five-year window, moving the start one month at a time: 271 QQQ windows starting April 1999 to October 2021, and 140 TQQQ windows starting March 2010 to October 2021. The balance includes unused cash. Money put in is what was actually spent on the stock, and value per dollar is the stock's value after five years divided by that money. The balance reflects how much went in; the value per dollar reflects when it went in. A losing window is one where the account ended below $40,000, and the last column is the largest gap between the money put into the stock and its value.

On QQQ the averaging-down ladder fired at least once in 155 of the 271 windows and all three times in 91. In the other 116 windows the $30,000 stayed in cash to the end, and the result matched buying only the first $10,000. Averaging down ended above putting all $40,000 in at the start in 100 windows (37%), all of them windows where the ladder fired. Counting only those windows, it led in 65%, and in the 91 windows where it bought all three times and so put in the same $40,000, it led every time. Buying the same money in pieces below the first price makes that inevitable. The catch is that no one knows in advance whether the ladder will go all the way. The median money put in was $20,000 for averaging down (mean $23,137) against $40,000 for everything at the start, and most of the gap in median balances comes from that. The median value per dollar put in after five years was 2.11x for averaging down, 2.03x for everything at the start and 1.75x for pyramiding.

All 140 five-year TQQQ windows ended with a gain. The ladder fired in 101 of them, and averaging down beat putting everything in at the start in 62 (44%). SPY (335 windows starting December 1993 to October 2021) pointed the same way. The five-year median was $50,694 for averaging down and $67,777 for everything at the start, and 10 and 49 windows ended with a loss. Much of averaging down's smaller count of losing windows comes from putting in less money. Its median money put in on SPY was $20,000, and the median value per dollar was 1.77x for averaging down and 1.69x for everything at the start.

Averaging down adds money only on the way down, so in windows where the market kept rising most of the budget stayed in cash. Pyramiding adds on the way up, so its median was higher than averaging down, but in windows that rose and then fell it met the decline holding a large sum at a high average cost. The worst QQQ pyramiding window started in February 2000: it made all three buys on February 8, March 2 and March 27, just before the dot-com bubble burst, and was worth $14,695 five years later.

Starting at a peak

Buying QQQ on the first trading day of March 2000, the ladder fired on April 13, April 17 and May 24. The decline ran until October 2002, and the $40,000 in the stock was at one point $30,736 short (−76.8%). Five years later the account held $17,360 with averaging down, $13,999 with everything at the start, and $33,500 with the first $10,000 only. Averaging down and everything at the start both put in $40,000, so averaging down did better by the amount it bought cheaper. It still ended more than $16,000 below putting in only $10,000.

Buying TQQQ on the first trading day of November 2021, the ladder fired on January 19, 21 and 26, 2022. TQQQ was already 30% below the first buy in January 2022, and kept falling until December 2022. By September 2026 the account held $101,130 with averaging down, $84,128 with everything at the start, $71,308 with pyramiding and $51,032 with the first $10,000 only. Apart from the first-$10,000 case, all three put in $40,000. Along the way, averaging down's largest paper loss was $29,988 (−75.0%). Buying SPY in October 2007, the ladder fired in January, July and October 2008, and five years later averaging down held $49,672 against $41,938 for everything at the start.

Averaging down with no cap

Without a cap on the ladder the arithmetic changes. Running "buy another $10,000 whenever the price is 10% below the average cost" with no limit on QQQ from March 2000, each $10,000 barely moves the average cost once the position is large, so the condition is met again every day while the price stays more than 10% below the average cost. By March 12, 2009 it had bought 518 more times, putting in $5,190,000, with a largest paper loss of $1,468,912. TQQQ from November 2021 bought 198 times through March 2023, putting in $1,990,000, with a largest paper loss of $632,111. Both later rose a great deal, but these are sums that a plan sized at $10,000 cannot carry. What decides an averaging-down result is not the size of the decline but the cash left.

What research found

In a 1998 Journal of Finance paper, Terrance Odean analyzed the 1987 to 1993 trades of 10,000 accounts at a US discount broker. The rate of buying more of a stock already held was higher for stocks at a loss (PLPA 0.135) than for stocks at a gain (PGPA 0.094). Odean read this as consistent with prospect theory's risk seeking in losses, or with a belief that prices revert. It comes from the same reference point as the disposition effect, which keeps losers and sells winners early (Investor psychology): the purchase price. A 1990 study by Richard Thaler and Eric Johnson found that people who have lost are drawn to choices that offer a chance to get back to even, and called it the break-even effect.

Trend followers wrote the opposite rule. In his 1940 book How to Trade in Stocks, Jesse Livermore forbade adding to a losing position more strongly than anything else (Livermore's trend breakout), and added in steps after a rise. The Turtles added one unit each time price moved ½N in their favor from the last fill, up to four units per market (Turtle trading). In both cases pyramiding came with a stop, and this article's pyramiding ladder has none.

When the idea holds and when it fails

• On broad indexes that recovered, averaging down became buying near the bottom: every figure above comes from US index ETFs that regained their highs. If the recovery is slow the budget runs out early, and if there is no recovery the arithmetic does not work at all.
• On single stocks that never recovered, the extra money was lost too: Enron's share price was $90.75 in August 2000, closed at $0.61 on November 28, 2001, and the company filed for bankruptcy on December 2 (Wikipedia). Lehman Brothers fell 73% in the first half of 2008 alone, dropped 45% to $7.79 on September 9, and filed for bankruptcy on September 15 (Wikipedia). SVB Financial fell more than 60% to close at $106.04 on March 9, 2023, and the next day trading was halted and the bank was closed (CNBC). With a ladder on stocks like these, the loss would have been $40,000 rather than $10,000.
• Pyramiding was weak where a rise turned down: with the average cost raised, a return to the first buy price is already a loss. That is why trend following pairs pyramiding with a stop.
• The purchase price has no bearing on future returns: both approaches take the first buy price as their yardstick. The real question is how much to hold if the stock were bought fresh today, and the purchase price changes that answer.

Related concepts and search terms

• Lowering the average, scaling in: other Korean names for averaging down. Unlike dollar-cost averaging, which splits buys by date, it buys only when the price falls (SPY monthly vs. lump sum).
• Pyramiding: buying more of a winning position in steps. The original sense buys less at each step; this article's ladder uses equal amounts.
• Martingale: the gambling rule of doubling the stake after every loss. Uncapped averaging down has the same structure and only works with unlimited money.
• Disposition effect, break-even effect: using the purchase price as a reference point and holding or adding to losing stocks.

Search terms: averaging down, pyramiding, martingale, break-even effect, disposition effect, mul-tagi, bul-tagi, when to average down, averaging down vs. stop loss.

Limits of these numbers

• QQQ, TQQQ and SPY are US index ETFs that regained their highs. There is no basis for expecting the same results from single stocks or other countries' indexes.
• Only one setting was computed: 10% steps, three steps, $10,000 each. Other settings give other results.
• Five-year windows overlap, so they are not independent samples. QQQ's worst windows all started in February to April 2000.
• Interest on cash, taxes and exchange rates are left out. Paying interest on the idle budget would lift averaging down and the first-$10,000-only case a little.