The overnight gain is real, no trade keeps it, and the rule that chases it buys wrecks

Buy at the close, sell at the open, sleep through the money. It has been in the journals since 2008 and in the newspapers since 2018. Two funds sold it to the public in 2022 and closed in 2023, and it keeps coming round with a table of small stocks whose overnight gains run to thousands of times, and a rule for picking them. Figure 1 is the whole mechanic: one night of the trade, and the two trades that pay for it.

What I did. Three censuses, with no registration: SPY split into its night and its day since 1993; every member of the S&P 500 split the same way since 2000, 10,921 name-years; and 1,046 US stocks under 300 million dollars split the same way since 2020. Then eight tests on the S&P 500, each 26 years long, July 2000 to July 2025, each year registered on the platform before it ran: every member of the index bought at every close and sold at every open, and the rule that picks the strongest recent night movers, at five trading costs from none to the estimated bid-ask spread, against holding the same names. The rule: every week, take each name's overnight gain over the last ninety sessions, keep the names that gained, weight them by the fourth power of that gain, and hold that book every night. It averages 15 names, the largest of them 36 percent of the book.

How to read the numbers. A year here is July to July. A trade is one buy or one sell. The costs are assumptions. Each is paid on every buy and every sell: nothing; 0.01 percent a trade; 0.05 percent a trade, the figure the claims usually assume; half a cent a share, what a per-share broker charges, which is a quarter of a percent on a two-dollar stock and the 0.01 percent floor on anything above fifty dollars; and the bid-ask spread estimated from each day's high and low, which overstates for liquid names and is the dearest of the five. Where a dollar's growth is quoted it is the 26 years chained, each year from its own registered run.

The universe is the S&P 500 as it stood on each first of July from 2000 to 2025, reconstructed from the vendor's change log, held fixed for the year that follows; a name that leaves the index in the year stays in the book until its series ends. The bars are one frozen table of 1,013 split-adjusted daily series, pulled once and never refreshed; SPY's dates are the calendar. On average 38 member-years a July have no series at the vendor and are counted as such. A dividend is paid on the night when the ex-date is the next session, from the payment record, never from an adjusted close.

The rule as walked: at the anchor and every seven sessions after it, each name's overnight growth over the last ninety sessions, R, is the product of open over prior close; names with R above one score (ln R) to the fourth power, weights are the scores' shares, weights under one percent are dropped and the rest renormalised. Only a name that printed at that close can be bought at it, and the ninety sessions are the calendar's, so a series that ended is not scored on its last bars. Held overnight means every name of the book is bought at the close and sold at the next open, every night, flat during the day, the cost paid on both trades. Every member means the same, equal weight across every name with a print at that close. Bought and held means equal weight at the anchor close to the end of the year, dividends reinvested, 0.05 percent once at entry and once at exit. The rule holds 15.3 names a night on average against 410 for every member.

Five costs, one test each. Nothing; 0.01 percent a trade; 0.05 percent a trade; half a cent a share, where a trade pays the larger of 0.01 percent and half a cent at the price that printed, recovered through the split record, so a two-dollar stock pays a quarter of a percent a trade; and the spread, where a trade pays the larger of those and half the Corwin and Schultz (2012) estimate from two consecutive days' highs and lows, the median of the last 21 sessions' estimates. The last overstates for a liquid name: the same estimator gives SPY, whose quoted spread is a cent, 10 to 28 hundredths of a percent in the years to 2010 and a median of zero in most years since 2011. The closing and opening auctions fill at one price, so a small order pays none of these, only the auction's own move; the five costs bracket what a real fill would cost. The registration record in section 3 prints the half-cent setting by its floor, 1.0 basis point a leg; the half cent applied throughout.

The engine's reliability gate runs on every series over the ninety sessions before the anchor and the year: refused when any close-to-close move is up more than 80 percent, when three or more days move more than 60 percent, when a close is under a cent, when one log move is at least 1.0 and eight times the next largest, when the series has a hole inside its own span, when an overnight move is over 100 percent, or when an overnight move over 50 percent reverses more than 30 percent by the same close. Symbols the vendor has since handed to another company are refused by name from the engine's taxonomy, with the evidence beside each: EP, WM and BBBY from earlier studies, and from replaying these walks KMG (the vendor's series is KMG Chemicals, the seat was Kerr-McGee's), SII (Sprott's series, Smith International's seat), GENZ (an ETF's series, Genzyme's seat), RMG (a seat the log gives to Cablevision, which traded as CVC) and ETFC, whose split record the vendor left empty so its pre-2010 prints cannot be priced at the half cent. On average 3.5 names a year are refused. Fewer than ten names would exclude a year; none was.

