AI-generated research, checked but not peer-reviewed.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

E859　 KILL　 UnderpoweredTested · October 2026

No detectable lead from Korean memory-chip moves to US chip perps

A move in Korean memory makers during the Seoul session showed no detectable lead into US semiconductor stock perps before the US open (slope +0.077, t +0.54), and the reverse direction failed out of sample.

cross-market read-through stock-perpetuals korea semiconductors lead-lag

Readable summary Full research record JSON recordFeedback on E859

People argue that when Korean memory makers move during the Seoul session, US semiconductor stocks, trading on thin overnight perpetual books, absorb only part of it before the US open. We tested this over 95 Seoul trading days. The Korean move predicted the US pre-open return with a slope of +0.077 (t +0.54), not distinguishable from a placebo window. Trading its sign after moves above 1% netted +10.2 bp per trade, with an interval from -39 to +59 bp, which is noise. The reverse direction, a US memory-chip move into Korean perps at the Seoul open, flipped sign out of sample.

Seoul trading days95Jun to Oct 2026

Slope on Korean move+0.077t +0.54; required t of at least 2.5

Rule net per trade+10.2 bp82 trades; 95% interval -39 to +59 bp

Reverse test, validation slope-0.137selection half was +0.168

01 · The idea being tested

A move in Korean memory makers during the Seoul session carries into US semiconductor stocks before the US market opens.

Why people believe it

The Seoul session runs from 00:00 to 06:30 UTC, while US semiconductor perpetuals trade against only a thin overnight reference. People argue that traders in those perpetuals do not watch Seoul, so part of the news would reach them late, as US liquidity builds toward the open.

Why it would persist The argument, as stated before testing

Overnight books are thin, and attention to a foreign session is scarce, so the delay would not be worth the cost of competing away.

02 · How it was tested

Each Seoul trading day we measured the average move of two Korean memory makers during their cash session. We then asked whether a US semiconductor stock perpetual, from the Seoul close to just before the US open, moved in the same direction after accounting for what that perpetual had already done during the Seoul session. A placebo repeated the regression with the Korean move from the evening window before Seoul (18:00 to 00:00 UTC, which includes the US afternoon session). The reverse direction was tested separately: we measured the US memory-chip move, net of the Korean perp's own move while Seoul is closed, and asked whether it predicted the Korean perp just after the Seoul open. Its trades pool each Korean stock on each day, and its placebo is the same test with the clock shifted back eight hours.

Pass required

Forward test: slope b on the US memory-chip target above 0 with t at least 2.5, slope on the US AI-chip target above 0, and the rule's net mean above 0 with a day-bootstrap interval above 0. Reverse test: validation slope above 0 with t at least 2.5, at least 15 threshold trades with net mean at least +30 bp and median above 0, selection slope above 0, placebo |t| below 2.

Killed if

Forward test: slope on the US memory-chip target at or below 0, or |t| below 1.5. Reverse test: validation slope at or below 0, or threshold net mean at or below 0.

Costs

Forward test: 2 x 5.5 bp per round trip (11 bp). Reverse test: 2 x 5 bp plus the median quoted spread of the instrument (7.3 and 19.8 bp).

Controls

Forward test: placebo K from the evening window 18:00 to 00:00 UTC, when Seoul is closed, in the same regression, plus the target's own move in the Seoul window (U). Reverse test: the same construction shifted back eight hours, where Seoul is closed at both ends.

03 · What happened

Why it was rejected

Forward test: the slope on the US memory-chip target had t +0.54, below the kill line of 1.5. Reverse test: the validation slope was negative (-0.137), the pre-registered kill condition.

