Dividend capture loses, the week before the ex-date pays, and holding the same stocks beats both

A dividend has one date that matters, the ex-date. Whoever owns the stock at the close before it receives the payment; whoever buys on the ex-date or after does not. The payment was fixed weeks earlier, at the declaration, so nothing is learned on the day. What happens on the day is that the price opens lower, by the amount paid if the market is doing its arithmetic, by less if buyers are keen, by more if they are not. The drop ratio in this paper is that fall divided by the payment: 1.00 means the price gave the whole payment back, 0.50 means half of it stayed with the holder.

Everything else is a choice of where to stand around that date, and the four trades here are the four choices. They use the same names and the same dates and differ in one thing only, what they do with the payment.

A, collect overnight. Buy at the close the day before the ex-date, sell at the close of the ex-date. One night. You receive the payment and you take the drop. If the drop equals the payment you have made nothing and paid the spread twice. This is the dividend game as people play it, and the drop ratio decides it: the trade pays only if the price falls by less than the payment.

B, the week after. Buy at the close of the ex-date, after the drop, and sell five trading days later. No payment. It pays only if the price climbs back after the drop, which is what anyone who calls the drop an overreaction is claiming.

C, collect and hold. Buy the close before the ex-date, hold through the date, sell five trading days later. This is A followed by B with one trip of the spread fewer, and it equals A then B by construction: the dividend investor who holds a week. It pays if either of the two above pays.

D, the week before. Buy six trading days before the ex-date and sell at the close the day before it, the last close at which the stock still carries the payment. Five days. It never receives a dividend and never takes the drop; it holds the stock only while the people who want the payment are buying it. It pays only if the price rises into the date. Figure 1 below draws the price path around the date and shades what each trade holds; Table 1 lists the four side by side.

Two ways of counting sit on top of these. Every event once: each trade is run on every dividend event of a list's names inside a window, one position per event, so the result is the trade's own record. The rotation: a fixed pot of ten equal slots, each slot taking the next upcoming ex-date of the list's names with the chosen trade and sitting in Treasury bills in between, which is what a person with one account would actually run. Holding the same names, buying every name of the list equal weight at the start of a window and holding through it with the dividends credited, is the passive version of the same names and the reference every trade is set against; the dividend funds are its buyable version. The registered records call each list a cut, as in a cut of the payers; it has nothing to do with a company cutting its payment.

Universe. Every company that was a member of the S&P 500 at some point from 2004, with membership at each January first rebuilt from the index's change-log, so a year's payers are the names that were in the index that January. Names that left the index and the vendor no longer prices are counted in the limitations; nothing is filled in for them.

Events. A cash dividend is one event; its ex-date is day zero. The amount is the vendor's split-adjusted payment, the closes are split-adjusted and not dividend-adjusted, the price a holder saw fall. Events over three times the company's usual payment (specials), windows with a bar-to-bar move over 25 percent (a split the vendor did not carry through), and payments over ten percent of the price are left out and counted.

The two lists, decided each January from the year before: high yield, the top fifth of the year's payers by prior-year dividends over the first close of the year; long record, the top fifth by the run of consecutive years in which the annual payment rose, ties by yield. The lists overlap in 14 names a year out of about 70.

The census. For every event: the drop ratio, the fall from the close before to the ex-date close over the payment; the same ratio with the market's move that day (SPY) credited to the ex-date; the calendar days from the last earnings announcement known before the ex-date, after-close announcements counting from the next morning, grouped 0 to 7, 8 to 14, 15 to 21, 22 to 35, 36 to 56, 57 and more days, and the days from the ex-date to the next announcement; the mean price path from six days before to five after against the close before; and the run-up over trade D's own span, the close six sessions before to the close before, plain and with SPY's move over the same bars taken out.

Checks and definitions. A then B with A's proceeds reinvested at the ex-date close equals C by construction, and the engine's identity check is part of the record; the data check that matters is bar continuity, printed as the count of rolled ex-dates (7) and of windows dropped for a broken bar. The per-event return in the tables is the simple return of one position from entry close to exit close with any dividend received, before the spread; the average week it is set against is the same names' holding return over the window taken to five trading days, also before costs. The platform's paired-window tables at the end pair the two arms of the paper's own walk bar by bar and start one bar after the window's first close, so their window means differ from Table 4 in the second decimal.