Eight walks were registered before any ran, each 26 years, each year a prospectus, a run and a frozen report: the rule against every member at each of the five costs, every member overnight against holding, and the rule at half a cent and at the spread against holding. The record keeps every pick's five largest names and weights and the nightly equity of both arms, so Tables 2 to 4 are read from the reports; the same rule replayed offline from the same bars matched every year's book size and its gross return within the dividends. The benchmark of an overnight arm is SPY's own night; the family figure chains the years, each from its own run.

Beside the walks: the census of the night and the day for every member-year and for SPY (Table 5), the microcap census from the vendor's screener (Tables 6 and 7; the list is today's, so a name that died since 2020 is not in it), and the two NightShares funds over their life against their index's night on the same sessions (Table 5).

The funds that tried it (Table 5). In June 2022 two exchange-traded funds opened to do exactly this for the public: NSPY held S&P 500 futures and NIWM Russell 2000 futures from each close to the next open and sat in cash by day, sold on the record that the index's return comes at night, the record the next paragraph reads. Futures were the right tool, a fraction of the cost any stock trader pays, and the funds did what they were built to do. They tracked the night with a correlation of 0.91 and 0.96. The night they got was down while the day was up. NSPY returned -5.6 percent over its fourteen months against 18.2 for SPY held round the clock, NIWM -8.5 against 10.9 for IWM. About 36 thousand dollars a day changed hands in NSPY. Both closed in August 2023.

The night's gain is real (Figure 6, Table 5). SPY held only overnight grew 13.7-fold from 1993 to 2026, held only through the day 1.27-fold, held round the clock 17.4-fold. Before 2009 the night carried the whole return and the day lost half. Since 2009 both are up, 3.20-fold against 2.57-fold. The members show the same: equal weight, their nights compounded at 10.7 percent a year against 5.5 for their days in 2000 to 2009, 5.3 against 6.2 in the 2010s and 8.4 against 1.8 since 2020.

You cannot collect it by trading (Table 1). Every member bought at every close and sold at every open made 7.7 percent a year before any cost, and holding the same names made 12.1. A dollar became 5.99 against 14.88. Then the costs. At 0.01 percent a trade the night makes 2.4 percent a year and a dollar becomes 1.62. At 0.05 percent it makes minus 16.2, a dollar shrinks to less than a cent, and it lost money in 25 of the 26 years. At the estimated spread it is minus 50.0. The night's gain on the index is a few hundredths of a percent a night, and a buy plus a sell costs more than that.

The doubling table is the spread (Figure 5, Tables 6 and 7). Among 1,046 US stocks under 300 million dollars, the median one turned a dollar into 3.75 at night since 2020 and into 0.06 by day. Close to close it ended at 24 cents. In 86 percent of them the night beat close to close, and in 30 percent the night made ten times or more on a stock that ended below its start. The largest night gains, SCWO, BRTX, IQST, LODE and the rest, run to millions of times on stocks that went nowhere: twelve of the fourteen are below their start, and they carry 24 reverse splits between them. The bounce is largest in stocks that trade 99 to 608 thousand dollars a day and smallest in the ones that barely trade, because a day without a trade has no bid and no ask to bounce between. A stock cannot rise ten-fold at night and end lower unless it gives the night back every day; that is the last print at the bid and the first at the ask, counted twice.

