A closed market still has a price

The New York equity session runs for six and a half hours. The other seventeen and a half, and the whole of the weekend, the tape is blank. Every practitioner knows the price is not blank, only unobserved: the stock reports earnings at ten past four, the world reprices it overnight, and at half past nine the following morning the exchange prints a number that is already several per cent away from the last one.

There is now a venue where that repricing is visible. A perpetual futures exchange quotes a set of American equities and indices around the clock, including weekends. This note asks a narrow question with a practical answer. If you read that venue five minutes before the opening bell, how much of the gap do you already know, and what exactly is the number you are reading.

A perpetual contract on a share that is not trading

A perpetual future has no expiry. It is held to its underlying by a funding payment that flows between longs and shorts whenever the traded price drifts from a reference, and by the reference itself, which the venue publishes and which is assembled from outside sources. So the quote has two parts that can be separated and tested one against the other: the reference feed, and the premium the venue's own participants are willing to pay over it.

That separation is the whole method of this note. If the traded price predicts the opening gap and the reference feed does not, the venue is discovering something. If both predict it equally, the venue is a conduit. The test is cheap and it is decisive, and it is run in section 07.

Seventeen and a half hours in which nothing was recorded

Nvidia reported on the evening of 26 August 2026. The regular session closed at 209.97. The next morning it opened at 222.89, six per cent higher. Between those two prints our own database, which stores every one-minute bar the exchange publishes, holds nothing at all for the name. Not a thin tape. Nothing.

The venue holds two hundred and twenty three observations over the same hours, and they are not a straight line between the two points. The stock was flat for the first twenty minutes after the close, jumped on the release, drifted higher through the European morning, and gave a little back in the last hour before the bell.

The two price scales are not directly comparable, because a perpetual contract trades at a basis to the underlying share. What transports is the ratio, not the level. The last column below applies the venue's move to our own closing price, which is the only arithmetic performed anywhere in this note.

Read the last two rows together. At 09:25, with the exchange still shut, the implied price was 223.62. The stock opened at 222.89. The error is thirty three basis points on a gap of six hundred.

The night is not one event

The overnight move is usually treated as a single jump that appears at the bell. Split into the windows a US market participant actually recognises, it resolves into four distinct periods with quite different weights. The extended session to eight in the evening carries roughly a fifth. The deep overnight hours, when European and Asian desks are the only ones awake, carry the largest single share. The pre-market carries about a third. The opening auction itself contributes the remainder.

The blue segment is the honest floor on this exercise. Between eight and nineteen per cent of the total movement happens in the opening print itself, when the auction clears against orders that were never visible anywhere. No amount of overnight data reaches it, and any claim to predict the open exactly should be read against that number.

Take the close, apply the night, compare

The measurement is deliberately crude, because a crude measurement that works is more persuasive than a fitted one. Take yesterday's closing price. Multiply by the venue's move from that close to five minutes before the next bell. Compare with the price the exchange actually printed. No regression, no coefficients, no parameters estimated on the sample being tested.

The Micron line is the one to hold on to. A typical overnight gap of 312 basis points, which on a large position is a material overnight result, is known to within 46 basis points before the exchange opens. That is not an edge, because everyone with access to the same screen sees the same thing. It is a measurement that our own data does not otherwise contain.

Where we can see the night ourselves

Everything so far compares a venue that trades overnight with an exchange that does not, which means the venue cannot be wrong in any way we could detect. Two instruments break that symmetry. Index funds quote from four in the morning to eight in the evening, so for SPY and QQQ we hold our own pre-market record and can put the two side by side over the identical window.

A few basis points of disagreement over a five and a half hour window is what two venues quoting the same exposure through different plumbing should look like. It is the closest thing to a validation available, and it is the reason the single-stock results, which cannot be checked this way, are stated as measurements rather than as claims about accuracy.

The table is also the argument for doing any of this. Four of the five instruments have an unbroken overnight record on the venue. Three of them have none at all in a conventional equity data set.

The traders there are not the ones deciding

The quote separates into a reference feed and a premium over it. Running the same regression twice, once with the traded price and once with the reference feed alone, produces almost identical results. Placed in the same regression neither survives the other, because they are the same variable.