The walks. For each list and each trade, one registered study with eighteen one-year windows anchored each January from 2008 to 2025. Two arms in every window: every name of the list, and the third of the list whose dividend-weighted drop ratio over the three calendar years before the anchor was lowest, computed from events strictly before the anchor, at least four events with a payment of at least 0.3 percent of price. Every event of the arm's names inside the window is traded; each open position runs at a constant tenth of book equity (equal split past ten), dividends are credited from the payment record, one position per name at a time, and every position pays its liquidity tier one way at entry and at exit: 1.5, 3 or 6 basis points by 63-day dollar volume before the ex-date. Returns are reported gross and net. A window with fewer than fifteen names with an event would be excluded; none was. The paper's own walk is high yield, trade D, the ranked third against every name, the record that carries the paper's claim about the ranking; trade C on the same list was registered first, as the trade people run, and stays a member of the family.

The rotation's mechanics. When more events arrive than free slots, the names ranked best in the training window go first; an idle slot earns the three-month bill; the two references are holding every name of the list equal weight through the window with dividends credited, and holding bills. The rotation was registered on trade C first, the trade people actually run, and again on trade D, the trade that held up in the walks (net positive in a majority of the eighteen windows for both lists).

The index layer, continuous from 2005 to 2026 and before costs: the two lists as equal-weight indices rebalanced on the first trading day of each quarter, daily equal weight over the names held. The shifted calendar holds a leaving name to the bar after its ex-date when that ex-date falls inside the week after the sale, and takes an entering name in at the bar before its ex-date when that falls inside the week before the purchase; one week, set once. At each pick the yield shown (the trailing year's payments over the pick price) is set against the yield collected in the quarter after, name by name, with the next payment classed against the last one before the pick. This layer has no registration record: its trades sit at the boundaries of any one-year window, so a registered window cannot hold them.

The fund layer, a measurement of the same kind: five dividend funds, each as total return with its distributions reinvested at the ex-date close from the payment record, set against the index of the list it resembles from the fund's first January (CAGR of each, the gap, the correlation of daily returns), and on the walks' own windows from that January against the rotation on trade D, holding the same names and bills.

The drop. All payers gave back a median 0.81 of the payment on the ex-date; with the market's move that day credited, 0.95. The high-yield list gave back 0.97 plain and 1.04 adjusted, the long-record list 0.87 and 0.99. By year the plain median runs from 0.60 to 1.24 with no trend; the adjusted one stays between 0.73 and 1.19. Across the 503 companies with eight or more events, the middle half of company-level ratios sits between 0.32 and 1.16; the extremes belong to payments of a few basis points of price, and the size-weighted ratio behind the ranking is the one to read. By liquidity tier the fall matches the payment for the most liquid names on the plain measure and exceeds it market-adjusted, a median 0.96 and 1.09 for names trading over 500 million dollars a day, and falls short for the thinnest, 0.75 and 0.86 under 100 million; 2 census events have no volume record and no tier.

The earnings calendar. The median event comes 35 days after the last announcement, half of them between 18 and 54 days, and 55 percent of companies keep a fixed calendar, their gaps within three days of their own median. By distance from the last announcement, the week after the ex-date fades from 0.30 percent inside a week of it to 0.10 at five to eight weeks, the shape of a post-earnings drift, and comes back to 0.26 past eight weeks. The first five groups fade the way an earnings pattern would. The last does not, and the next announcement is not the reason: it lands inside the week after for 18.2 percent of that group's events, and without them the group's week after is 0.27, no lower.

The four ways to stand around the date, net of the spread, mean book return over the eighteen windows, every name of the list. Collecting overnight, trade A: -1.88 percent a year for the high-yield names (6 of 18 windows positive) and -1.42 for the long-record names; gross it is 0.15 and 0.43, a wash, so the payment bought nothing and the spread took the rest. The week after without the payment, trade B: 1.24 and 0.61, under an average week per event (0.11 and 0.12 against 0.18 on either list), so the drop is not made back. Collecting and holding a week, trade C: 1.52 and 1.04, A plus B. The week before, leaving the day before the payment, trade D: 6.37 and 5.75, positive in 11 and 14 of 18 windows. Collecting the payment added nothing on either list, and the week before earned 2.9 times per event what the week after earned on the high-yield list and 2.3 times on the long-record list, 0.32 against 0.11 and 0.29 against 0.12, before the spread. The rule used to pick the trade for the second rotation: net positive in at least ten of the eighteen windows with a positive mean, for every name of the list. The trades that pass: D in high yield, C in long record, D in long record.

