AI-generated research, checked but not peer-reviewed. It may contain errors.

E190　 KILL　 Below costsTested · October 2026

Fading a Korean coin premium on global perpetuals loses

Shorting a global perpetual after a coin-specific Korean premium of 5% or more lost 130 bp on average over 4 hours net of costs and funding (416 events; interval -235 to -24 bp), and 158 bp over 24 hours.

premium korea funding perpetuals event-study reversion

Readable summary Full research record JSON recordFeedback on E190

Some coins trade at a sudden premium on large Korean won exchanges. People say it is temporary, so shorting the same coin on a global perpetual futures venue should pay when it closes. We found 416 such spikes of 5% or more between 2023 and 2026 and shorted the perpetual at the next hourly open. After costs and funding the average trade lost 130 bp over 4 hours (interval -235 to -24 bp) and 158 bp over 24 hours. The short also paid funding: 61 bp over 4 hours and 198 bp over 24 hours. None of the four pre-registered cells passed.

Fade, 4 hours, net-130 bpn 416 events; 95% interval -235 to -24 bp; median -61 bp

Fade, 24 hours, net-158 bp95% interval -352 to +41 bp; median -25 bp

Funding paid by the short198 bpper 24 hours, 95% interval 152 to 255 bp

Primary cells passing0 of 4none cleared the pre-registered criteria

01 · The idea being tested

When a coin trades at a sudden premium on a Korean exchange, the premium fades, so short it on a global perpetual and collect the fall.

Why people believe it

People argue that Korean retail buying pushes one coin above its global price, that capital controls and exchange limits slow the arbitrage that would close the gap, and that the premium therefore decays within hours. Shorting the global perpetual of the same coin is the cheapest way to bet on that decay without owning or borrowing the coin.

Why it would persist The argument, as stated before testing

Common belief, not tested here: moving coins and won across the border is slow and restricted, so the gap may stay open long enough for a short to be paid.

02 · How it was tested

For each coin listed on a large Korean won exchange and also on a global perpetual futures venue, we compared its won price with its perpetual price every hour. To remove exchange rates and the market-wide Korean premium, we subtracted the median of the same gap across all coins at that hour, leaving a coin-specific premium. When it first reached 5% after at least 24 hours below, we shorted the perpetual at the next hourly open and held 4 or 24 hours. We charged trading costs and the real funding payments, and compared each event with 20 random hours of the same coin.

Pass required

Per primary cell: day-clustered interval above zero net of 7 bp per side, corrected q below 0.05, mean above zero at 20 bp per side, median above zero, validation and holdout means above zero, event-minus-control interval above zero, and no dependence on events within 72 hours of a new listing.

Killed if

No separate kill line was recorded; a primary cell that fails any pass criterion is rejected. The four primary cells are premium short and discount long, each at 4 and 24 hours.

Costs

7 bp per side (stress case 20 bp per side) plus the actual funding settlements between entry and exit.

Controls

20 random valid hours per event for the same coin, same direction and hold, with no event of that coin within 72 hours; the event-minus-control difference is reported.

03 · What happened

Why it was rejected

All four pre-registered cells failed: the 4-hour premium short lost 130 bp with a 95% interval entirely below zero, and the 24-hour cell lost 158 bp with an interval of -352 to +41 bp.

All intervals are day-clustered bootstrap 95% intervals. The premium short averaged -130 bp over 4 hours (n 416 events on 242 days; interval -235 to -24 bp; median -61 bp). Minus the matched controls, which averaged -3 bp, the difference was -127 bp (interval -230 to -27 bp). A negative number means the short lost money.
Over 24 hours the short averaged -158 bp (interval -352 to +41 bp; median -25 bp; event minus control -166 bp). The worst single trade lost 2,983 bp at 4 hours and 8,201 bp at 24 hours.
By time window the 4-hour short lost 160 bp, 129 bp and 43 bp in selection, validation and holdout; the 24-hour short lost 206 bp, 112 bp and 61 bp. No window was positive.
A 10% threshold (143 events) gave the same picture: -169 bp at 4 hours and -192 bp at 24 hours. Entering one hour later gave -91 bp at 4 hours (interval -226 to +24 bp) and -60 bp at 24 hours (interval -220 to +119 bp).
The perpetual short paid funding of 61 bp over 4 hours (interval 48 to 79 bp) and 198 bp over 24 hours (interval 152 to 255 bp; median 61 bp). Over 4 hours the perpetual itself rose about 54 bp on average, so the short lost on price as well. Funding paid is already inside the net figures above.
The discount events (cs of -5% or lower) number 73 on 14 days, and 58 of them fall on two days: a global crash in which perpetual prices fell while the Korean exchange did not, and a night of political turmoil in Korea. The 4-hour long averaged -60 bp (interval -318 to +325 bp) and the 24-hour long +600 bp (interval -1,114 to +2,770 bp, corrected q 0.5), noise from two days, not evidence.
Corrected q values for the four primary cells were 0.06, 0.25, 0.7 and 0.5; none was below 0.05. The 4-hour premium short missed the pass test because its result is negative, not merely insignificant.
Only 3% of premium events stayed above the threshold for the next 24 hours, and only 4 of 416 events were within 72 hours of a new listing on the Korean exchange. Results were unchanged when requiring won turnover of at least 0.2 or 1 billion won per bar.