On 95 Seoul trading days, a 1% Korean memory-maker move is followed by +0.077% in the US memory-chip target from the Seoul close to the US pre-open (t +0.54). The US AI-chip target gives +0.032 (t +1.17) and the US semiconductor ETF +0.065 (t +0.92, 54 days).
The placebo, using the Korean move from the evening before Seoul opens, gives -0.05 (t -0.65) for the US memory-chip target. A slope of that size in either direction is what noise looks like here.
The rule (follow the sign of Korean moves of at least 1%, 82 days) earned +10.2 bp net per trade on the US memory-chip target (interval -39 to +59 bp; median +14.7 bp, worst trade -766 bp). Net means after the 11 bp cost; the interval spans zero, so the result is indistinguishable from noise.
On the US AI-chip target the same rule lost 7.3 bp net per trade (interval -24 to +9 bp); on the US semiconductor ETF it lost 14.4 bp (45 trades, interval -53 to +25 bp).
Reverse direction: a US memory-chip move over the Seoul close to 23:50 UTC predicted the Korean perps' return to 01:00 UTC with slope +0.168 (t +1.13, 62 threshold trades, +37.5 bp net, interval -46.0 to +123.9 bp, which includes zero) in the selection half, but -0.137 (t -1.27, 51 trades, -50.2 bp net, median -36.4 bp) in validation. The placebo-clock slopes were +0.003 (t 0.09) and +0.021 (t 0.78).
At 30 minutes after the Seoul open and at the Seoul close, the validation correlation between G and the Korean perp's return was near zero (between -0.18 and +0.03 by name); to 01:00 UTC it was -0.15 for one Korean name and -0.12 for the other.

Trading the Korean move does not clear zero

Net return per trade of following the sign of a Korean memory-maker move of at least 1%, from the Seoul close to the US pre-open, with a 95% day-bootstrap interval, after 11 bp of cost. All three intervals contain zero.

Chart data (bp)

	Mean	95% low	95% high
US memory-chip target (82 trades)	10.2	-39	59
US AI-chip target (82 trades)	-7.3	-24	9
US semiconductor ETF (45 trades)	-14.4	-53.2	25

Same slope in the real window and the placebo window

Seoul session movePlacebo: evening before Seoul

Regression slope of the US pre-open return on the Korean move. The real window (Seoul session) and the placebo (evening before Seoul opens, no Korean session) are of similar size and mixed sign; no real-window slope reaches t 1.2.

Chart data (slope (US return per unit of Korean return))

	Seoul session move	Placebo: evening before Seoul
US memory-chip target (95 days)	0.08	-0.05
US AI-chip target (95 days)	0.03	0
US semiconductor ETF (54 days)	0.07	0.08

Reverse direction: positive, then negative

Net return per trade into Korean perps at the Seoul open after a US memory-chip move, with 95% day-bootstrap intervals. The selection half is positive but includes zero; the validation half is negative; the placebo clock loses roughly the cost. The selection bar reversed to -50.2 bp in validation (a pre-registered split).

Chart data (bp)

	Mean	95% low	95% high
Selection half (62 trades)	37.5	-46	123.9
Validation half (51 trades)	-50.2	-100.8	5.5
Placebo clock, selection (72 trades)	-24.6	-50.9	0.5
Placebo clock, validation (74 trades)	-18.3	-34.3	-2.1

04 · What it means

Why it failed Interpretation · not established

Interpretation, not established: the data fit the idea that US semiconductor perpetuals already price the Korean move while the Seoul session is running, so nothing is left for the hours before the US open, and that the Korean perps already carry the US memory-chip move by 23:50 UTC. The tests do not establish this: with 95 days a slope below about 0.40 could not be separated from zero, so a small real read-through is not ruled out.

Lesson

In short samples, a positive in-sample half (here +37.5 bp) with an interval spanning zero is not evidence of a lead: it reversed to -50.2 bp in the next half.

Scope of this result

Covers the mean of two Korean memory makers' perps, three US semiconductor targets, one window in each direction, and 95 Seoul trading days from June to October 2026. The ETF forward result uses 54 days; the reverse test uses 92 trading days.

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

Stock perpetual 5-minute closes: two Korean memory makers and three US targets (a US memory-chip maker, a US AI-chip maker, a US semiconductor ETF) — largest USD-margined crypto perpetual venue, public market-data archive; 5 minutes; 2026-06..2026-10; public | Stock perpetual 5-minute closes for the reverse test: the same two Korean names and the US memory-chip maker — a large crypto derivatives venue, public candle history; 5 minutes; 2026-06..2026-10; public; a public holiday calendar drops Korean holidays

Universe

Forward test: Korean memory-maker perps as the signal; US memory-chip maker (primary), US AI-chip maker (secondary) and a US semiconductor ETF (reported) as targets. 95 Seoul trading days; the ETF perp has 54 usable days. Seoul holidays drop out through a volume filter (Korean quote volume in the Seoul window at least 25% of its trailing 20-day median).

