The Hour Before Funding Settlement: A Real Effect That Costs More Than It Pays

Every eight hours, perpetual futures settle funding: one side pays the other. When the rate is high, that payment is not trivial, and traders on the paying side have a reason to close before the clock runs out. If enough of them do it at once, the price should drift against them in the last hour, and bounce back once the payment is behind them.

It's a clean, mechanical story, and settlement times are fixed and public. That makes it worth testing properly. This article walks through the test, the result, and why a statistically real effect can still be a losing trade.

The hypothesis, written down before the run

Settlements on Binance happen at 00:00, 08:00 and 16:00 UTC. For every settlement and every coin, we asked:

• Who pays? If funding is positive, longs pay shorts; if it is negative, shorts pay longs.
• The bet: one hour before settlement, take the side against the payers (short if longs pay, long if shorts pay) and exit exactly at settlement.
• Only when it matters: the pre-registered test covers events where the estimated funding rate was at least 0.10% per 8 hours, about ten times the default rate.
• Market-adjusted: each coin's move is measured relative to the average move of all eligible coins in the same hour, so a market-wide swing doesn't count as an effect.

Two more arms were fixed in advance. A control arm runs the identical rule at low funding (0.02–0.05% per 8 hours), to separate a funding effect from anything that simply happens at these hours. A post-settlement arm checks the hour after settlement in the reverse direction: the bounce.

Estimating the rate without peeking

The final funding rate is only fixed at settlement, so using it would be look-ahead. Instead we estimated it the way the exchange does, from the hourly premium-index readings, and stopped one hour before settlement, when the position would be opened.

Our first draft used the raw premium index and got the sign wrong in more than 70% of low-funding cases, because the rate is usually pinned at the default +0.01% even when the premium is slightly negative. We caught that before computing a single return and switched to the exchange's formula with its interest-rate clamp. After that fix, the estimate had the same sign as the final rate in 98.9% of the tested events. Fixing a design flaw before the run is allowed. Changing the design after seeing results is not.

The result

9,277 events at 2,711 different settlements, July 2020 to August 2026, across 371 perpetuals including delisted ones:

Basis points per event, market-adjusted. Net = gross minus a 10.5 bp round trip; double costs = minus 21.0 bp. Cluster t groups all events of the same settlement.

It is real. To test whether the pattern could be chance, we moved every settlement, with all its events, to the same hour on a random other day 7 to 90 days away, and recomputed the average. We did that 20,000 times. Not a single shifted run reached the actual +5.74 bp; the highest was +4.56 bp. The difference to the control arm is significant too (p = 0.0001), so it is the funding level that matters, not just the hour.

It grows with the funding rate. The more expensive funding is, the stronger the drift against the payers:

Why a real effect still loses money

A taker round trip costs about 10.5 bp. The effect is +5.74 bp. Even the top of its 95% confidence interval, 11.9 bp, only just clears the cost line, and the bottom sits at 2.2 bp. After costs, the average event loses −4.76 bp; with doubled costs, −15.26 bp. The bounce after settlement is a little larger, at +7.28 bp, and still nets −3.22 bp.

Trading only the most extreme funding looks tempting: at 0.20% or more, the hour before averages +13.6 bp over 3,013 events, above the cost line. But that threshold was chosen after seeing the data. Picking the slice that happens to clear costs is exactly the kind of search that produces backtests which fail live, so we did not adopt it.

Where the effect actually lives

Averages hide a lot here. Splitting the main arm shows that the effect is concentrated, and not quite where the story predicts:

• One side carries it. Almost all of the effect comes from negative funding, when crowded shorts are paying. The short side is close to zero.
• One settlement carries it. The 00:00 UTC settlement is strongest, while 08:00 UTC is indistinguishable from zero. If the driver were simply payers closing, all three settlements should look alike. A concentration at the daily rollover fits a day-change or rebalancing effect at least as well.

The time windows tell the same story. The first walk-forward window (2020-09 to 2021-11) contains 61% of all events and averages only +1.4 bp:

Removing any single coin leaves at least +5.2 bp; removing any single year leaves at least +4.1 bp. The first half of the sample averages +4.2 bp, the second +12.6 bp from less than a fifth of the events. The pre-registered criteria asked for more than the cost in both halves, in at least three of four walk-forward windows and after removing any single coin or year. Only the second half and the third window get there.

In stress episodes the picture is noisy, with few events each, and the worst one hit the side that carries the effect:

During the Luna collapse, buying coins with crowded shorts an hour before settlement lost −30.6 bp per event on average.

What we take away

1. Significant is not the same as tradable. A placebo p-value below 0.0001 and a net loss per trade can be true at the same time. Check effect size against costs before anything else.
2. Pre-register a control arm. The low-funding control earned +1.28 bp from the same rule at the same hours. Without it, we could not tell how much of the effect is about funding at all.
3. A plausible story is not a confirmed mechanism. "Payers close before settlement" fits the sign of the result, but not its concentration in one side and one settlement hour.
4. The slice that clears costs is a trap. Some sub-bucket will always look profitable after the fact. Choosing it is a new hypothesis that needs new data, not a conclusion.
5. Look at where the average comes from. One year with a few hundred events can carry a result that most of the sample doesn't show.

What this means if you trade around funding

Settlement hours are not quiet. When funding is expensive, a few basis points of drift against the paying side in the last hour are normal, and so is a partial bounce afterwards. Measured against a taker round trip of about 10.5 bp, though, that drift is noise from a trading point of view.

The same arithmetic decides whether a funding payment is worth collecting in the first place: a positive rate is only interesting if it clears the cost of getting in and out. That's the subject of The Fee-Coverage Math.

Methodology