Rounding bottom

Reversal shape3 completed examples in the record

What the shape is

A slow, curved base rather than a sharp turn. A parabola fits a V-shaped reversal well enough that curvature alone cannot separate the two, so both models are fitted and the shape is only called rounded when the parabola wins.

What followed it

HorizonExamplesTextbook directionMedian changeMiddle half
5 bars3too few examples+16.07%+13.56% … +17.38%
10 bars3too few examples+13.16%+4.55% … +14.87%
20 bars3too few examples+10.65%+3.73% … +12.80%
60 bars3too few examples+17.71%+9.53% … +19.23%

10 instruments, 1d / 4h, 2017-08-17 to 2026-07-25 — 182,341 bars and 661 completed formations.

engine v1.0.0 · Record generated 2026-07-25

  • These figures describe past occurrences of the same shape. They are not a projection of this chart.
  • 661 formations is not a large sample, and the instruments are correlated USDT pairs, so the effective sample is smaller still.
  • Below 30 examples no rate is printed at all.
  • Changes are close-to-close and exclude fees, funding and slippage.
  • Medians and quartiles pooled across instruments are sample-weighted approximations, not exact pooled quantiles.
  • A formation is timed from the bar its break became observable, and every change is measured strictly after that bar.
  • Pooled across every instrument and both backtest timeframes.

The same shape, split by interval

Interval 4h

HorizonExamplesTextbook directionMedian changeMiddle half
5 bars3too few examples+16.07%+13.56% … +17.38%
10 bars3too few examples+13.16%+4.55% … +14.87%
20 bars3too few examples+10.65%+3.73% … +12.80%
60 bars3too few examples+17.71%+9.53% … +19.23%

The backtest covered the daily and 4-hour intervals only. Only one shape reaches the publication floor on the daily, against eight on the 4-hour — so every pooled rate above is in practice a 4-hour statistic. The console scans five intervals; the three not listed here have no historical record at all.

How it is detected

Formations are matched against swing pivots rather than raw bars, so a single spiky candle cannot create one. A pivot is committed only once price reverses by more than 2.5 times ATR(14), floored at 0.8% of price. A detection is reported only at confidence 0.72 or above — the library default, unmodified.

Rounding bottom in crypto: the measured record | Chart Intel