The differences are in the third and fourth decimal place of R-squared. The premium, which is the only component carrying information originated at the venue, adds nothing to the prediction of the opening gap on any instrument tested. The single exception with a t-statistic above two is SPY at the four o'clock pre-market open, and with thirty four observations and near-total collinearity that is a suggestion rather than a result.

The venue is not forecasting the open. It is repeating what an overnight equity market already said.

Which raises the obvious question of where a reference feed for a single American share obtains a price at three in the morning. The answer is that overnight equity trading venues exist, are consolidated by data vendors, and are simply absent from most institutional equity data sets, including ours. The value of the observation in this note is therefore not that a crypto venue has predictive power. It is that a cheap, public, continuously polled feed provides a serviceable window onto a market we do not otherwise carry.

Useful precisely where nothing else is

The equity index futures trade the same overnight hours and are carried by every institutional data set. Where a future exists, it should dominate, and it does. Where none exists, there is nothing to dominate.

On SPY the venue alone explains 96.4 per cent of the opening gap, and placed beside the front E-mini contract its coefficient turns negative and insignificant while the future's is strongly positive. The venue was never adding anything on the index; it was reflecting the same overnight consensus by a longer route.

On the single names the picture inverts, for a reason that is mechanical rather than statistical. There is no listed future on Micron. The E-mini tells you what the market did overnight, not what Micron did, and on the night the company reports those are entirely different questions. The venue's coefficient sits at one with a t-statistic above twenty three, and the future's at zero.

A result that passed every control and was still wrong

The obvious follow-up question is whether the venue leads or lags within the session, when both markets are open. Run as a five-minute cross-correlation with the equity index future, the answer arrived clean and strong: the venue led, with a t-statistic above twelve on every pair tested.

It also passed the negative control. Advancing the venue's series by one, two, three, five and seven whole days, which preserves the time of day and the intraday structure and destroys only the true relationship, dropped the mean absolute t-statistic from 7.92 to 0.91 and the share of significant cases from all of them to six per cent.

A perpetual contract does not lead the deepest futures market in the world by five minutes. When a result survives every control and remains impossible, the fault is in the measurement.

The honest conclusion is neither direction. The answer changes ownership within a single bar, so no intraday precedence can be measured in this data at all. It is recorded here because it is the kind of finding that publishes easily and should not, and because every level-based result in this note was re-run reading the venue five minutes before the event, which is immune to a one-minute ambiguity and is in any case the only actionable form: a price known at 09:30:00 does not predict the print at 09:30:00.

Install the window, do not trade the signal

The opening print follows the overnight venue almost exactly, and that fact is worth very little on index funds and a good deal on single names. The distinction is not statistical. It is that a Micron future does not exist, so for seventeen and a half hours a position in the name has no observable price in a conventional data set, and this one does.

Because the number is the reference feed rather than the venue's own participants, it should be treated as a data source and not as a view. It tells you where the stock is. It does not tell you anything the rest of the market does not also know, and the section above is the reminder of how easily a measurement error can be mistaken for the opposite.

How the measurements were made

Prices on our side are one-minute exchange bars; the venue is polled at five-minute cadence. A one-minute bar labelled 09:34 covers the interval beginning at 09:34, so its close is the price at 09:35, and our value at any instant is the close of the bar labelled one minute earlier. The opening price is the open of the 09:30 bar, which sits within three basis points of the official opening price on the daily record. The close of the first five-minute bucket, which is the more convenient choice, sits between forty two and eighty five basis points away and measures something else.

All returns are logarithmic and quoted in basis points. Levels are never correlated: a timezone error that produces a correlation of 0.028 in returns produces 0.986 in levels and passes an inattentive review. Empty five-minute buckets are dropped rather than filled, because a repeated price is a zero return and zero returns make the other side appear to lead. The venue is read five minutes before each event throughout.

The negative control is temporal displacement of one series by whole days rather than the more obvious choice of economically unrelated pairs, which fails: gold and the Nasdaq genuinely co-move, so an unrelated-pair control registers a false positive. Where models are compared, the sample is forced common across them; comparing one model's error over thirty four nights with another's over twenty seven compares nothing.

Statistics are computed on 35 sessions and 7 weekends between 10 July and 27 August 2026, gross of all costs. Sample sizes are stated wherever a figure is given.