Thirteen Percent of the Volume, a Third of the Gap: What the Overnight Tape Is Worth at 04:00 ET

TL;DR - US equities now trade a genuine overnight session - 20:00 to 04:00 ET, on the Blue Ocean ATS - and it is invisible on almost every retail chart, which begins at 04:00. Across 34,320,713 overnight prints in 28 US names over 240 sessions, $299.9bn changed hands overnight against $55.2tn in the regular session. That is 0.35% of a name's regular-session dollars on the median night, and 13.8% of everything traded between the 20:00 open and the bell. The size varies enormously and predictably by name: 3.44% of regular-session dollars in Alibaba down to 0.02% in JPMorgan, a 172x spread that follows which time zone actually owns the stock. That thin tape does real work. Measured leg by leg, the mean close-to-open gap of 132.4 bp is built 50.5 bp in the post-market, 80.4 bp in the overnight window, 85.8 bp in the pre-market and 7.3 bp in the auction; on the median large gap the overnight window alone carries 34.8% of it. The price of doing that on so little volume is brutal and measurable: a five-minute bar moves 9.67 bp per $1m traded overnight against 0.42 bp at midday - 23x - so $1.0m moves a name 1% at 02:00 where $23.8m is needed at 13:00 and $60.1m into the close. Then we tried to trade it. Fourteen distinct rules, each at three to four thresholds and four horizons: fade the overnight move at the open (+3.42 bp, t = 0.67), fade the whole gap, fade the pre-market leg, trade the nights the morning overturned, fade the night into the pre-market, and treat the overnight high and low as levels - which held 50.0% of the time against a distance-matched placebo's 50.4%. All flat, against a 8.7 bp round trip in half-spreads. One specification did light up: gaps made on a busy night reversed and gaps made on a quiet night continued, a 113.5 bp spread at t = 6.67. It was our own look-ahead - the denominator was that day's regular-session volume, which you only know at 16:00. Point-in-time, the same sort gives 20.4 bp at t = 1.22. What survived every check is not a direction signal at all. Overnight volume against its own trailing 20-night median forecasts the day's range - rank correlation +0.230 to the full session and +0.263 to the first hour, positive in 28 of 28 names, in both halves of the sample. Quietest quintile: 0.89x the name's normal range and a 19.2% chance of a big day. Busiest: 1.16x and 43.8%. You have that number at 04:00 ET, five and a half hours before you have to size anything.

There is a session in US equities that most people trading US equities have never seen.

It runs from 20:00 to 04:00 Eastern, five nights a week, on the Blue Ocean ATS, and it has quietly stopped being a curiosity. In our sample it printed 34.3 million trades. Every major US retail broker now routes to something like it. And it is missing from almost every chart, because the standard extended-hours chart starts at 04:00 - which is precisely when this session ends.

That makes it interesting for the obvious reason and the non-obvious one. The obvious reason is that if a price is being discovered where nobody is looking, there might be something in it. The non-obvious one is that if you trade from Hong Kong, Singapore, Tokyo, Dubai or Riyadh, this is not the middle of the night. It is your working day.

So we pulled it apart.

The session, measured

Across 28 US names over 240 sessions - 2025-09-02 to 2026-08-14, mega-cap tech through to China ADRs, crypto proxies and four deliberately boring large caps - the overnight tape carried $299.9bn against $55.2tn in the regular session.

Per name-night the median is $12.55m overnight against $2,975m in the regular session: 0.35%. Measured against everything that trades between the 20:00 open and the 09:30 bell, the overnight window is 13.8% of the shares and 13.1% of the dollars. The rest belongs to the pre-market, and most of that belongs to the last ninety minutes of it.

The tickets are small. The average overnight print is 36.3 shares against 64.5 in the regular session - this is a tape made of odd lots.

The size of the session is not uniform, and the variation is not random:

Name	Overnight dollars, % of its regular-session dollars
BABA	3.44%
NIO	1.23%
MSTR	0.92%
IONQ	0.91%
NVDA	0.44%
SPY	0.14%
KO	0.06%
XOM	0.04%
WMT	0.03%
JPM	0.02%

That is a 172x range across a single panel, and it sorts almost perfectly by who is awake and who cares. Alibaba trades 3.44% of its US volume overnight because the people with the strongest opinion about Alibaba are at their desks. JPMorgan trades 0.02% because nobody outside New York is repricing a US money-centre bank at 23:00.