The fade loses, and the controls explain none of it

Mean net return of the premium short with day-clustered 95% intervals, and the same minus matched random hours. At 4 hours both intervals lie entirely below zero; at 24 hours they straddle zero but the means are -158 and -166 bp.

Chart data (bp per trade, net of costs and funding)

	Mean	95% low	95% high
Short, 4 h	-130	-235	-24
Short, 4 h, minus controls	-127	-230	-27
Short, 24 h	-158	-352	41
Short, 24 h, minus controls	-166	-341	43

The short pays funding while the position is open

Average funding paid by the premium short over the 4-hour and 24-hour holds (416 events). Funding of 61 bp at 4 hours and 198 bp at 24 hours is included in the net losses of -130 and -158 bp.

Chart data (bp paid by the short per trade)

	Funding paid
4 h hold	61
24 h hold	198

No time window is positive

SelectionValidationHoldout

Mean net return of the 5% premium short by hold and time window (60/20/20 split). All six bars are negative; the 4-hour holdout is the smallest loss at -43 bp.

Chart data (bp per trade, net)

	Selection	Validation	Holdout
4 h hold	-160	-129	-43
24 h hold	-206	-112	-61

04 · What it means

Why it failed Interpretation · not established

Likely explanation, not established: traders who carry coins into Korea may hedge with perpetual shorts. When the funding rate is negative, shorts pay longs, which would make a short expensive to hold exactly when the premium is high; the short paid 198 bp over 24 hours. A spot short, which would not pay funding, would need borrowable coins that these altcoins generally lack. Premium events also cluster in altcoin rallies, which adds market beta to the first hours of the short.

Lesson

A position on a perpetual pays or receives funding that can be as large as the move being traded, so event tests on perpetuals need funding inside the return and should report it separately from price and costs.

Scope of this result

Covers 242 coins that were listed on a large Korean won exchange at the time of data collection and have a perpetual on the largest USD-margined venue, hourly data from 2023 to 2026, premium and discount events defined against the cross-sectional median, shorts and longs held 4 or 24 hours, at 7 and 20 bp per side plus actual funding. Coins delisted before data collection are missing, so the sample is survivors only. The premium to the dollar price was computed only from mid-2024, when a dollar-stablecoin won market became available; the coin-specific premium needs no exchange rate.

A rejected experiment applies to the tested conditions and does not prove an idea can never work. bp = basis point, 0.01%.

1. Test design

Pre-registered

Yes: rule and criteria fixed before the test data was examined

Data

Hourly won prices of 243 coins against the Korean won (242 matched to a perpetual contract; 2 dropped as different tokens) — the public quotation API of a large Korean won exchange; 1 hour; 2023-01..2026-10; public API, no login | Hourly perpetual prices and funding payments for the matched coins — the public klines and funding feeds of the largest USD-margined crypto perpetual venue; 1 hour; funding settlements; 2023-01..2026-10; public download, no login

Universe

Won markets listed at the time of data collection whose coin has a USD-margined perpetual (242 coins). Delisted markets return no candles, so the sample contains only coins that survived. A coin is dropped if its median absolute coin-specific premium exceeds 3% over all valid hours (this removed 2 coins that are different tokens).

Signal

cs = ln(won price) - ln(perpetual price, adjusted for contract multiplier), minus the cross-sectional median of the same quantity over valid coins (at least 20 coins). A bar is valid if it exists without forward fill, won turnover is at least 50 million won, the previous hour also exists and the perpetual bar has volume. Event: the first bar close with cs of at least +5% (primary) or +10% (secondary), after at least 24 hours below the threshold with at least 12 valid bars in that window. Mirror event: cs of at most -5%.

