The Lab

How we tested the first two

• Spot, starting with 10,000, with a 0.1% fee each way.
• Three windows, never pooled:
  • 14 coins, April 2024 to September 2026, mostly falling: BTC, ETH, SOL, ZEC, NEAR, XRP, DOGE, GTC, STRK, RSR, PEPE, WIF, BONK, SHIB, the same coins as our grid study.
  • BTC, ETH and SOL, October 2020 to September 2026, with up years.
  • BTC alone, October 2017 to September 2026.
• Each window is split in two halves, each run on its own. Settings are picked on the first half and judged on the second, which they never saw.
• Entries at random times: each ladder starts on buy signals at random hours, a set number in each coin-month, drawn 20 times. Every setting gets the same draws, so only the ladder differs.
• A reference ladder in each window, to compare against:
  • falling window: 8 orders, 5% apart, each 1.5 times the last;
  • up years: 3 orders, 5% apart, equal sizes.

The take-profit is at one step over the average price.

• Prices are Binance spot hourly candles. Orders rest on the hourly lows.
• Every setting was fixed before the run, and every result was printed, not only the best.

The maths, step by step

What a ladder does to the average price.

• Buy 1,000 worth at 100 (10 coins). The price falls 5%, and the next order buys 1,500 worth at 95 (15.8 coins).
• The holding is now 25.8 coins for 2,500: an average price of 96.94.
• A take-profit one step (5%) over the average sells at 101.79.
• The price only has to come most of the way back, not all of it.

First half and second half. Each window is cut in two.

• We try every setting on the first half and pick the best.
• Then we run that pick on the second half, which played no part in choosing it.
• The best of many settings on old data always looks good, partly by luck. The second half shows how much of it was real.

The median coin. Line the coins up from worst result to best and take the middle one. With 14 coins, it is halfway between the 7th and the 8th.

• Example: the 2x perps pick in section 3, falling window, second half. Its 14 coins ran from −100% (twice) up to +82%.
• The middle two were +18.9% and +16.8%, so the median coin is +17.9%.
• The plain average was only +5.4%: the two coins that lost everything pull it down. The median shows the typical coin. The average shows the damage.

Points. A difference between two percentages is given in points. +22% against +15% is 7 points apart.

Ruin (section 3). A run is ruined if the account is liquidated, or if at the lowest point in any hour it is worth under 20% of its start.

• The ruin figure is the share of runs ruined, over every coin and draw.
• Example: 14 coins × 3 draws = 42 runs. If 8 are ruined, the ruin figure is 8 ÷ 42 = 19%.
• That pick's +17.9% median coin came with that 19% ruin.

Leverage (section 3) is how many times the account the ladder holds once every order has filled.

• At 2x, an account of 10,000 holds 20,000 of coin.
• Then a 10% fall below the average price costs 20% of the account.

1. Fewer orders, wider gaps, more patience

We tried 288 settings: 2–4 orders, gaps of 8–15%, a take-profit of 8–15%, equal or growing sizes, and a time limit on each round (none, 30 days or 90 days, after which it sells at the close).

What we found:

• Wider and fewer did not beat the reference ladder.
  • In the falling window, the first-half pick (4 orders, 15% gap, 8% take-profit) lost 15% in the second half.
  • The reference ladder (8 orders, 5% apart, ×1.5) made +22% and beat it on 10 of 14 coins.
  • In the up years the pick made +12% against the reference's +27%.
• Patience as time is the setting that matters, and it cuts both ways.
  • In the falling window, a 30-day limit lost least: −23% against −39% with no limit. Giving up took a loss each time, but freed the money for cheaper rounds.
  • In the up years and on BTC, no limit was best: +34% against +14%, and +46% against −15%. Waiting paid because prices came back.
  • With no limit, the money sat at a loss 71–89% of the time.
  • Waiting pays only if the market recovers, and nobody knows that in advance.
• Patience as depth barely mattered. Going from an 8% gap to a 15% gap moved the result under 8 points either way.
• More orders beat fewer in the falling window (4 against 2: about +17 points). The take-profit width and growing sizes made no steady difference.

2. Widening gaps and a split take-profit

Two features of our DCA engine:

• Gaps that widen down the ladder: each step is 1.3 or 1.6 times the last.
• A take-profit in two halves: half sells at one step over the average price, half at two steps.

