The Lab

How we tested

• Bitcoin, Ether and Solana, from public exchange candles:
  • 1-minute candles over 30 days;
  • 15-minute candles over a year;
  • 1-hour and 4-hour candles over two years.
• A turn is a swing low or high: a candle lower (or higher) than a few candles each side. We only count it from the candle where that becomes known, never earlier.
• Leverage from 2x to 125x.
• Two controls:
  • the same test from random moments;
  • fake markets, built by cutting the real candles into blocks of 50 and shuffling the blocks (a "block bootstrap", after Künsch, 1989): 120 fake markets for the 1-hour chart, 240 for the 4-hour. They keep each candle's shape and how jumpy each stretch is, but lose any longer pattern.
• A trade test: trade toward the bounce side's liquidations, stopped by the trade's own liquidation, after fees (0.11% in and out).
• The first run was on 9 October 2026. It was re-run on 14 October on the latest data, and the figures here are the re-run's.

The maths, step by step

1. Where a position is liquidated. Here we use the plain arithmetic: a position at L times leverage is liquidated a move of 1 ÷ L away (the exchange's small margin cushion is left out, so leverage is the only thing that varies).

• Example: a swing low at 100.
  • A 25x short opened there is liquidated at 100 × (1 + 1/25) = 104.
  • A 25x long is liquidated at 100 × (1 − 1/25) = 96.
  • At 100x the levels are 101 and 99. At 2x they are 150 and 50.

2. The race. From the moment the low is known, which level does the price reach first: the shorts' above (the bounce side) or the longs' below? Some turns reach neither within the time allowed.

3. How far the bounce goes. For each turn, how far does the price bounce before it makes a new low? The share of turns whose bounce reaches 1 ÷ L is the chance a fresh L-times position on the wrong side would have been liquidated.

4. The fake-market test. Build many fake markets from the real candles and run the same test on each. If the real market sits in the middle of the fakes, the effect comes from how much prices move, not from anything the levels do.

• Example: a real figure of 64.6% against fakes that average 63.6% sits at the 82nd percentile: 82% of the fakes were lower. That is well inside the fakes' range.

5. A 95% interval. Resample the trades with replacement thousands of times (Bradley Efron's "bootstrap", 1979), and take the middle 95% of the averages. If the interval spans zero, the average can't be told from zero.

What we found

1. The bounce side does come first more often, but fake markets do the same. After a turn on the 1-hour chart at 10x, the price reached the bounce side's liquidations first 64.6% of the time. Fake markets did it 63.6% of the time on average.

A turn is measured from the extreme by its very definition, so a bounce from it is partly geometry. "The knife stopped" is how a random walk looks at this volatility, not a pull toward the levels.

2. Trading toward the bounce side lost after fees.

• On the 1-hour chart every leverage tested lost on average (−0.07% to −0.14% a trade).
• From 25x up, the 95% interval sat wholly below zero.
• On the 4-hour chart, 5x and 10x lost too, with intervals spanning zero.
• At 2x, 3x and 7x on the 4-hour chart the average was slightly positive (+0.05% to +0.17%). But in most of those trades neither level was reached in time, and the 5x interval ran from −0.64% to +0.47%.

3. Resampled into runs of 200 trades, each risking a tenth of the money as margin, every leverage lost at the median. The share of runs ruined:

4. No clustering at liquidation distances or at Fibonacci ratios.

• Turning points sat within ±2% of the move at Fibonacci ratios 25.5% of the time on the 1-hour chart. They sat at the same-width control levels just beside them (the ratios moved by 0.06) 26.6% of the time.
• At liquidation distances, 47.2%; at distances 15% nearer or further, 46.0%.
• No timeframe showed either.

5. Nothing predicted the bounce on unseen data. A model using the size of the fall, volume and RSI, fitted on the first 70% of the history, was no better than the plain base rate on the last 30%. Its score was within about ±3% of the base rate, mostly worse.

A risk table worth keeping

How far the price bounced after a turn: the share of turns whose bounce reached a fresh position's liquidation distance. This is true of fake markets too, so it is a fair map of volatility against leverage, not a signal.

In plain words: on the 4-hour chart, a 25x position opened against a fresh turn was liquidated by the bounce about 6 times in 10. At 100x it was almost always liquidated.

Update, 27 October 2026: where the idea comes from

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

• The idea that liquidation levels pull the price toward them is trader lore. It is behind the liquidation "heat maps" on charting sites. No one person's method was copied here: the levels were built from swing points with the plain 1 ÷ L arithmetic.
• Why it might be true. A liquidation is a forced market order, so many of them at one price are a pile of orders waiting to be set off. Carol Osler studied the nearest thing in currency markets: stop-loss orders cluster at round numbers (2003), and once the price reaches them, they can push it further (2005).
• The catch: a swing point is a bounce by definition.
  • A swing low here is a candle lower than a few candles each side. So it is known only once those later candles have printed.
  • For example: on the 1-hour chart, with 4 candles each side, by the time the low is known the 4 candles after it have all had higher lows.
  • So every turn starts with a bounce already under way. The fake markets have the same built-in bounce, which is why they are the fair comparison.
• Fibonacci ratios and where they come from are on our theory page "Fibonacci and Elliott waves: where the numbers come from".
• The next step this page named, real liquidation records, is under way. That follow-up's first look is reported, not judged. Its judged looks are fixed for January and April 2027.
• Theirs and ours. The bootstrap (Efron), its block form (Künsch), the RSI (Wilder) and the Fibonacci sequence are theirs, credited below. The swing-point levels, the race, the fake markets of 50-candle blocks and the trade test are ours.

Credits

• Bradley Efron, "Bootstrap Methods: Another Look at the Jackknife", Annals of Statistics, 1979.
• Hans R. Künsch, "The Jackknife and the Bootstrap for General Stationary Observations", Annals of Statistics, 1989: the block bootstrap.
• J. Welles Wilder Jr., New Concepts in Technical Trading Systems, Trend Research, 1978: the RSI.
• Leonardo of Pisa (Fibonacci), Liber Abaci, 1202: the sequence behind the ratios.
• C. L. Osler, "Currency Orders and Exchange Rate Dynamics: An Explanation for the Predictive Success of Technical Analysis", Journal of Finance, 2003: orders that cluster at round numbers.
• C. L. Osler, "Stop-Loss Orders and Price Cascades in Currency Markets", Journal of International Money and Finance, 2005: forced orders that push the price further.
• The idea that liquidation levels act as price targets is trader lore, with no known originator. The tests are our own design.
• Data: Bybit's public candles.
• PerpGremlin and Claude, Liquidation levels after a turn: targets, or just distance?, TerminalGremlin, 2026: the tests, the code and the words.