The night itself has a shape. 56.9% of overnight volume prints between 20:00 and 00:00 ET - the Asian working day - and 43.1% between 00:00 and 04:00, as the Gulf comes in and London opens. The single busiest hour is the first one.

What that thin tape actually builds

Here is the part that matters to somebody who never trades overnight at all.

Take the close-to-open gap - the thing you are handed at 09:30 and have to make a decision about. Decompose it into the four windows that build it, measured leg by leg from each window's own endpoints. Mean absolute contributions across the panel:

Window	Mean absolute contribution	Median share of the final gap
Post-market, 16:00-20:00 ET	50.5 bp	9.9%
Overnight, 20:00-04:00 ET	80.4 bp	34.8%
Pre-market, 04:00-09:30 ET	85.8 bp	46.3%
The opening auction	7.3 bp	0.0%
The gap you trade	132.4 bp	100%

A third of the gap is built in a window carrying an eighth of the pre-open volume, by participants on the other side of the world, while the desk that will trade it is asleep.

That is the fact worth internalising even if you never place an overnight order. The gap is not made in the pre-market. The pre-market mostly confirms a decision that was taken hours earlier.

The price of trading there

If a third of the gap is built on 13.8% of the volume, the arithmetic has an unavoidable consequence: price impact per dollar overnight has to be enormous. It is.

Take every five-minute bar, measure the absolute price move in it and divide by the dollars that traded. Median across the panel:

Window	Price move per $1m traded	Dollars to move the price 1% in five minutes
Overnight, 20:00-04:00	9.67 bp	$1.0m
Pre-market, 04:00-09:30	5.57 bp	$1.8m
Post-market, 16:00-20:00	2.07 bp	$4.8m
The open, 09:30-10:00	0.50 bp	$20.1m
Midday, 12:00-14:00	0.42 bp	$23.8m
Into the close, 15:30-16:00	0.17 bp	$60.1m

Twenty-three times the impact of midday, and fifty-seven times the impact of the closing half hour.

Two things follow. The first is a cost warning: a million dollars moves a liquid US large cap a full percent at 02:00 ET. If you are working real size overnight you are not taking liquidity, you are making the print. The quoted spread agrees - 10.17 bp at our 03:59 marks, against 4.55 bp at 10:30.

The second is the more interesting one, and it cuts against the intuition. The obvious read of a thin, high-impact tape is that its prices are junk - a few small orders pushing an unwatched book around, to be handed straight back when real liquidity shows up. That is a testable claim. It is also wrong.

Fourteen ways of being wrong about direction

We built the trade that the thin-tape story implies, and then we built thirteen more.

The discipline is the same one we apply to our own signals:

• Every return is measured in excess of the cross-sectional median of the panel over the identical window, so nothing here is a disguised market bet.
• Every conditioner is point-in-time. Nothing in the sort may use a number you could not have had at the moment of the decision. We violated this once, on purpose-by-accident, and it is the most instructive result in the study.
• Every rule is compared against the right null, which is almost never zero. The control for "fade the overnight move" is "fade the move regardless."
• Costs are real: the median quoted spread is 12.83 bp at the open and 4.55 bp at 10:30, so an open-to-10:30 round trip pays about 8.7 bp in half-spreads before anything else.

Here is what came back.

Fade the overnight move at the open, exit 10:30. The core trade. At a one-sigma overnight move: +3.42 bp, n = 1,494, t = 0.67. At 1.5 sigma it reaches +16.49 bp (t = 2.24), and at 2 sigma it falls back to +18.73 bp (t = 1.91) - a shape that is what a threshold sweep produces from noise, not a signal. Split by date, the two halves give +1.51 bp (t = 0.20) and +4.77 bp (t = 0.69).

Fade the whole close-to-open gap instead. Negative at every threshold - -11.17 bp at one sigma (t = -1.93) - meaning gaps in this panel mildly continued. Whatever the overnight leg is doing, it is not the classic gap fade.

