Does a moving average running through an old pivot hold better? Across 4,808 tests of the EMA 25 on 57 Bybit perpetuals, confluence does not make the level hold more often. It does reduce how far price overshoots it, in every configuration tested.
A level is stronger when several reasons to watch it overlap. That is the idea of confluence: a moving average running exactly through an old floor or ceiling is supposed to be worth more than an average on its own. The idea is everywhere, almost always illustrated with a well-chosen chart, and rarely counted.
The article on the moving average gap fill left the question open: does an average that coincides with a horizontal level pull price harder? The measurement below answers it, and the answer is not the expected one. Confluence does not make a level hold more often. It changes how far price overshoots it.
To count, everything has to be defined without judgement, and without borrowing anything from the future.
There is confluence when a pivot sits within half a standard deviation of the average. The reaction is observed over the next ten days: does the level hold (tenth close still on the right side), how far does price bounce, and how far does it overshoot to the other side.
Measured on 57 Bybit perpetuals among the sixty most traded, with up to 1,000 daily candles each: 4,808 tests of the EMA 25, 2,181 from above and 2,627 from below.
The trap is to simply compare tests "with a pivot" and "without a pivot". A pivot near the average mostly appears when price is ranging, and in that case the average behaves differently anyway. The level would then get credit for an effect that belongs to the market regime. Hence three groups: pivot on the average, pivot offset by 0.5 to 3 standard deviations, and pivot far or absent. The offset group serves as the control: the same structured market, without the coincidence.
| Holding after 10 days | Support | Resistance |
|---|---|---|
| Pivot on the average | 50.5 % | 58.9 % |
| Pivot offset | 42.9 % | 58.1 % |
| Pivot far or absent | 52.3 % | 59.2 % |
On the support side, confluence seems to help: 50.5 % holding against 42.9 % with an offset pivot. But with no pivot at all, the level holds just as often (52.3 %). On the resistance side, all three groups sit at the same level, around 58 %.
The gap between support and resistance has nothing to do with the level. Over the period measured, most altcoins fell: resistance therefore holds more often than support, whatever surrounds it.
Does the support-side result survive the settings? Not really:
| Setting | Support: on / offset | Resistance: on / offset |
|---|---|---|
| Pivots over 5 candles | 47.6 / 43.7 % | 58.1 / 58.6 % |
| Pivots over 10 candles | 50.5 / 42.9 % | 58.9 / 58.1 % |
| Pivots over 20 candles | 47.1 / 46.3 % | 61.2 / 57.7 % |
| 5-day horizon | 49.4 / 47.5 % | 62.9 / 55.4 % |
| 20-day horizon | 48.0 / 42.6 % | 60.1 / 61.8 % |
On the support side, confluence always does at least as well, but the gap ranges from under one point to nearly eight points depending on the setting. On the resistance side, it changes sign from one row to the next. An effect that depends this much on the settings does not support the claim that confluence makes a level hold.
Holding is a binary measure: the level held on day ten or it did not. It says nothing about what happened in between. The adverse excursion measures how far price went beyond the average before making up its mind. Here the result is clear:
| Median overshoot, in standard deviations | On the level | Offset | Far |
|---|---|---|---|
| Support, pivots over 5 candles | 2.15 | 2.32 | 2.36 |
| Support, pivots over 10 candles | 1.91 | 2.35 | 2.25 |
| Support, pivots over 20 candles | 1.95 | 2.17 | 2.36 |
| Resistance, pivots over 5 candles | 1.84 | 2.06 | — |
| Resistance, pivots over 10 candles | 1.77 | 2.03 | 1.96 |
| Resistance, pivots over 20 candles | 1.57 | 1.95 | 2.04 |
In all six configurations, the overshoot is smaller when a pivot coincides with the average: 7 to nearly 20 % less than with an offset pivot. The result does not hinge on a few contracts. Taking the median contract by contract, the gap is 2.01 against 2.37 standard deviations on the support side, and 1.67 against 2.04 on the resistance side.
The ratio between bounce and overshoot points the same way: it is better on the level in every configuration tested, for support and resistance alike.
Confluence reduces the typical overshoot, not clean breaks. The share of tests in which price goes more than three standard deviations beyond the average is about one third in every group: 32 to 35 % on the level, 33 to 37 % with an offset pivot. When a level really gives way, the fact that it coincides with an old pivot changes almost nothing.
One last figure is worth keeping, and it holds for every group. The median overshoot is close to two daily standard deviations. In half the tests, price therefore went at least that far to the other side of the average within ten days, including when the level eventually held.
The same measurement on the EMA 200 yields 1,089 tests, but 86 % of them have a pivot within one and a half standard deviations. The long moving averages move slowly, and pivots pile up around them. Almost no test is left without a nearby level to serve as a comparison. Where the comparison is possible, the gap in holding is nil on the support side (44.8 % against 44.6 %) and small on the resistance side (53.7 % against 51.2 %).
Confluence should not be read as a higher probability that the level holds. It says something else, more useful than it looks:
The screener shows where price stands relative to its averages. To add a horizontal level, you have to draw it yourself. This measurement shows what to expect from it: a narrower margin of error around the level, not a better chance that it holds.
A pivot defined over 5, 10 or 20 candles is only one way among others to draw a level. A trader who draws supports by hand, taking volume or wicks into account, is not looking at exactly the same ones. The measurement covers the EMA 25 and nearly three years of daily candles, a single market cycle.
It leaves open the second half of the original idea: a level's strength is supposed to come from the number of independent reasons to watch it. Yet on crypto perpetuals, reasons are rarely independent: most contracts move with Bitcoin. How much of their movement truly belongs to them? That is the subject of the article on correlation.
What it does establish is enough to correct a common reading. Confluence does not make a level more reliable, it makes it more precise: price overshoots it less, but does not hold there more often.
This is not investment advice. This article sets out technical-analysis concepts for educational purposes. Past performance is no guide to future performance, and trading crypto-assets carries a risk of losing the entire capital committed. Legal information.