Ranking by past drop ratio. The third of each list with the lowest past drop ratio kept more of the payment in the capture trade: 97 percent probability of beating the whole list for high yield and 90 for long record, per event 0.036 against -0.004 percent. It did not pay the spread: -0.22 and -0.49 percent a year net. And the same ranking loses the run-up: every name beats the ranked third on trade D with 97.8 and 98.3 percent probability, a paired Sharpe of -0.49 on the paper's own walk. The names whose price gives back least on the ex-date are the names that run up least before it.

The week before, checked. The run-up week beat an average week of the same names in 10 of 18 windows on the high-yield list and 10 on the long-record list, 0.32 against 0.18 percent per event and 0.29 against 0.18, before the spread. On the census events the same five days returned 0.33 percent for the high-yield list, 0.10 more than SPY over the same bars, positive against the market in 51 percent of events; long record 0.29 and 0.05. Split after 2013, the year Hartzmark and Solomon published (Table 6): the high-yield premium shrank, 0.45 against 0.18 for an average week through 2013 and 0.26 against 0.18 from 2014, and the long-record one showed up, 0.13 against 0.15 and then 0.37 against 0.20. Event by event, with the ex-date week as the unit so that events sharing a week count once (10,000 draws): against SPY over the same five bars, on the whole census, the high-yield run-up is 0.10 percent with a 95 percent interval of -0.02 to 0.23, the long-record 0.05, -0.04 to 0.15. Against the year's average week, on the census events of the walk years 2008 to 2025 with every event weighted once (Table 6's figures are window means, hence 0.11 there and 0.13 here for long record), 0.14 for high yield, -0.05 to 0.33, and 0.13 for long record, -0.03 to 0.30; 1.4 and 1.6 standard errors from zero. Split at three weeks from the last earnings announcement (Table 6, last two columns), the high-yield premium is not post-earnings drift: inside 21 days the run-up is 0.03 against an average week and 0.00 against SPY, nothing, and from 22 days on it is 0.19 against an average week and 0.15 against SPY, 0.01 to 0.29, clear of zero, and 0.16 against SPY with the next announcement kept more than ten days away. On the long-record list the split changes nothing, 0.14 against an average week inside 21 days and 0.13 beyond, neither clear of zero. The premium is the dividend one; on the long-record list it is positive in every cut and settled in none.

The rotation. On trade C the high-yield rotation earns 2.25 percent a year net (gross 3.98), in trades 53 percent of the time, 227 trades a year, 1.67 percent a year paid in spread, against 10.88 for holding the same names and 1.37 for bills; long record 2.21 net against 10.45 holding. On trade D the rotation earns 6.17 and 6.75 percent net, still under holding at 10.88 and 10.45, because it is in trades less than half the time, and 1.8 and 1.6 times an average week over less than half the year is less than a full year of average weeks.

The funds. The long-record index built here and the Dividend Aristocrats fund (NOBL) move together at 0.98 daily correlation from 2014, the fund 1.03 percent a year behind the index before costs; the high-yield index and the S&P 500 High Dividend fund (SPYD) at 0.97, the fund 4.20 behind. On the walks' own windows from each fund's first full year, the long-record rotation on trade D earned 9.27 percent a year net against 10.65 for holding NOBL and 10.84 for holding the same names, with the rotation's money in bills 55 percent of the time; it beat the fund in 5 of 12 windows. Against SPYD the high-yield rotation earned 7.75 against 9.88. Chained from 2016, the first January both funds exist (Figure 3), the long-record rotation grew 2.85x against 2.65x for NOBL; from 2014, NOBL's first January, the fund is ahead, 3.05x against 2.76x.

The 2020 window. The long-record list had 324 ex-dates in 2020 against 308 a year on average over 2016 to 2019, so suspensions did not empty the rotation. Its 34.7 percent that year came in the rebound months: March, April, June, August and November together made 31.4 percent, with the book in bills between its trades while the fund held through the fall and the recovery, 1.35x against 1.08x over the year.

The index layer. The high-yield index shows 5.27 percent yield at its picks and collects 4.75 in the quarter after, per name and pick, and 3.7 percent of the picks paid nothing in that quarter (Table 7). Of that half point, 0.28 is the 2.4 percent that paid nothing for two quarters (suspensions), 0.07 the 1.3 percent whose ex-date slipped past the quarter end, and 0.16 the 5.2 percent that cut the payment; the long-record index shows 2.64 and collects 2.70. Holding the high-yield names on the shifted calendar instead of the fixed one is worth 4.7 percent of final value over the span, 4.81 against 4.62 percent a year in dividends collected; the long-record index gains 0.06 percent, nothing. The difference is turnover: 98 names entered the high-yield index inside the week before an ex-date over the span against 14 for the long-record index.

