Overnight: Why Does So Much of the Equity Return Arrive While the Market Is Closed?

The trading day is shorter than the day. A share listed in the United States has its price discovered continuously for six and a half hours, from the opening auction at 9:30 to the closing auction at 4:00 Eastern time, and then not at all for the remaining seventeen and a half. Companies do not stop reporting when the market closes, governments do not stop publishing, and markets elsewhere do not stop trading. Whatever happens in that interval is priced all at once when the market reopens. The return from one close to the next open, the overnight return, is therefore a different kind of object from the return from the open to the close, the intraday return, and the two can be studied separately.

When they are, they look nothing alike. A long academic literature has documented that, for broad equity indices and for most individual stocks, a large share of the long-run return has accrued between the close and the open, while the session that holds nearly all of the trading has contributed most of the variance and comparatively little of the drift.^[1] This piece sets out the decomposition, shows in a stylized model how the split moves as more information migrates outside regular hours, asks what a clock of information arrival looks like, and weighs the explanations that have been offered for the gap. None of them is sufficient on its own, which is itself informative.

Two Halves of a Day

The arithmetic of the decomposition is simple. A day's close-to-close return is the product of two legs: the overnight leg, from yesterday's close to today's open, and the intraday leg, from today's open to today's close. Compounding each leg separately over many days produces two cumulative series that together reproduce the full return. Nothing is lost and nothing is added; the decomposition only asks which part of the clock the return was booked in.

Figure 1 shows what the two series look like in a stylized model in which the drift is placed almost entirely in the overnight leg and the variance almost entirely in the intraday leg. The parameters are chosen to illustrate the shape reported in the literature rather than to estimate it, and the wobble is a deterministic sum of sinusoids rather than a draw from any distribution. The point is the contrast: one line climbs with modest fluctuation, the other fluctuates around a level it never really leaves.

Two things about the picture deserve emphasis. The first is that the intraday series is not dead. It moves a great deal; it simply does not go anywhere. The session is where the price is discovered, where liquidity is provided and consumed, and where most of the risk is borne. It is not, in this stylized account, where the compensation for bearing equity risk shows up. The second is that the overnight series is smoother than intuition suggests. The gap between close and open is where the largest single moves occur, since the backlog of a night's information is absorbed in one print, but those moves are rare, and between them the overnight leg carries, in the model, a small positive drift that compounds.

Where the Variance Lives

Return and variance can be decomposed on the same clock, and the variance decomposition is the more stable of the two. A classic strand of the literature observed decades ago that the variance of returns over hours when the market is open is much larger, hour for hour, than over hours when it is closed, and concluded that a good part of the variance is produced by the act of trading itself rather than by the arrival of public information.^[2] That is one of the reasons the intraday series in Figure 1 is so much noisier than the overnight one.

The split is not fixed. Over the past several decades a growing share of scheduled information has moved outside the session. Earnings releases that once arrived during trading hours now cluster before the open and after the close; the main macroeconomic releases are published before the open; and, for a US listing, a full trading day in Asia and most of one in Europe pass between one close and the next open. Figure 2 stylizes the consequence with a three-parameter model. Let f be the share of a day's information events that arrive while the market is closed, let k be the price impact of an event absorbed in the opening auction relative to one absorbed in continuous trading, and let τ be the variance generated by trading itself, the term the classic strand identified, which accrues to the session whether or not any news arrives. The overnight share of daily variance is then f·k divided by f·k plus (1 − f) plus τ. We let f rise by decade, hold τ fixed, and hold k slightly above one, on the reasoning that an event priced in a single auction, with no opportunity to trade against it in stages, moves the price by somewhat more than the same event priced over a liquid session.

The direction is the point, not the levels. If a growing fraction of information arrives while the market is closed, the overnight leg must carry a growing fraction of the variance, even while the session, with the variance that trading itself generates, keeps the larger share; and the character of overnight risk changes with it. Gap risk, the risk of a position that cannot be adjusted until the market reopens, becomes a larger part of total equity risk. A risk model that treats a day as a single interval will misstate the risk of a portfolio held through the close, and on this reasoning it will misstate it by more with each passing decade.

