Six out of ten trading signals are not signals
What 113 310 graded market situations on Binance futures actually show
We record every market situation the moment it fires and grade what happened next — including every time nothing happened. Across 98 708 directional situations, 59.8% produced no move at all, and the rest were a coin flip. Here is what beat its control, what did worse, and why confluence does not help you pick a side.
Updated November 3, 2026
Full table, controls and intervals →
Most writing about trading setups reports the times they worked. We run a system that records a market situation the moment it fires, then grades what happened next — every time, including all the times nothing happened. No setup is dropped for being embarrassing, because the record is written before the outcome exists.
This is what 113 310 of those records say. 109 208 of them have been graded so far.
A snapshot of a rolling 30-day window, taken 2026-11-03. The live table moves every day as the window rolls, so expect it to differ from the figures below by a point or two. The method does not change; the numbers do.
The first number is the uncomfortable one
Of 98 708 directional situations graded, 58 985 (59.8%) produced no qualifying move at all — not a loss, nothing. Price stayed inside the noise band until the horizon expired.
Of the 39 723 that did move, 20 369 went the way the situation implied and 19 354 went the other way. A 51.3% hit rate: a coin flip, for any practical purpose.
Six in ten "signals" are not signals. They are observations that happen to have a name.
Method
- Source. Binance USD-M perpetual futures, recorded live from the liquid universe. No backfill and no survivorship selection — the situation is written when it fires, graded later, and cannot be withdrawn in between.
- Window. Rolling 30 days, snapshot taken 2026-11-03. Collection started 2026-08-24.
- Threshold. A move counts when price crosses a threshold scaled to the instrument’s own volatility (ATR) within 24 hours, so a 1% move on BTC and on a small-cap alt are not counted as the same event.
- Control. Each type is measured against a matched baseline — the same grading applied to randomly picked coins at the same moments. The baseline is not one global figure, which is why it is printed on every row. A rate means nothing without it.
- Verdicts. Wilson 95% intervals on the rate. Edge: the interval lies entirely above its baseline. Anti: entirely below. Noise: the interval covers the baseline. Thin: under 30 observations, no verdict.
Two different questions, never mixed
Some situations imply a side; others do not. These are measured differently, and the two numbers are not comparable — so they get two separate tables rather than one ranking.
- Any move. For a side-less situation the rate is the share of ALL resolved observations where price crossed the threshold in either direction. Its control is how often a randomly picked coin moves at all: 41.8%.
- Implied side. For a situation that does imply one, the rate is the share of DECIDED cases where price crossed the threshold on the implied side BEFORE crossing it against. Observations where neither threshold was hit are dropped from both the numerator and the denominator. Its control is how often the market itself goes that way first.
The denominators differ, so a 60% in one table and a 60% in the other mean different things. Ranking one against the other is the mistake this split exists to prevent.
Of 43 types: 21 are indistinguishable from their control and 2 are too thin to judge. The honest summary of that line is that 21 of 43 named situations do nothing measurable at all.
Does price move at all?
Side-less situations only. The question is whether anything happens, not which way.
| Situation | n | moved | control | lift (pp) | 95% CI |
|---|---|---|---|---|---|
| Volatility spike | 469 | 64.8% | 41.8% | +23 | 60.4–69 |
| Open interest surge | 671 | 62.1% | 41.8% | +20.3 | 58.4–65.7 |
| Volatility squeeze | 107 | 57.9% | 41.8% | +16.1 | 48.5–66.9 |
| Active pump | 1 285 | 51.2% | 41.8% | +9.4 | 48.5–53.9 |
| Volume spike, no side | 3 318 | 38.8% | 41.8% | −3 | 37.2–40.5 |
| Bollinger bands squeezed (TTM) | 1 157 | 32.2% | 41.8% | −9.6 | 29.5–34.9 |
Volatility and open interest lead by a wide margin. Both measure participation — how much is actually happening — rather than interpreting the shape of a chart.
The two bottom rows are the interesting ones. “Bollinger bands squeezed” is the TTM condition: the bands (SMA20 ± 2σ) sitting entirely inside the Keltner channel (EMA20 ± 1.5×ATR20). It comes out −9.6 pp on 1 157 observations — after that squeeze, price moves LESS often than normal. “Volatility squeeze” at the top of the same table is a different measurement of the same word — the ratio of short to long realised range, ATR(14)/ATR(50) ≤ 0.9 — and it comes out +16.1 pp. Same concept, two formulas, opposite results, and here they are directly comparable because both answer the same question.
That is a warning about every claim in this field that names a concept instead of a formula, including the ones below.
Does price go the way the situation implies?