Signal

K = mean return of the two Korean perps from 00:00 to 06:30 UTC. U = the target's return over the same window. Y = the target's return from 06:30 UTC to 09:25 New York time (about seven hours). Primary regression: Y = a + b K + c U, Newey-West standard errors with 5 lags. Trading rule: take the sign of K when |K| is at least 1%. Reverse test: G = beta x (US memory-chip return) - (Korean perp return) over the previous 06:30 UTC to 23:50 UTC, beta fixed in advance at 1.0 and 0.5 for the two Korean names; outcome y = Korean perp return from 23:50 to 01:00 UTC.

Execution

Forward test: enter at the Seoul close, exit at 09:25 New York time. Reverse test: trade the sign of G when |G| is at least 100 bp, from 23:50 to 01:00 UTC.

Costs

Forward test: 2 x 5.5 bp per round trip (11 bp). Reverse test: 2 x 5 bp plus the median quoted spread of the instrument (7.3 and 19.8 bp).

Controls

Sample split

Forward test: the full period, no split. Reverse test: selection before mid-August 2026, validation from mid-August to October.

Pass criterion

Kill criterion

Forward test: slope on the US memory-chip target at or below 0, or |t| below 1.5. Reverse test: validation slope at or below 0, or threshold net mean at or below 0.

Power

Forward test: the Newey-West standard error of the slope is about 0.141 (slope divided by t), so only slopes above about 0.40 would have been detectable with 95 days; the rule's 82 trades have a net standard error of about 24.7 bp (223 bp standard deviation divided by the square root of 82), a minimum detectable effect near 69 bp.

2. Result

Observations

Instruments

Net mean

+10.2 bp

95% interval

-39 bp to +59 bp

Median

+14.7 bp

Worst

-766 bp

Finding 1

On 95 Seoul trading days, a 1% Korean memory-maker move is followed by +0.077% in the US memory-chip target from the Seoul close to the US pre-open (t +0.54). The US AI-chip target gives +0.032 (t +1.17) and the US semiconductor ETF +0.065 (t +0.92, 54 days).

Finding 2

The placebo, using the Korean move from the evening before Seoul opens, gives -0.05 (t -0.65) for the US memory-chip target. A slope of that size in either direction is what noise looks like here.

Finding 3

The rule (follow the sign of Korean moves of at least 1%, 82 days) earned +10.2 bp net per trade on the US memory-chip target (interval -39 to +59 bp; median +14.7 bp, worst trade -766 bp). Net means after the 11 bp cost; the interval spans zero, so the result is indistinguishable from noise.

Finding 4

On the US AI-chip target the same rule lost 7.3 bp net per trade (interval -24 to +9 bp); on the US semiconductor ETF it lost 14.4 bp (45 trades, interval -53 to +25 bp).

Finding 5

Reverse direction: a US memory-chip move over the Seoul close to 23:50 UTC predicted the Korean perps' return to 01:00 UTC with slope +0.168 (t +1.13, 62 threshold trades, +37.5 bp net, interval -46.0 to +123.9 bp, which includes zero) in the selection half, but -0.137 (t -1.27, 51 trades, -50.2 bp net, median -36.4 bp) in validation. The placebo-clock slopes were +0.003 (t 0.09) and +0.021 (t 0.78).

Finding 6

At 30 minutes after the Seoul open and at the Seoul close, the validation correlation between G and the Korean perp's return was near zero (between -0.18 and +0.03 by name); to 01:00 UTC it was -0.15 for one Korean name and -0.12 for the other.

3. Interpretation

Why it failed

Lesson

In short samples, a positive in-sample half (here +37.5 bp) with an interval spanning zero is not evidence of a lead: it reversed to -50.2 bp in the next half.

Scope

4. Reproducibility

Steps

1) Download 5-minute closes of the Korean and US stock perpetuals; price at time T is the close of the bar that ends at T. 2) For each weekday with Korean quote volume in 00:00 to 06:30 UTC at least 25% of its trailing 20-day median, compute K, U and Y as defined. 3) Regress Y on K and U with Newey-West errors (5 lags); repeat with K taken from 18:00 to 00:00 UTC. 4) Trade the sign of K when |K| is at least 1% and subtract 11 bp; bootstrap days for the interval. 5) For the reverse test, compute G and y on weekdays that are not Korean public holidays, split at mid-August 2026, take the slope with a day-clustered bootstrap t, and repeat shifted back eight hours.

Grade

COMPLETE