When Information Arrives

The split by decade is a coarse view. A finer one asks when, on the 24-hour clock, information arrives at all. Figure 3 stylizes the arrival intensity for a US-listed equity, hour by hour in Eastern time, as a floor plus a set of bumps: a broad, low bump through the Asian session; a somewhat higher one through the European morning; a sharp pre-open cluster of corporate releases; a narrow spike for scheduled macroeconomic data before the open; the opening minutes, in which the backlog of the night is priced; an elevated but flat level through the session; the closing minutes; and a large post-close cluster of releases. Bars are shaded by whether the hour falls within regular trading hours.

The construction is deliberately simple, and its shape is what matters. Under it, more than half of the day's information arrives while the market is closed, and the two largest arrival windows sit immediately on either side of the session: the 8:00 hour before the open and the 16:00 hour after the close. Prices, however, can be discovered only while the market is open. The mismatch between when information arrives and when it can be priced is the whole reason overnight returns exist as a distinct category. The opening auction and the minutes after it are where the night's backlog is cleared, and the closing auction is where positions are squared before the next backlog begins to accumulate. Both are the subject of companion pieces; here it is enough to note that the two ends of the session are, in an information sense, the busiest parts of the day.

A distinction hides in that description. Information arrival and price discovery are not the same thing. The clock in Figure 3 is a clock of arrival: when the facts become available. The clock of discovery is the session. Between the two sits the question of who is willing to hold a position through the interval, and at what price. That question is where the explanations for the return gap come in.

Four Explanations and a Skeptic

Why should the overnight leg carry the drift? The candidates fall into a few families, and Figure 4 lays them out along with what each predicts and where each strains. We find it useful to keep them side by side, because each explains part of the pattern, and the temptation is to stop at the first one that fits.

News timing is the most natural explanation for the variance split and the least satisfying for the return split. Announcements cluster outside the session, so the price reactions to them are booked overnight; but announcements are surprises in both directions, and a clock that tells us when surprises arrive does not tell us why the average surprise should be favorable. Liquidity provision goes further. Intermediaries who absorb order flow during the day hold inventory they would rather not carry across a gap they cannot trade through, and a rational intermediary demands compensation for doing so. That compensation is paid to whoever does hold through the close, and it shows up as a return between close and open. The flow explanation is related but distinct: it holds that different populations trade at the two ends of the session, that their net demand pushes prices in different directions, and that the overnight and intraday legs are in a persistent tug of war whose sign depends on which population is buying and which is selling.^[3]

The closed-market explanation reframes the same observation as a risk premium. A position held through hours in which it cannot be exited bears a risk that a position held only during the session does not, and a premium for that risk should accrue exactly across the gap. The prediction is testable in principle by comparing assets with near-continuous trading against those without, and the answer is a matter of data rather than argument. The skeptic's row is a reminder that the open and close are auction prints, that bid-ask bounce and stale quotes can move return from one leg to the other without any economic content, and that any decomposition built on those prints inherits their flaws.^[4]

Our own view is that the explanations are not competitors. News timing sets the variance clock. Liquidity provision and flows set the sign of the drift. The closed-market premium is the name given to that sign once it has persisted long enough to be treated as a feature rather than an anomaly. And measurement determines how much of what is observed is real. A research process that tests each hypothesis in isolation, expecting one to win, will keep finding partial support for all of them and concluding that the question is open. A process that treats them as a layered account will find that the layers fit together.

Two Clocks, Two Models

The practical lesson we draw is that a day is not one interval. It is two, with different sources of return, different sources of risk, and different populations of participants. We therefore find it cleaner to model the overnight and intraday legs separately than to model the close-to-close return and hope the mixture averages out. A signal may forecast one leg and not the other. A risk model built for a session, in which positions can be adjusted continuously, understates the risk of the gap, in which they cannot. And an execution decision is not complete until it says whether a position is to be carried across the close at all; a trade that looks reasonable on a close-to-close horizon can look different once the choice of which leg to hold is made explicit.

The deeper lesson concerns information. The clock of Figure 3 is a clock of arrival, and a research process that date-stamps information by when the price reacted to it, rather than by when it became available, will misattribute the night's arrivals to the morning's session and draw the wrong conclusions about who knew what and when. We try to keep the two clocks apart: to record when a fact arrived, when it could first have been traded on, and when the price finished absorbing it, as three separate timestamps. Much of what looks like a mystery in the overnight return is, on that view, an accounting question about which of the three clocks one is reading.

None of this settles why the market has paid for the night. It does suggest that the question is better posed as several smaller ones, and that the answers to those are within reach of anyone willing to keep the clocks straight.