Of the four trades, the one that only collects the payment, A, earns less than the spread, and collecting and holding a week, C, is positive by the week after alone; the payment part earns nothing. On the ex-date the price gives the payment back: for the names people actually pick, the high-yield list, 0.97 of it on the plain measure and 1.04 with the market's move credited, so capture's gross return is the noise of one day and its net return is the spread. For all payers a fifth of the payment stays with the holder on the plain measure, and most of that is the market's own drift on those days.

What persists is smaller than the trade needs. Companies whose price gave back less in the past give back less again, with 97 percent probability of the ranked third beating the whole high-yield list, and the effect is worth 0.039 percent per event, under a round trip of the spread for all but the most liquid names. The week before the ex-date returns 0.32 percent per event on the high-yield list against 0.18 for an average week, and it belongs to the names that give back most on the day.

This is the dividend month premium of Hartzmark and Solomon (2013): abnormal returns in the months a dividend is predicted, tied to buying by investors who want the payment, larger before the ex-day and reversed after it. Harris, Hartzmark and Solomon (2015) name the buyers, funds that hold names through their ex-dates to raise the yield they report. After their paper appeared the premium shrank on the high-yield list and showed up on the long-record list; it beat an average week in 10 of 18 windows on each, and against the market over the same five days it is 0.10 percent per event for the high-yield list, positive in 51 percent of events. Split at three weeks from the last earnings announcement, the high-yield premium is absent inside and 0.15 percent against the market beyond, clear of zero, so it is not post-earnings drift.

A book that holds only those weeks earned 6.75 percent a year against 10.45 for holding the same stocks over the eighteen years, because it is in trades less than half the time. It moves with the Dividend Aristocrats fund (Figure 3) and kept up with it only from 2016, 2.85x against 2.65x, with 2020 supplying the margin, 1.35x against 1.08x in that year, made in the rebound months with the book in bills between them. From 2014 the fund is ahead, 3.05x against 2.76x.

For an index builder the shifted calendar is a small gain on the high-yield side and nothing on the long-record side. The high-yield index gains because its names change often, 98 entries inside the week before an ex-date over the span against 14 for the long-record index, which barely changes its names. The number that matters more to a dividend investor is the gap between the yield shown at the picks and the yield collected in the quarter after, 5.27 against 4.75 percent, and over half of that half point is names that paid nothing for two quarters after they were picked.

What would change the conclusion: quoted spreads instead of tiers, which would lower every net number, not raise it; a longer hold after the ex-date, which becomes a different question; or a list of names with payments large relative to price, where the census's tails live and where the ratio is far from one for reasons of liquidity rather than of the dividend.

The vendor prices none of the 117 names that left the index between 2004 and 2026, and 116 of them have no dividend record with it either, so none entered a list; the hole is 560 member-years out of about eleven thousand. If payers that went bust are among them, holding, which keeps a name through its fall, loses more than a five-day trade does, so the gap between the trades and holding is narrower than shown, and the fund comparison is made against an index that could not hold those names. The spread is charged by a three-tier proxy from dollar volume, not from quoted spreads, so the net numbers are a floor on the cost for thin names. Each yearly window drops the events whose entry falls before its first bar or whose exit falls after its last, counted per window. Taxes are not modelled; a holding of a day or a week does not qualify for the reduced US dividend rate, so the after-tax gap to holding is wider than shown. The index layer is a measurement without a registration record and before costs. The fund layer runs from each fund's first January, 12 of the 18 windows for the Aristocrats fund and 10 for the high-dividend one, and sets a fund net of its own fee against an index before costs.

1. Elton and Gruber (1970), Marginal stockholder tax rates and the clientele effect, Review of Economics and Statistics: the ex-date drop as a fraction of the dividend, measured at 0.78 for NYSE stocks 1966 to 1967.
2. Kalay (1982), The ex-dividend day behavior of stock prices, Journal of Finance: the drop ratio read as a bound set by transaction costs rather than as a tax clientele.
3. Boyd and Jagannathan (1994), Ex-dividend price behavior of common stocks, Review of Financial Studies: dividend capture by taxable investors and the role of costs.
4. Hartzmark and Solomon (2013), The dividend month premium, Journal of Financial Economics: abnormal returns in the months a dividend is predicted, larger before the ex-day and reversed after it, read as price pressure from investors who want dividends.
5. Harris, Hartzmark and Solomon (2015), Juicing the dividend yield: mutual funds and the demand for dividends, Journal of Financial Economics: funds that buy names ahead of their ex-dates to raise the yield they report.