A stock-perp long can hand the dividend to the short while the usual funding rule sees nothing wrong. Seo, Cha, Son, Lee, Sung and Lee prove that tuning a standard premium-index rule cannot fix the problem. Their atom result needs no data, and the proof is airtight. The authors make no trading claim. We see a possible trade in an overnight window covered in general terms by their Corollary 1, though we did not find that window measured in the paper. Their event study works at hourly resolution.
Where does the dividend go?
The paper's thesis is that basis-only funding is not dividend fair. For a long entering and exiting at any time, dividend fairness means receiving the cash flows a shareholder would. On a standard venue, the short instead keeps the long's lost dividend: the short gains from the ex-dividend price drop without paying for it.
The authors pair continuous-time fairness results for a financed spot-long, perp-short book with an event study. Its exchange-native records come from Binance, Hyperliquid (13 symbols), OKX and Bitget. The sample spans 12 ex-dates across 7 names between May and June 2026; 5 have both a Binance leg and a standard-venue leg. Binance posts a funding atom of -D/(Mark-D) at each covered ex-date. The standard venues post none, leaving a 6 to 22bp difference for a long held across the event.
Take AAPL. A shareholder receives D in cash when the stock falls by D at the ex-date. A perpetual long takes the mark-to-market loss without that cash; the short gains. Venues calculate funding from the contemporaneous basis e = P - S, the perp mark less the index. At the ex-date, P and S fall together, leaving the basis unchanged. The funding rule misses the dividend.
The authors call a rule weakly fair when the residual of a financed spot-long, perp-short book is a martingale, so an entry and exit pair earns no predictable carry. Weak fairness requires the basis to jump by D in conditional expectation at the ex-date; strong fairness requires that jump pathwise. Ex-dividend marking produces a zero basis jump. The condition fails by the full dividend.
Smooth funding misses the event
A funding spike immediately before the ex-date sounds like a repair. The paper shows why it cannot cover every holding interval. A trader accumulates funding only between entry and exit, and the largest gap across all such intervals between a smooth (absolutely continuous) funding stream and the dividend's point mass is at least D. Even the best smooth stream has a worst gap of exactly D. As an interval around the ex instant shrinks, its smooth funding tends to zero while the dividend remains whole. The authors put it plainly: "Continuity of the funding rule and fairness across a predictable dividend are mutually exclusive." Settlement needs a discrete event credit of size D.
This is the paper's strongest result. It needs no backtest.
Other parts of the theory remain open, as the authors acknowledge. They never estimate the fair target Y_t. Testing whether the basis alone predicts it, and finding the projection floor for the best basis-only rule, remain a program. The sign-reversal result is conditional on an OU model: under mean reversion, fair funding can decline as a positive premium rises if mean reversion dominates the spot-basis covariance term. There are observed inputs for that condition. Hyperliquid's own premium has an hourly AR(1) of 0.72 and a half-life near two hours, while the deployed funding slope is positive on all 13 names.
What do the five ex-dates show?
The dividend calendar is the only external input. The authors use Binance equity-perp funding and mark, Hyperliquid venue-own funding and premium on 13 symbols, OKX records, and the Bitget funding tape. Twelve ex-dates across seven names fall in the window. Table 1 gives five Binance ex-dates with a standard-venue comparison: AAPL, MSFT, NVDA, GOOGL and META, from 11 May to 15 June 2026. META has no Hyperliquid leg in that window, so its standard-venue leg comes from elsewhere.
Every one of those Binance events has a single funding entry equal to -D/(Mark-D), within 5×10^-6. For MSFT, the entry is -0.00218 against a computed -0.00217, both rounded to five decimals in Table 1. Its dividend is 21.8bp as D/P in Table 1 and 21.7bp as D/Mark in the text. Hyperliquid and OKX post nothing. Binance's match is close to circular: -D/(Mark-D) is its published rule, and the records confirm that it follows the rule. The standard venues' silence carries the comparison.
Size alone will not identify an atom. Ordinary equity funding runs 3 to 12bp per interval, alongside atoms from 6.0bp (GOOGL) to 21.8bp (MSFT). The authors match entries by exact equality at the ex timestamp. Bitget poses a further trap for a funding-only screen: it pays a snapshot credit outside its funding history and would therefore appear to have no atom.
The residual heatmap makes the transfer visible. On Hyperliquid, the authors score every MSFT entry and exit pair within ±24 hours of the ex-date. Straddling pairs average +17bp against a D/Mark of 21.7bp; pairs that stay on one side average -0.5bp. Add the atom and the straddling block falls to -4.5bp, with ±3.8bp of speckle left. Its size tracks the dividend: ORCL 36bp, MSFT 22bp and NVDA 12bp. For NVDA's $0.01 dividend, it falls to 0.5bp, a built-in null.
These figures are settlement differences rather than trading P&L. The authors call the 6 to 22bp cross-venue transfer "a settlement-leg size, not an arbitrage net of fees." Long Binance and short Hyperliquid across the ex-date would collect that gap before fees and margin. Carry is separate and depends on the symbol: cumulative long-favoured sub-windows reach +294bp on TSM, +92bp on COST and +70bp on AAPL. The 6 to 22bp range describes the five Binance names. ORCL's 36bp heatmap block belongs to a name without a Binance leg.
From 20:00 ET to the open
NVDA's atom arrives at 2026-06-04 00:00 UTC, or 20:00 ET on 3 June. The cash stock reprices at the ex-date open, the 09:30 ET open, roughly thirteen and a half hours later. Under the paper's own Corollary 1, a single settlement atom is fair for every holding interval only when the price adjustment is a single atom at exactly the same instant. The corollary names overnight gaps and a mark that adjusts at a different time from the oracle among its failure cases.
For Binance, that result implies fairness across the event as a whole and unfairness for intervals that begin or end inside the gap. If Binance's mark waits until the open to move, a long entering just before 20:00 and leaving before the index reprices receives the credit without the drop. On the same assumption, a long entering after 20:00 suffers the drop without the credit. Tradability depends on when Binance's mark and index actually move overnight, and on the sequence of adjustments in Binance's published methodology. We did not find this window measured in the paper. The authors acknowledge that hourly resolution cannot separate the ex-open drop from noise: dividends of a few basis points face roughly 6bp hourly moves.
We could not test the window ourselves. We hold no crypto-listed equity perpetual price or funding series. Stock prices and a dividend calendar cannot stand in for venue funding records and their timestamps, where the entire effect lies.
The atom result is finished. The trade is our construction in the 20:00-to-open interval, rather than a verdict on the authors' work. Testing it calls for minute-level marks and funding prints from each venue across that interval. We did not find minute-level data in the paper, which works at hourly resolution, and we have none either.