We tried 3, 5 or 8 orders, a first gap of 2, 3 or 5%, and equal or growing sizes. Each was compared with the plain version on the same random entries.

What we found:

• Selling in two halves helped almost everywhere. It beat the single take-profit in 52 of 54 settings in the falling window (+3.6 points on the median coin), 53 of 54 in the up years (+6.2), and 41 of 54 on BTC alone (+4.7). The second half rides part of each bounce further.
• The exception is the deep ladder: 8 orders, 5% apart, ×1.5. In the falling window's second half, halves made it much worse: +4.7% on the median coin against +21.5% with one take-profit.
  • It did worse on 7 of 14 coins, and the losses were large (PEPE −55, BONK −37, RSR −32, GTC −30 points). The gains were small, at most +21.
  • Waiting for the second half left it holding at a loss 75% of the time instead of 52%.
• Widening gaps barely mattered. At the same first gap they moved the result by about 2 points or less. At the same total depth they were slightly better (+0.1 to +1.2 points). Most of what wider steps do is reach deeper, and depth wasn't the edge either.

3. How much leverage can a ladder take?

We ran the same ladder with borrowed money, three ways, at 2, 3, 5 and 10 times the account (the leverage once every order has filled):

• Perpetual futures (perps): Binance's perpetual hourly prices and its real funding payments, with Bybit's fees (0.02% for a resting order, 0.055% for a market order). A position is liquidated at Bybit's maintenance margin, and a liquidation takes the whole account.
• Spot with a margin loan: spot bought partly with borrowed money, at 10% a year interest (5% and 20% were tried too). It is liquidated when the holding is worth less than 1.1 times the loan.
• Plain spot, with no loan, to compare.

How:

• Ruin means the account was liquidated, or at its lowest point in an hour it was worth under 20% of what it started with.
• 240 ladder settings in each of the 9 markets (spot, and perps and margin at each leverage):
  • 3 to 12 orders, the first 1–4% apart;
  • gaps equal or each 1.5 times the last;
  • sizes equal or each twice the last;
  • a take-profit of 1, 2 or 4%, in one part or three.
• Long only, one round at a time. Entries at 30 random hours a month, drawn 3 times.
• Two windows: the 14 coins from April 2024 to September 2026 (mostly falling), and BTC, ETH and SOL from October 2020 to September 2026 (with up years).
• Each window is split in two halves. Settings are picked on the first half and judged on the second.
• Everything was fixed before the run.

What we found:

• Leverage did not pay once ruin is counted.
  • The best first-half setting at each leverage made a lot in the first half. In the second half it was ruined on 45–83% of coins and draws in the falling window, and 11–56% in the up years.
  • In the falling window, no setting at 2x or more avoided ruin in the second half (0 of 240 at every leverage). On plain spot, 50 of 240 did.
  • In the up years, the settings with no ruin in the second half were: 25 of 240 at 2x on perps, 16 at 3x, 4 at 5x and 1 at 10x. On margin it was 20, 15, 3 and 1. On plain spot, all 240 had no ruin.
• Picking only among settings with no ruin in the first half, judged on the second half (the median coin's result, and how often it was ruined):

"None" means no setting avoided ruin in the first half. Plain spot did best in both windows.

• The leveraged settings that survived were ladders with growing sizes: 8–12 orders, 2–4% apart, each order twice the last. They hold little until the price is deep, so they act like far lower leverage until the bottom.
• "Wait until every order has filled before liquidation" was not enough. Almost every ladder fills fully before its liquidation price (234 of 240 at 2x, 180 of 240 at 10x). The ruin comes after: the price keeps falling past a full ladder.
• Funding was a real cost in the up years. The best first-half 2x perps setting paid 26% of the start in funding over the second half, against a net of +217% (and it was ruined on a third of coins and draws). Margin interest was small next to the ruin; 5% or 20% a year changed the picks by a few points.
• Textbook entry signals did not help. At each market's pick, each signal was tested against 10 random entry draws with the same number of entries each month. In the falling window, a coin beating all 10 draws in a half happened this many times out of 126 (chance gives about 11):

• The same coin beat in both halves at most 2 times.
• In the up years (27 each), 0–7 beat.
• "10% under the 7-day high" looked good in the second half, but went the other way in the first.

