For Bitcoin trades that were eligible to end in at least two zeros, the round-size share fell by 0.04 standard deviations at an ordinary minute opening and by 0.20 at the top of the hour, making the hourly effect five times larger.

A standard deviation describes how far an observation moves from its usual range, so those numbers aren’t the percentage of trades placed by machines. They show that the market moved farther from its normal preference for round quantities exactly when trading activity jumped, giving the authors a behavioral fingerprint of heavier automated participation.

Trade size still can’t identify the source of every order. Large institutional executions and forced liquidations can produce irregular quantities, and funding arbitrage can do the same, so the paper uses roundness as indirect evidence associated with machine activity.

The authors ran several checks to see whether another recurring event was creating the pulse.

Binance processed funding payments at 00:00, 08:00, and 16:00 UTC during the sample, but removing those windows left the quarter-hour result largely intact, and the pattern at minutes 15, 30, and 45 survived when every top-of-hour observation was removed. A separate analysis of Bybit data produced a similar structure on another exchange.

Those checks describe a broad form of electronic coordination. Any trader can choose any interval, but exchange data, chart settings, and common indicators pull many systems toward the same boundaries, with the strongest concentration appearing at the clock points that receive the most shared attention.

A price forecast worth less than the trading fee

Once the researchers established that the pulse repeated, they asked whether data available before each quarter-hour could forecast the price move during its first ten seconds.

Their rolling model studied earlier quarter-hour returns alongside familiar price and volume indicators, then made a fresh out-of-sample forecast using information available at the time.

Across the six contracts, the model chose the correct direction 56.6% of the time. Its average out-of-sample R-squared was 3.4%, meaning it explained a small portion of the variation in those ten-second returns, while its area-under-the-curve score was 0.60 on a scale where 0.50 is a random guess, and 1.00 represents perfect classification.

In a market with enormous noise over ten-second intervals, those modest figures establish that the pattern contains repeatable information.

However, they don’t establish an easy trading strategy because the predicted move was tiny. Trading in the model’s chosen direction at every quarter-hour produced an average gross return of 0.51 basis points per trade before fees, equal to about 0.0051%, or roughly 51 cents on a $10,000 trade.

During the sample, Binance’s base fee was 5 basis points for a taker order, which executes immediately against an existing quote, and 2 basis points for a maker order, which provides a quote for someone else to accept.

A $10,000 taker trade therefore cost about $5 to open and another fee to close, while the model’s average gross return was roughly one-tenth of the first charge alone.

Given how small the gains are here, the most useful result from this dataset is the gap between statistical predictability and the money an ordinary trader can capture.

A pattern can repeat often enough to survive formal analysis while the expected move stays too small to cover basic trading costs, which is one reason highly automated markets can contain recognizable patterns without making any profits.

Market makers and large traders can still use the finding because they face a very particular problem.

If a company is quoting both sides of the market, it could demand a wider spread during those ten seconds or reduce how much it offers when one-sided flow becomes easier to anticipate, while a trader working through a large order may release pieces at less crowded points on the clock to reduce the price movement caused by its own activity.

The first ten seconds also carried information over a longer horizon. When buyer-initiated volume exceeded seller-initiated volume at a quarter-hour boundary, that imbalance was associated with returns over the next four to 12 hours, and the reverse relationship appeared when sellers dominated.

Order imbalance here means the difference between aggressive buying and aggressive selling relative to the total volume in that window, giving the researchers a way to measure which side was pushing harder.

At the four-hour horizon, much of the relationship came from earlier quarter-hour flow carrying into later boundaries. At eight and 12 hours, ordinary price and volume indicators explained more of it, which fits a market where algorithms use the quarter-hour as a shared moment to process information that has already been building across the wider market.

That longer-horizon result needs to be taken with a grain of salt because the four-, eight-, and 12-hour return windows overlap, allowing one market move to appear in several observations.

The authors used block-bootstrap methods designed for dependent data, though aggregate trade records still can’t show whether the initiating orders contained private information, reacted to the same public inputs, or moved prices as market makers absorbed an uneven flow.

Nonetheless, the larger idea is easier to understand and eventually implement than the statistical machinery behind it.

Crypto removed the closing bell and made trading continuous, then its APIs, chart intervals, and automated strategies rebuilt miniature openings throughout the day.

Every 15 minutes, thousands of independent systems reach the same clock boundary, and for a few seconds a market designed to run without interruption behaves like a crowd pushing through the same door.