The rule's whole result sits between two cost assumptions (Figures 2 and 3, Tables 1 to 4). At what a per-share broker charges, half a cent a share, it beat holding in 17 of 26 years, median 24.4 percent a year against 13.3, and a dollar became 480. At 0.05 percent a trade, the cost the claims assume, a dollar became 301. Both figures are mostly one year, July 2008, explained below. Take that year out and the two costs part company. At half a cent the rule still turns a dollar into 86.15 against 14.88 from holding, up in 20 of the 25 years and ahead of holding in 16. At 0.05 percent it turns a dollar into 10.81, ahead of holding in 9 of 25 years. The difference between the two costs is about a fifth of the money a year, and that is the whole margin. Split the walk at 2010 (Table 10) and the lead is almost all in the first half, a mean gap of 58 points a year before 2010 against 5 after it, and the first half is mostly the one year. After 2010 the rule and holding the same names are within a few points of each other a year, and the rule pays five hundred trades a year for that. The last ten years (Figure 3) are the plainest view. A dollar in the rule at 0.05 percent became 0.75, and in every member held overnight at that cost 0.24, while holding the same names made 3.36 and SPY 4.17. At half a cent the rule made 4.44, against 4.17 for SPY, ahead of holding in 6 of the ten years, on a book that leaned on Micron, Nvidia, AMD, Gilead, American Airlines, Coterra, Biogen, Broadcom and Oracle in turn (Table 3) with a third of the money in one name on an average night. That green line is the rule's best case, a per-share broker and auction orders, before tax. Five hundred trades a year, and after ten years it stands where SPY stands. The names it leaned on over the 26 years say what the rule is: Ciena at 2.22 dollars in 2004, JDSU at 2.68 dollars in 2002, Citigroup at 4 dollars in 2009, Office Depot at 5 dollars in 2010, and when the index had no wreck the year's biggest gapper, Nvidia in 2023, Broadcom in 2024, Oracle in 2025. Outside 2008 its median night is -0.03 to 0.27 percent, a drip. Neither cost line pays the bid-ask spread or the order's own push on the price. At the estimated spread the rule is above water in none of the 26 years.

What happened in 2008 (Table 4). The crash turned members of the S&P 500 into penny stocks, and a penny stock with news at night is what the rule is built to find. Freddie Mac closed June 2008 at 16.40 dollars. The government took it over on Sunday, September 7. It opened Monday at 2.50, closed at 0.88, and traded between 26 cents and two dollars for the rest of the year, ten to a few hundred million shares a day. The government's moves came at night and the stock gapped up on them: 97 percent at the open of September 19, the morning of the short-sale ban and the rescue plan; 49 percent on March 19, the morning after the Fed enlarged its purchases of the agencies' bonds; and the day sold each gap back. Over the year its night leg multiplied a dollar 135 times while close to close it lost 96 percent. The rule, which weights by the strongest ninety-session night gain to the fourth power, put 93 percent of the book into it on September 30 and held it at 60 to 96 percent into late January. Over those four months the stock fell from 1.71 to 0.59 and the rule's night leg on it made 4.6 times. Ford, Unisys and AIG did the same at smaller weights, and twelve of that year's nights made more than ten percent. At 0.05 percent a trade the year is 2,680 percent because 0.05 percent of a fifty-cent stock is a fortieth of a cent, a cost no one has paid; at half a cent a share it is 457 percent. Freddie Mac had left the index on September 10; the walks hold the July list for the year, and the limitations say what the live list gives.

For the index it is not worth doing. The night's gain on the index is the same money as holding, split differently: whoever holds a stock through the night gets the night for free and the day on top. Buying every evening and selling every morning keeps the night, drops the day, and pays five hundred trades a year for the privilege. On the S&P 500 the day was worth 4.5 percent a year on top of the night, and the trades cost more than the night paid. That is why the funds closed. They held the night of the index in a year in which the night was down and the day was up, at the lowest cost anyone can get, and there was nothing left over. The idea was sold on the long record of SPY, which is true, and delivered the one thing the record does not promise, that the night beats the day every year.

For the rule that picks the strongest night movers, the answer is the same, and its best case is the reason. A stock whose last ninety closes sat below its next opens is either quoted so wide that the bars alternate bid and ask, which is the microcap table, or falling all day and gapping back at the open, which is a stock in trouble. The rule cannot tell the two apart, and in the S&P 500 it lands on the second. Those stocks do pay at night: Berkman and his co-authors found the night's gain concentrated in exactly the volatile, watched names the rule picks, and the record here agrees, at half a cent a share, in 16 of 25 ordinary years. The rest of the story is the cost and the concentration. An account at a per-share broker, half a cent a share, that buys at the closing auction and sells at the opening one pays the half-cent line and would have had the green line in Figures 2 and 3. That is the rule's best case, and it ends the last ten years where SPY ends, 4.44 against 4.17, after five hundred trades a year with a third of the money in one name on an average night, before tax. Anyone paying the 0.05 percent the claims assume would have had the red one, below holding without 2008 and down 25 percent over the last ten years while holding made 3.4 times the money (Figure 3). A plainer version tried after the walks, the ten strongest names equal weight at 0.05 percent, made a dollar into 24, 22 of them in 2008, and 2.20 in the other 25 years together. A cap on any one name's weight is the next thing to test; the record says most of the lead is the concentration. The best case of the rule ends level with the man who did nothing, and the usual case loses money.

