How often does Bitcoin reverse in the last 30 seconds of a 5-minute window?

Short answer: About one window in nine. In 8,633 clock-aligned 5-minute Bitcoin windows (27 Sep – 26 Oct 2026), the side that was ahead of the opening price 30 seconds before the close finished on the other side 11.2% of the time. With 10 seconds left it was 7.8%, and with 5 seconds left 6.4%. How far ahead the price is matters far more than the clock: a lead under $10 at T−30 s was overturned 31% of the time, one over $50 less than 1% of the time. Settling on 60-second averages, as Polymarket and Kalshi now do, halves the T−30 s rate to 5.7%.

Fixed-window up/down markets ask one question: will Bitcoin finish this 5-minute window above or below the price it opened at? Everyone who trades the window shares the same finish line, so the last seconds draw a crowd. This post measures how often those last seconds actually change the answer.

The numbers#

A flip is a window whose side at T−x (above or below the open) differs from its final side.

The random-walk column is the textbook baseline: for a driftless random walk, the chance that the sign at time s differs from the sign at the end T is arccos(√(s/T))/π. Bitcoin tracks it closely a minute out and flips more often than it in the final seconds (more on why below).

How we measured it#

• Data: Binance's BTCUSDT 1-second klines from its public data archive, 30 full days from 27 Sep to 26 Oct 2026 (2,592,000 seconds, none missing).
• Windows: aligned to the clock (:00, :05, :10 …), the way fixed 5-minute markets are cut. 8,633 windows, after skipping the first 35 minutes (the volatility measure needs 30 minutes of history).
• Prices: the window's opening price is the last trade at or before the open; the final price is the last trade at or before the close. The same rule gives the price at T−60, T−30, T−10 and T−5 seconds.
• Ties: 48 windows (0.6%) ended exactly at their opening price, and a few dozen were exactly at it at T−x. They are counted apart, not as flips. Venues settle a tie by their own rules.
• Reproduce it: bun studio/research/btc-reversals.ts in Rizon's repository rebuilds every number and chart here.

The lead matters more than the clock#

The flip rate at T−30 s depends mostly on how far the price already is from the open:

Over the last 30 seconds the price typically travelled $9 (median), and $36 at the 90th percentile. So a lead under $10 is close to a coin flip with a thumb on it, and a lead past $50 has needed an unusually large 30-second move to undo. About 38% of windows reached T−30 s with less than $25 between price and open: those are the windows where the end still decides things.

Volatility changes what a safe lead is#

The same dollar lead is safer in a calm market than in a fast one. We split the windows into thirds by realised volatility over the 30 minutes before each window opened (from 10-second returns, annualised): under 17%, 17–27%, and over 27%.

Overall flip rates barely change across the thirds (12.2%, 10.9% and 10.5% at T−30 s), because faster markets also build bigger leads. What changes is the meaning of a given lead. A $25–50 lead in a quiet hour flipped about one time in twenty; in a fast hour, about one time in nine.

Why the last seconds flip more than a random walk says#

At 10 and 5 seconds out, Bitcoin flipped about 2 percentage points more often than the random-walk baseline. Two things in the data account for it:

• Tiny leads. When the price is less than $1 from the open, trades alternating between the best bid and the best ask can move the last price across the line and back. Those windows flipped 22.6% of the time at T−5 s. Without them, the T−5 s flip rate drops from 6.4% to 6.1%, and the T−30 s rate from 11.2% to 10.7%.
• Fat tails. Measured in sigmas of the remaining 30 seconds (from trailing volatility), leads of 2 sigmas or more flipped 2.6% of the time, where a normal random walk gives 0.3%. Bitcoin's short-horizon moves have more large jumps than a bell curve, and volatility can change inside a window.

In 13.5% of windows the price came back to the open, or through it, at some second in the last 30, though only 11.2% ended on the other side. A line that gets touched isn't always a line that's crossed at the close.

What averaging the last minute does#

Fixed-window venues no longer settle on a single price. Polymarket's crypto up/down markets resolved on one Chainlink price at each end until August 2026; since 14 August 2026, both the price to beat and the final price of its 5-minute markets are a 60-second Chainlink TWAP (time-weighted average price), and Up wins ties (changelog, market rules). Kalshi's 15-minute Bitcoin markets compare the simple average of 60 seconds of CF Benchmarks' Real Time Index before the close with the same average before the open (rules). Polymarket's change followed research and reporting on large spot orders in the final seconds before settlement (CoinDesk, August 2026; Dai, Jia and Yu, 2026).

We reran the same 8,633 windows with a 60-second average at both ends (the mean of the 60 one-second prices up to each moment):

Averaging cuts late reversals hard: 30 seconds out, half of the final average is already fixed, and the side ahead on the average lost 5.7% of the time, half the single-price rate. Five seconds out it almost never lost (0.9%). The catch is which number you watch. A minute out, the last trade is the better guide, because the final average hasn't started. In the last seconds the average is, and the last trade misleads: at T−5 s it pointed the wrong way 5.4% of the time, because the average lags the price.

What this means if you trade fixed windows#

• The end is where the contest is. Every trader in a fixed window shares one close, so the final seconds concentrate the decisions, the order flow and the automated strategies that trade the last moments.
• The price of the leading side already reflects the lead. Near the close, a window with a $60 lead isn't a 50/50 question anymore, and its price shouldn't be. What's left to trade is the small set of windows where the lead is thin.
• Your feed isn't the resolution feed. These numbers use Binance's last trade. Polymarket settles on Chainlink's BTC/USD TWAP stream and Kalshi on CF Benchmarks' index, and in close windows a few dollars of difference between sources decides the side.

On Rizon, contracts don't share a close. Each one expires a fixed time after its own fill (15 seconds, 1 minute or 5 minutes), and its strike starts next to the price when you trade, so nobody arrives at a window that's already mostly decided. The data above still applies to every contract's own last seconds: a thin lead near expiry is closer to a coin flip than it looks.

FAQ#

How often does a 5-minute Bitcoin window reverse in the last 30 seconds?#

In our sample of 8,633 windows from 27 Sep to 26 Oct 2026, 11.2% (95% interval 10.6–11.9%). In the last 10 seconds it was 7.8%, and in the last 5 seconds 6.4%.

Is the side that's ahead with 5 seconds left safe?#

Usually, not always. It lost 6.4% of the time overall, and much more often when the lead was under a dollar (22.6%). With a lead over $50 at T−30 s, a reversal happened in under 1% of windows.

Does volatility make late reversals more likely?#

For a given dollar lead, yes: a $25–50 lead at T−30 s flipped 4.8% of the time in the calmest third of windows and 11.2% in the most volatile third. Across all windows the rate barely changes, because volatile markets also build bigger leads.

Which price feed did you use, and does it matter?#

Binance's BTCUSDT 1-second klines (the last trade in each second). Prediction markets settle on their own feeds, so their exact flip rates differ. Close windows are the most sensitive to the feed.

Do 60-second averages change the answer?#

Yes. With a 60-second average at both ends, the side ahead on the average at T−30 s lost 5.7% of the time (single prices: 11.2%), and at T−5 s 0.9% (single prices: 6.4%).