Situations that imply a side. The rate counts only cases that resolved one way or the other.
| Situation | n | implied side first | control | lift (pp) | 95% CI |
|---|---|---|---|---|---|
| Support zone broken | 763 | 63.3% | 58.7% | +4.6 | 59.8–66.6 |
| Price at resistance | 2 472 | 62.3% | 58.7% | +3.6 | 60.3–64.1 |
| Open interest change | 1 199 | 57.4% | 54% | +3.4 | 54.6–60.2 |
| Price crossed EMA50 | 1 372 | 54.1% | 50.7% | +3.4 | 51.4–56.7 |
| EMA trend | 2 354 | 56.2% | 53.1% | +3.1 | 54.2–58.2 |
| Equal lows (liquidity below) | 3 066 | 60.9% | 58.7% | +2.2 | 59.2–62.6 |
| False breakout / sweep | 5 460 | 51.2% | 49.7% | +1.5 | 49.9–52.5 |
| Equal highs (liquidity above) | 2 851 | 39.1% | 41.3% | −2.2 | 37.4–40.9 |
| Price at support | 2 324 | 37.1% | 41.3% | −4.2 | 35.2–39.1 |
| RSI divergence (1h) | 394 | 43.4% | 49.6% | −6.2 | 38.6–48.3 |
| Double bottom, confirmed | 232 | 33.2% | 41.3% | −8.1 | 27.4–39.5 |
| Oversold | 164 | 32.9% | 41.3% | −8.4 | 26.2–40.4 |
| RSI divergence (4h) | 309 | 37.5% | 46.8% | −9.3 | 32.3–43.1 |
| Overheated, late stage | 159 | 45.9% | 58.7% | −12.8 | 38.4–53.7 |
The confirmed double bottom underperforms its own control. The reversal pattern every course teaches came in at 33.2% against a 41.3% baseline on 232 decided cases — and “confirmed” here means the detector already waited for the second low and the break of the intervening high. Waiting for confirmation did not help; comparable observations without the pattern resolved upward more often.
RSI divergence is negative, and it gets worse on the higher timeframe. The 1h version reads 43.4% against 49.6%; the 4h version 37.5% against 46.8%. The usual advice — ignore the noise, trade divergence on a higher timeframe — points the wrong way in this sample.
Note what the positive rows have in common: breaks and trend-following reads beat their control, reversal reads do not. The largest lift in this table is under five points, so none of this is a strategy — it is a ranking of which readings are least wrong.
Confluence buys movement, not direction
The standard belief is that stacking confirmations improves entries. This is directly testable, because every record carries how many other situations were live on the same instrument at the same moment.
| signals stacked | n | moved at all | n (decided) | implied side first |
|---|---|---|---|---|
| 0 | 509 | 44% | 4 891 | 53.8% |
| 1 | 2 130 | 43.7% | 19 500 | 50.5% |
| 2 | 4 254 | 40.8% | 9 660 | 51.4% |
| 3 | 2 199 | 43.8% | 3 981 | 52% |
| 4 | 934 | 51.1% | 1 246 | 49.8% |
| 5 | 331 | 54.1% | 336 | 53.9% |
| 6 | 86 | 61.6% | 86 | 57% |
| 7 | 44 | 56.8% | 18 † | 38.9% |
| 8 | 10 | 50% | 5 † | 20% |
The chance that something happens does rise — but not from the first confirmation. It is flat between zero and three stacked signals (44%, 43.7%, 40.8%, 43.8%), then climbs: 51.1% at four, 54.1% at five, 61.6% at six. The first three confirmations buy nothing even on this measure.
The chance of being right about the side does not. Zero confirmations scores 53.8%; five scores 53.9%. Everything in between sits within a few points of a coin flip. The one row that looks different is six signals at 57% — on 86 observations, where the interval is wide enough to cover the coin flip comfortably. The two deepest rows fall far outside the band, but at n=18 and n=5 they carry no information, and we are not going to pretend otherwise.
Confluence is a filter for activity, not for direction. If your system stacks signals to decide whether to engage, this sample supports you. If it stacks them to decide which way, this sample offers nothing.
The same discipline, turned inward
We built a rule engine on top of these situations and tested it against 30 days of Binance archive. Expectancy came out at −0.37R per trade. There was no edge, so we switched the delivery off instead of shipping it.
It belongs in this article for two reasons. It is the same measurement pointed at our own work, and a study that only ever confirms its author is not worth reading.
Limits
- One venue. Binance USD-M perpetuals. Cross-venue behaviour may differ.
- One grading definition. The horizon is 24 hours and the threshold is ATR-scaled; both are our choices, and different ones would move every number here.
- Rolling 30-day window. Regimes change, and verdicts move between a weekly and a monthly window for a sizeable share of types — a reason to distrust any single-month claim, this one included.
- Our detectors, our names. "Double bottom" means our detector’s definition of one. A different implementation produces a different population and possibly a different answer. This is the point of the compression example above.
- Multiple comparisons. 43 types tested at once; verdicts near the interval boundary will include some chance.
- Rates, not returns. Everything here is frequency, not expectancy. A 65% hit rate with poor payoff loses money, and nothing in these tables measures payoff.
- Dropped observations. In the implied-side table, cases where neither threshold was hit are excluded. The control is computed the same way, so the comparison holds — but the rate is conditional on something happening, not unconditional.
If you grade the same situations differently and reach a different answer, we would like to see it.