What the literature already knew. Cooper, Cliff and Gulen (2008) found the U.S. equity premium since 1993 earned overnight; Berkman, Koch, Tuttle and Zhang (2012) traced the night's gain in small, volatile stocks to retail buying at the open that is given back by the close; Lou, Polk and Skouras (2019) showed most anomalies earn their return either at night or by day, rarely both; Knuteson (2020) asked why the night of the whole market is so persistent. This study adds the costs. The night is real, a trade cannot keep it, and the rule that chases it buys wrecks.

What would change the verdict: a fill model from the tape, effective spreads and auction imbalances by name and date, would replace the five assumed costs with a number; a walk on the Russell 2000, where retail attention is larger and the spreads wider, would show where between the S&P 500 and the microcap list the bounce takes over. The Russell walk is the next one to run.

The two blocks this study runs on, the overnight book and the overnight hold, were written for it as custom primitives on the engine, the way an account holder writes one, and they stay on the platform with the frozen bars they read. A reader with an account can wire them to a universe of their own and run the five costs on it.

The costs are assumptions, not fills. The half-cent charge reads the price that printed through the vendor's split record, and where that record is empty the price is wrong; one such series, E*Trade before 2010, was found and refused, and the delisted names are where the record is thinnest. The cent is used throughout, though before 2001 most spreads were a sixteenth of a dollar, so the first two years are undercharged. The spread estimate overstates for liquid names by the amount stated in the method. On a two-dollar stock half a cent is a quarter of a percent a trade, the whole of a one-cent spread over the round trip, so on the names the rule leaned on most, the two-dollar banks of the crisis, the half-cent line charges at least what a fill costs and possibly more; that is in the rule's favour and is said here. A small order in the closing and opening auctions pays none of these and gets the auction price, which is the print; that is the no-cost row, and it holds for an account too small to move the auction in a two-dollar stock.

Membership is fixed at the anchor, so a name that leaves the index in the year stays eligible until its series ends; that is how Freddie Mac was held into 2009, twenty days after S&P dropped it, and the rule as it circulates has no index at all. The gate refuses a series with a hole in it but not a series that ends, so a takeover target is held to its last print and drops out at the next pick; between the two, up to six nights, its weight sits in cash.

The July list is a convention, and the record keeps it as registered. Replayed offline with each name leaving the index on the day of its removal, the rule holds Ford and then AIG in 2008 instead of Freddie Mac, and that year is 989 percent at half a cent. Over the 26 years the live list chains to 694 dollars at half a cent against 380 for the same replay on the July list, and to 153 at 0.05 percent against 241, so the convention understated the rule at the one cost and overstated it at the other; the replay carries no dividends, which is why its July-list run sits below the record's 480. Without 2008 the live list gives 63.70 and 6.74 against 68.59 and 8.72. The verdict does not move.

The microcap list is today's, so the names that died since 2020 are missing and the survivors' close-to-close figures carry the fall that made many of them microcaps; the split between night and day does not depend on that. The microcaps were censused and not walked, because the walk would be the bounce. No shorting anywhere.

1. Cooper, Cliff and Gulen (2008), Return differences between trading and non-trading hours: like night and day, working paper: the U.S. equity premium since 1993 is earned overnight; the trading day's return is flat to negative.
2. Berkman, Koch, Tuttle and Zhang (2012), Paying attention: overnight returns and the hidden cost of buying at the open, Journal of Financial and Quantitative Analysis: retail buying at the open lifts the opening price of attention-grabbing stocks, which give it back by the close; the effect is largest in small, volatile, high-turnover names.
3. Lou, Polk and Skouras (2019), A tug of war: overnight versus intraday expected returns, Journal of Financial Economics: most anomalies earn their return either overnight or intraday, with the two legs often opposed.
4. Knuteson (2020), Strikingly suspicious overnight and intraday returns, working paper: the persistence of the overnight leg across markets and years, and the case that it is made by someone's order flow at the open.
5. Corwin and Schultz (2012), A simple way to estimate bid-ask spreads from daily high and low prices, Journal of Finance: the spread estimator from two consecutive days' ranges used as the dearest of the five costs; the paper reports it overstating for the most liquid names.