You have seen the claim: a fixed share of daily highs form in one particular hour. We measured it on ten crypto pairs and 1,825 complete days each. The honest answer is that the hour tells you nothing, and the shape everyone is looking at is arithmetic.
The daily high lands in the hour after midnight UTC on 15.2% of days, and in the hour before midday on 2.0%. The daily low behaves the same way: 14.4% and 1.7%. If the hour made no difference at all, every hour would hold 4.17% of them.
| UTC hour | Daily high | Daily low |
|---|---|---|
| 00 | 15.2% | 14.4% |
| 01 | 5.9% | 6.9% |
| 02 | 4.4% | 4.6% |
| 03 | 3.3% | 3.2% |
| 04 | 2.7% | 2.6% |
| 05 | 2.3% | 2.1% |
| 06 | 2.5% | 2.1% |
| 07 | 2.7% | 2.5% |
| 08 | 3.0% | 2.4% |
| 09 | 2.8% | 2.3% |
| 10 | 2.7% | 2.1% |
| 11 | 2.0% | 1.7% |
| 12 | 3.4% | 2.7% |
| 13 | 3.7% | 3.7% |
| 14 | 4.5% | 5.1% |
| 15 | 4.5% | 4.4% |
| 16 | 4.3% | 4.7% |
| 17 | 3.1% | 4.0% |
| 18 | 3.2% | 3.9% |
| 19 | 3.2% | 4.1% |
| 20 | 4.1% | 4.3% |
| 21 | 3.7% | 4.1% |
| 22 | 5.2% | 4.8% |
| 23 | 7.8% | 7.4% |
So far this looks like a strong pattern, and it is the point at which most people stop.
The line on the chart is not fitted to our data. It is the arcsine law: for a price that wanders at random, the highest point of any stretch is most likely to occur at the very start or the very end of it, and least likely in the middle. Simulated over 20,000 days, a random walk with no pattern in it at all puts its high in the first hour 13.1% of the time and in the last 12.9%, with a trough of 2.5% in the middle.
Our measured curve is that curve.
If midnight matters, the peak should stay at midnight when we cut the day somewhere else. So we cut the same prices into 24-hour days starting at other hours:
| Day starts at | Share landing in the first hour | Share landing at midnight UTC |
|---|---|---|
| 00:00 UTC | 14.8% | 14.8% |
| 06:00 UTC | 11.5% | 4.6% |
| 12:00 UTC | 13.6% | 3.9% |
| 18:00 UTC | 14.1% | 4.2% |
Midnight stops being special the moment the day stops starting there. What stays special is being first. The first hour of any 24-hour day holds the high between 11.5% and 14.8% of the time, whichever hour it happens to be.
They do, and barely. Hourly ranges are widest from 14:00 to 16:00 UTC and narrowest around 04:00, a difference of 1.6 times. Yet the daily high lands in those busy afternoon hours only 3.4% to 4.5% of the time, which is about the flat rate. Re-running the simulation with each hour carrying its real volatility moves the first-hour figure to 15.1%, against the 15.2% we measured.
Volatility across the day moves this number by about one point. Position in the day moves it by eleven.
A competitor publishes that 89% of daily highs and lows are set between 00:00 and 01:00 UTC. We measure 21.1% for those two hours.
The difference is what is being counted. Count the running high — the highest price so far — and the first hour of any day contains it 100% of the time, by definition, the second hour 46% of the time, the third 31%. Counts built that way can be true of several windows on the same day, which is why four of their windows that cannot overlap add up to 313% when a day has one high and one low and the total cannot pass 200%.
One thing in our own measurement is unexplained. The daily high lands in the final hour of the day on 7.8% of days, where a random walk says 12.9%, the same walk carrying real hourly volatility says 11.8%, and a model with volatility clustering and fat tails says 10.9%. This is not a finding. Each more realistic model has closed part of that gap, which is what happens when the rest of the gap is also our model falling short rather than the market doing something. It stays in our research notes until some model closes it.
The hour of the day does not tell you where the high will be. Anyone selling you that number should show you what it would be if the hour meant nothing — and ours is on the chart above, as the line.
Ten crypto pairs: BTC, ETH, SOL, BNB, XRP, ADA, AVAX, LINK, DOT and DOGE. Hourly bars, 1,825 complete UTC days each, 11 Aug 2021 to 11 Aug 2026. The rule, the null and the thresholds were written down and committed before the measurement ran. Running the same method on data with the hours deliberately scrambled produced zero findings, which is how we know the method does not manufacture them.