What it means: on these windows, more leverage meant a bigger first half and a much higher chance of losing the account. The only leveraged ladders that survived were the ones that barely used their leverage until deep. If leverage is used at all, read the chance of ruin before the profit.

What it all means

• No ladder wins in every market. A ladder that sells everything on a bounce does well when prices fall. Waiting longer pays only when prices come back.
• The price of a DCA is time spent at a loss, often months, with the money tied up the whole time.
• Wider gaps and fewer orders didn't help. The time limit was the setting that mattered, and the right one depended on a market nobody can know in advance.
• Selling in two halves usually adds a few points. On a deep ladder in a falling market, though, it can cost a lot.
• Leverage on a ladder mostly adds the chance of losing everything. Plain spot did best once ruin was counted.

Limits

• Few windows, and BTC since 2017 is a single coin.
• Our DCA rests its orders on hourly lows. That flatters it in crash wicks, such as 10 October 2025.
  • A wick is the thin line above or below a candle's body: how far the price went during the candle before coming back. For example, a candle that opens at 100, dips to 95 and closes at 99 has a wick down to 95.
• The split take-profit here has no breakeven stop.
• The leverage test:
  • Only 3 random draws.
  • The coins move together, so they are not independent.
  • The liquidation and margin models are simple: one line each.
  • Hourly lows are not the exchange's mark price.
    • The mark price is the price a futures exchange uses to value positions and to decide liquidations. It is built from several exchanges' spot prices, so one wild trade can't liquidate anyone. For example, a sell that knocks the traded price down 3% for a second may barely move the mark.
  • The liquidation tiers are today's.

Update, 27 October 2026: where the idea comes from

Added after this page was published. The text above is left as it was.

• Two ideas share the name.
  • Dollar-cost averaging buys a fixed sum at fixed times, whatever the price. Benjamin Graham described it for ordinary investors in 1949. Its theory is on our page "Steady buying or all at once: the theory of dollar-cost averaging".
  • A DCA ladder, tested here, is different. It buys only as the price falls, in growing sizes, and sells on a bounce. It comes from trading bots' "safety orders", with no known inventor.
• Growing sizes are a martingale. The martingale is an old French betting system: double the stake after each loss, so the first win wins back every loss plus one stake.
  • For example: stakes of 1, 2, 4 and 8. Lose three, win the fourth: −1 − 2 − 4 + 8 = +1.
  • Jean Ville brought the word into probability in 1939, for a fair game.
• The catch, worked through for our reference ladder (8 orders, each 1.5 times the last):
  1. The money it needs: 1 + 1.5 + 2.25 + … up to the 8th order adds up to about 49 times the first order.
  2. Where it goes: the 8th order alone is about 17 times the first. Most of the money goes in near the bottom.
  3. So: a ladder wins small amounts often, and loses a large amount when the price falls past its last order and stays there. That is the martingale's shape. The ruin figure in section 3 measures it.
• Theirs and ours. Graham's averaging, the martingale, Wilder's RSI, Bollinger's bands and Fisher's random comparison are theirs, credited below. The ladder's settings, the first-half and second-half test and the ruin measure are ours.

Credits

• A DCA ladder with "safety orders" is a common trading-bot design, and averaging down with growing sizes echoes the old martingale betting system. Neither has a known inventor. Our ladder settings are our own.
• Benjamin Graham, The Intelligent Investor, Harper & Brothers, 1949: dollar-cost averaging for ordinary investors.
• Jean Ville, Étude critique de la notion de collectif, Gauthier-Villars, 1939: the martingale in probability.
• J. Welles Wilder Jr., New Concepts in Technical Trading Systems, Trend Research, 1978: the RSI.
• John Bollinger introduced Bollinger Bands in the 1980s; his book is Bollinger on Bollinger Bands, McGraw-Hill, 2001.
• R. A. Fisher, The Design of Experiments, Oliver and Boyd, 1935: the idea behind comparing with random choices (the random entries).
• Data: Binance's public hourly prices and funding history; Bybit's fee table.
• PerpGremlin and Claude, Buying the dip in steps: three tests of our DCA ladder, TerminalGremlin, 2026: the tests, the code and the words.