Fade the pre-market leg. Also negative: -7.05 bp (t = -3.19) unconditionally. The 04:00-09:30 move continues into the first hour. That is the closest thing to a directional effect anywhere in the study, and it belongs to the visible session, not the hidden one.

Trade the disagreement. The night and the morning are two different populations. When the morning overturns what the night decided, somebody with more capital has arrived with an opinion - so trade with the morning. Both legs moving at least one sigma, disagreeing: -15.56 bp, n = 230, t = -1.10. Agreeing: -1.99 bp, t = -0.11. Nothing.

Fade the night during the pre-market, entering at 05:15 and covering at 09:15 so that no shared price can manufacture the result: -4.67 bp at one sigma (t = -1.29). The sign says the night's move mildly persisted into the morning. Thin does not mean wrong.

Condition the fade on how thin the night was. Quartiles of overnight participation, one-sigma moves: -5.28, -7.20, +6.83, +19.33 bp, top t = 1.75. Nothing that clears a spread.

The levels nobody can see

Then the structural idea, which we liked more than the directional ones.

62.0% of sessions in the panel end the night with at least one overnight extreme outside the pre-market range - a high or a low that a chart starting at 04:00 physically cannot draw. If price levels matter because participants see and defend them, a level that nobody can see should do nothing. If instead it marks where real supply or demand sat, it should still bite.

So: take every session where the overnight high sits above the open and above the pre-market high, within 2% of the open. Walk the first hour. Find the first touch. Measure the next thirty minutes.

Against a distance-matched placebo - the same measurement taken the first time price travels the same distance from the open on sessions with no level anywhere near it:

Level	n	30 minutes after the touch	Held
Overnight extreme, invisible on a normal chart	1,401	+3.64 bp (t = 1.08)	50.0%
Overnight extreme, inside the pre-market range	2,788	+2.42 bp (t = 0.89)	50.3%
Pre-market extreme - the level everyone draws	6,632	+0.49 bp (t = 0.26)	50.8%
Placebo: same distance, no level there	4,074	+1.83 bp (t = 0.60)	50.4%

The invisible level minus the placebo is +1.81 bp, t = 0.40. Matched inside four buckets of distance from the open, the differences are -4.44, -10.80, +6.21 and +1.29 bp, every t below 1.4.

Note the row that should bother you more than ours. The pre-market high and low - the levels drawn on every day trader's chart in the world - held 50.8% of the time against a placebo's 50.4%. In this panel, at this horizon, they are also a coin.

The one that looked like a discovery

And then something worked.

Sort one-sigma gaps by how busy the overnight session was and the continuation splits apart: gaps built on a quiet night carried on into the first hour, gaps built on a busy night reversed hard. Quartiles, in basis points of continuation to 10:30: +61.60, +34.54, -2.56, -51.92. Spread quietest minus busiest: 113.52 bp at t = 6.67.

It survived the checks we ran next. Both halves of the sample. Every horizon from 10:00 to the close. Every gap-size tercile in a double sort. Within-symbol ranking, so no single name drove it. The regression put the effect on the night's participation, not on gap size.

It was our own look-ahead, and it was sitting in the denominator.

"How busy was the night" had been computed as overnight dollars divided by that day's regular-session dollars - a number you do not have until 16:00. A day that continues its gap is a day with heavy regular-session volume, which shrinks the ratio and lands the session in the "quiet night" bucket by construction. The sort was a bet on today's volume wearing a disguise. That bet is real and it is strong - sorting on today's regular-session volume against trailing ADV gives a -131.23 bp spread at t = -7.63 - and it is completely untradeable at 09:30.

Replace the denominator with something you actually possess at 04:00 and the result dissolves:

How "a busy night" is measured	Q1	Q2	Q3	Q4	Spread	t
Overnight $ / that day's regular-session $	+61.60	+34.54	-2.56	-51.92	113.52	6.67
Overnight $ / its own trailing 20-night median	+15.14	+37.14	-5.25	-5.25	20.39	1.22
Overnight $ / trailing 20-day regular-session ADV	+24.04	+33.91	+3.27	-19.44	43.48	2.42

The rank correlation between the contaminated ratio and that day's regular-session dollars is only -0.093. That is the part worth remembering. The contamination was not visible as a large marginal correlation between the sort variable and the outcome; it lived in the structure of the ratio, whose variation is dominated by its denominator. A correlation check would have passed it. Only asking "when do I know each number in this expression" caught it.