Execution

Premium events: short the perpetual at the open of the next hourly bar. Discount events: long. Hold 4 hours or 24 hours. Simple returns, floored at -100%. Variants: 10% threshold; entry one hour later.

Costs

7 bp per side (stress case 20 bp per side) plus the actual funding settlements between entry and exit.

Controls

20 random valid hours per event for the same coin, same direction and hold, with no event of that coin within 72 hours; the event-minus-control difference is reported.

Sample split

60% / 20% / 20% by time: selection, validation, holdout. Day-clustered bootstrap 95% intervals (2,000 draws); Benjamini-Hochberg correction at q 0.05 across the 4 primary cells.

Pass criterion

Kill criterion

No separate kill line was recorded; a primary cell that fails any pass criterion is rejected. The four primary cells are premium short and discount long, each at 4 and 24 hours.

Power

416 premium events on 242 distinct days; the 4-hour interval spans -235 to -24 bp and the 24-hour interval -352 to +41 bp, so differences smaller than roughly a hundred bp could not be told from zero. The discount cells have 73 events on 14 days and are not a usable sample.

2. Result

Observations

416

Instruments

159

Net mean

-130 bp

95% interval

-235 bp to -24 bp

Median

-61 bp

Worst

-2,983 bp

Finding 1

All intervals are day-clustered bootstrap 95% intervals. The premium short averaged -130 bp over 4 hours (n 416 events on 242 days; interval -235 to -24 bp; median -61 bp). Minus the matched controls, which averaged -3 bp, the difference was -127 bp (interval -230 to -27 bp). A negative number means the short lost money.

Finding 2

Over 24 hours the short averaged -158 bp (interval -352 to +41 bp; median -25 bp; event minus control -166 bp). The worst single trade lost 2,983 bp at 4 hours and 8,201 bp at 24 hours.

Finding 3

By time window the 4-hour short lost 160 bp, 129 bp and 43 bp in selection, validation and holdout; the 24-hour short lost 206 bp, 112 bp and 61 bp. No window was positive.

Finding 4

A 10% threshold (143 events) gave the same picture: -169 bp at 4 hours and -192 bp at 24 hours. Entering one hour later gave -91 bp at 4 hours (interval -226 to +24 bp) and -60 bp at 24 hours (interval -220 to +119 bp).

Finding 5

The perpetual short paid funding of 61 bp over 4 hours (interval 48 to 79 bp) and 198 bp over 24 hours (interval 152 to 255 bp; median 61 bp). Over 4 hours the perpetual itself rose about 54 bp on average, so the short lost on price as well. Funding paid is already inside the net figures above.

Finding 6

The discount events (cs of -5% or lower) number 73 on 14 days, and 58 of them fall on two days: a global crash in which perpetual prices fell while the Korean exchange did not, and a night of political turmoil in Korea. The 4-hour long averaged -60 bp (interval -318 to +325 bp) and the 24-hour long +600 bp (interval -1,114 to +2,770 bp, corrected q 0.5), noise from two days, not evidence.

Finding 7

Corrected q values for the four primary cells were 0.06, 0.25, 0.7 and 0.5; none was below 0.05. The 4-hour premium short missed the pass test because its result is negative, not merely insignificant.

Finding 8

Only 3% of premium events stayed above the threshold for the next 24 hours, and only 4 of 416 events were within 72 hours of a new listing on the Korean exchange. Results were unchanged when requiring won turnover of at least 0.2 or 1 billion won per bar.

3. Interpretation

Why it failed

Lesson

Scope

4. Reproducibility

Steps

1) Download hourly won candles for all won markets of a large Korean exchange and the matching perpetual klines and funding history, 2023-01 to 2026-10, with multiplier-adjusted symbols. 2) cs = ln(won price) - ln(perpetual price) minus the cross-sectional median over valid coins at each hour; keep bars with at least 50 million won of turnover and drop coins with median |cs| above 3%. 3) Event: first close with cs of at least +5% after 24 hours below, with at least 12 valid bars in that window. 4) Short the perpetual at the next hourly open, exit after 4 or 24 hours; subtract 7 bp per side and the actual funding settlements in (entry, exit]. 5) Controls: 20 random valid hours of the same coin, no event within 72 hours. 6) Day-clustered bootstrap intervals, 60/20/20 time split, Benjamini-Hochberg at q 0.05 across the 4 primary cells.

Grade

COMPLETE