Every ratio you sort on has a denominator. Date it.

What actually survives

Strip out everything that failed and one thing is left standing, and it is not a direction signal. It is a range forecast, and it arrives early.

Take a name's overnight volume and divide it by its own median over the previous 20 nights. That is a pure point-in-time number: it is complete at 04:00 ET, it uses no information from the session it predicts, and the normaliser is trailing.

Against the day's regular-session range, expressed as a multiple of that name's own trailing 20-day median range:

Overnight volume vs its 20-night median	n	Median day range	Median first-hour range	Big day (>1.25x)	Dull day (<0.8x)
Q1 quietest night (0.42x)	1,232	0.89x	0.89x	19.2%	40.0%
Q2 (0.69x)	1,232	0.94x	0.93x	23.9%	33.9%
Q3 (0.99x)	1,230	1.00x	1.01x	29.8%	28.2%
Q4 (1.45x)	1,232	1.05x	1.07x	32.7%	24.5%
Q5 busiest night (2.71x)	1,232	1.16x	1.20x	43.8%	17.1%

Monotone in every column. Rank correlation +0.230 against the full session's range and +0.263 against the first hour's, n = 6,158. The odds of a big-range day go from 19.2% to 43.8% - a factor of 2.29 - and the odds of a dull one from 40.0% down to 17.1%.

It holds up where the direction results did not:

• Positive in 28 of 28 names, from AMD at rho +0.38 to KO at +0.06.
• Both halves of the sample: +0.258 and +0.194, with the quintile ladder running 0.89x to 1.18x and 0.91x to 1.12x.
• Overnight trade count (+0.208) and overnight range (+0.221) carry nearly the same information, so it is not an artefact of one measurement.
• It is a modest effect honestly stated: R² 0.058 against the day's range and 0.078 against the first hour's, symbol-demeaned. This shifts a distribution. It does not tell you what today is.

The reason we care about it more than the R² suggests is when you get it. At 04:00 ET the pre-market has printed almost nothing - the 04:00-09:30 window does 86% of its volume after 07:00, and 35% of it in the 08:00 hour alone. The overnight number is the only real read on today's activity that exists five and a half hours before the bell, and for a desk in Singapore, Dubai or Riyadh it lands in the middle of the working day rather than at 4am.

One thing it does not do is pick your strategy for you. We tested that directly, because it is the obvious next step: opening-range breaks, 09:30-09:45 range, first break between 09:45 and 11:00, follow-through measured forward from the break level. Base rate across 5,667 breaks: 48.4% reached half an opening range before falling back through the range midpoint. Busiest-night quintile 49.2%, quietest 45.5%, a 30-minute follow-through difference of +5.23 bp (t = 1.05). The direction of that tilt repeats in both halves of the sample - 49.4% vs 46.2%, then 48.9% vs 44.8% - which is worth noting and not worth trading on its own. A busy night tells you how far price will travel. It does not tell you whether it will travel in a straight line.

How the desk uses it

• Size from the night, not from yesterday. A quintile-one night says today is likely to be a 0.89x day with a 40% chance of being genuinely dull; a quintile-five night says 1.16x with a 44% chance of a big one. That is a stop-distance and a target-distance decision, and a position-size decision, and you can make all three at 04:00 rather than at 09:45.
• Read the gap's birthplace before you read the gap. A third of it was built between 20:00 and 04:00. Before you form a view on why a name is up 2%, look at when it went up 2% - a gap that was in place by 22:00 ET and drifted sideways for six hours is a different animal from one that appeared at 08:15 on a headline, even though both print the same number at 09:30.
• Do not take size overnight, and do not read the overnight price as a price you can transact at. A million dollars moves a liquid large cap a full percent at 02:00. The overnight quote is real information and a poor execution venue, and the two facts are the same fact.
• Date every denominator. Before any sort, write down the timestamp at which each number in the expression becomes knowable. The single most convincing result in this study - t = 6.67, monotone, out-of-sample, robust to four other cuts - was created entirely by a denominator that arrives at 16:00. A correlation check did not catch it. An audit of when did.
• Check what your own levels are worth. The pre-market high and low held 50.8% against a placebo's 50.4% in this panel. If you are building a trade around a level, you owe it a matched null - price reaching the same distance with no level there - before you let it set a stop.

The takeaway

The overnight tape is not the noise the thin-volume story predicts. Fourteen ways of trading against it came back flat, including the direct fade at every threshold and the mirror-image trade of trading with the morning that overturned it. Whatever gets decided between 20:00 and 04:00 ET on 13.8% of the pre-open volume mostly stands, all the way through the pre-market and into the first hour.

What it will not do is tell you which way. It will tell you how far - and it will tell you at 04:00, which is early enough to change how you size, where you put a stop and how far you are willing to hold for, hours before the pre-market has printed enough to have an opinion.

The night is not a signal about direction. It is the earliest honest read you get on the size of the day.

And the most expensive number in the study was the one that agreed with us. A 113 bp spread at t = 6.67, monotone across quartiles, stable out of sample, robust to four separate controls - and worth nothing, because one of its inputs had not happened yet.

Related reads

The Night Shift · The Pre-Market Echo · The 62-Session Mirage · The Comb in the Fourth Decimal · The Gap Map.

Joining the desk

If your first instinct on reading "t = 6.67" was to ask what was in the denominator, you think the way this desk does. The trader application takes about ten minutes; serious applicants hear back within five business days.

---

Methodology and sources: The panel is 28 US common stocks and ETFs (AAPL, AMD, AMZN, AVGO, BABA, COIN, GOOGL, HOOD, IONQ, JPM, KO, MARA, META, MSFT, MSTR, MU, NIO, NVDA, PDD, PLTR, QQQ, RKLB, SMCI, SOFI, SPY, TSLA, WMT, XOM) over the 240 regular sessions from 2025-09-02 to 2026-08-14, giving 6,720 symbol-sessions. Overnight data is the Blue Ocean ATS tape retrieved from Alpaca Market Data v2 (/v2/stocks/bars, feed=boats) in five-minute bars; regular, pre- and post-market data is the consolidated SIP tape from the same endpoint (feed=sip), and quoted spreads are single NBBO marks from /v2/stocks/quotes sampled at 09:30:00, 10:30:00 and 03:59:00 on 20 randomly chosen sessions across all 28 names (1,680 marks). The overnight session is defined on the Eastern clock as 20:00 on session D-1 through 04:00 on session D and is therefore DST-correct by construction; the mapping was validated by checking that the last overnight print and the first pre-market print agree to a median of 0.7 bp. Bars are split- and dividend-adjusted. The four gap legs are measured from each window's own first and last prints, so venue handoffs are attributed to the window that receives them. Returns are in basis points and, except where explicitly labelled raw, in excess of the cross-sectional median return of the panel over the identical window. Price impact is the median across five-minute bars of the absolute open-to-close move in the bar divided by the dollars traded in it, which is a descriptive ratio rather than an estimated Kyle lambda and should not be read as an execution cost curve for any specific order. The level test defines a touch on five-minute bar extremes, restricts levels to between 0.05% and 2% from the open, requires the touch inside the first hour and at least 30 minutes of subsequent bars, and matches each event against a synthetic level drawn uniformly between 0.15% and 1.5% from the open on sessions where every real level is at least 0.3% away. Opening-range breaks use the 09:30-09:45 range with the first break between 09:45 and 11:00, and follow-through is measured forward from the break level, never from the open. Sigma thresholds are within-symbol standard deviations of the relevant leg. Fourteen distinct trade rules were tested, most at three to four thresholds and four horizons; no correction for multiple comparisons has been applied, which is why the isolated 1.5-sigma result at t = 2.24 is reported as a threshold-sweep artefact rather than a finding. Confidence intervals are normal-approximation on the per-trade distribution. A sample of 240 sessions cannot exclude a small directional effect; what it excludes is one large enough to clear the 8.7 bp round trip measured here. The look-ahead described in the text was ours, was found during this study, and is reported because the corrected result is the useful one. Compiled from public market data - VCG Research.

#MarketMicrostructure #OvernightTrading #PriceDiscovery #GapTrading #Liquidity #PositionSizing #QuantResearch #Backtesting #DayTrading #PropTrading #VortexCapitalGroup