Data as of: Korean names are 2026-07-31 KRX closing prices; U.S. names are the 2026-07-30 official close plus 2026-07-31 pre-market quotes (NOT closing prices — see "Limits of this method" below). Every ratio in this piece can be recomputed from the raw percentage moves given in the text; nothing depends on undisclosed data.
Bottom line
Two headline numbers printed on the same day, and neither can be read at face value: one is contaminated by market beta, the other by calendar misalignment. After two pieces of arithmetic, "Korean memory names exploded +30%" shrinks to a real thematic excess of +16.84%, and the "extreme U.S.–Korea divergence of −11.52pp, 99th percentile" signal goes to zero — the aligned spread is just 4.67pp, entirely normal. Both conclusions reverse.
Contamination layer 1: how much of the headline was just the Korean market?
On 2026-07-31 three Korean memory-complex names all pushed toward the +30% single-day limit. But non-memory Korean large caps rallied hard the same day, and that portion has nothing to do with the memory narrative.
Memory complex (2026-07-31 KRX close)
| Name | Close | 1-day move |
|---|---|---|
| SK Hynix (000660.KS) | KRW 1,718,000 | +29.95% |
| Samsung Electronics (005930.KS) | KRW 262,500 | +26.81% |
| Hanmi Semiconductor (HBM bonding equipment) | KRW 214,000 | +27.98% |
| Median | — | +27.98% |
Non-memory Korean large-cap control basket (same day)
| Name | 1-day move |
|---|---|
| Samsung SDI | +10.74% |
| Hyundai Motor | +10.54% |
| LG Electronics | +9.53% |
| Kia | +9.19% |
| LG Energy Solution | +2.50% |
| Median | +9.53% |
The decomposition
Using the non-memory median of +9.53% as that day's market baseline:
| Name | Headline move | Share of headline that is market beta | Relative excess (multiplicative) |
|---|---|---|---|
| SK Hynix | +29.95% | 31.8% | +18.64% |
| Hanmi Semiconductor | +27.98% | 34.1% | +16.84% |
| Samsung Electronics | +26.81% | 35.5% | +15.78% |
Key points:
- Roughly one third of the +30% headline (31.8%–35.5%) is broad Korean market beta, not the memory trade.
- Additive excess = 27.98 − 9.53 = 18.45pp. Multiplicative excess = 1.2798 ÷ 1.0953 − 1 = +16.84%. Cross-market work should use the multiplicative form because it is insensitive to baseline drift.
- The point is not that the rally was fake. The point is that the true memory-specific excess is about +17%, not +30% — and that 13pp gap will distort any position sizing or valuation anchor that takes the headline as an input.
Contamination layer 2: why the "extreme U.S.–Korea divergence" was a fake signal
The same day, a cross-market alignment tool printed what looked like a very strong signal: a U.S.–Korea memory spread of −11.52pp, sitting at the 99th historical percentile. That reading is wrong, and the cause is calendar misalignment.
What broke
- The Korean leg used the 7/31 close (Korea had already finished trading).
- The U.S. leg used the 7/30 close (the U.S. had not yet opened on 7/31).
- The two legs are a full trading day apart. Subtracting them does not measure cross-market mispricing — it measures the one-day time offset.
The fix: two-day window alignment
Restate both markets as cumulative 7/30 × 7/31 returns, and the windows match:
| Name | Market | 7/30 | 7/31 | Two-day cumulative |
|---|---|---|---|---|
| Samsung Electronics | KRX | −0.72% (implied) | +26.81% | +25.90% |
| Micron (MU) | US | +18.36% (close) | +4.07% (pre-mkt) | +23.18% |
| SK Hynix | KRX | −5.64% (implied) | +29.95% | +22.62% |
| Western Digital (WDC) | US | +15.37% (close) | +5.08% (pre-mkt) | +21.23% |
Key points:
- Aligned max-minus-min spread = 25.90 − 21.23 = 4.67pp.
- Raw reading −11.52pp / 99th percentile → aligned reading 4.67pp / unremarkable. The signal disappears completely.
- All four names land in a tight +21.23% to +25.90% band — the U.S. and Korea moved essentially in lockstep, and there was no arbitrageable cross-market dislocation.
- Korea's implied 7/30 returns were negative (Hynix −5.64%, Samsung −0.72%) while U.S. names surged on 7/30. That up-day and down-day landing on different calendar dates is precisely what manufactured the fake divergence.
A reusable two-step check
Before drawing any conclusion from a cross-market or cross-sector move comparison, clear these two gates:
- Strip the beta. Divide the name's move by the median of a same-market, same-day, different-theme control basket. What remains is the thematic excess. Reading absolute moves alone will always overstate it.
- Align the calendar. Confirm both legs cover exactly the same time window. If either side is unclosed or unopened, restate both as multi-day cumulative returns before comparing.
Rule of thumb: when a cross-market spread signal suddenly prints at the 99th percentile, suspect the calendar before you suspect the market. An extreme percentile is more often produced by a misaligned data pipeline than by a genuine dislocation.
What actually changes
| Untreated reading | Treated reading |
|---|---|
| Korean memory +30% in a single day | True thematic excess +16.84%; ~1/3 is market beta |
| Extreme U.S.–Korea divergence, 99th pct | Spread 4.67pp, no dislocation |
| A cross-market convergence trade exists | It does not — the signal came from calendar offset |
| "A pure memory rally" | The whole Korean market rallied; memory was the amplifier, not the sole driver |
Limits of this method (what we did not do)
To keep the conclusion falsifiable, these constraints have to be disclosed alongside it:
- The control basket is not an index. The non-memory baseline is the median of five Korean large caps, not the KOSPI itself. Using the actual index would shift the exact beta shares; the order of magnitude holds, the precise figures would not.
- The U.S. 7/31 leg is a pre-market quote, not a close. Pre-market liquidity is thin and reversible. The two-day figures must be recomputed after the official 7/31 U.S. close.
- This is a one-day snapshot, not a statistical inference. It is a cross-sectional decomposition of a single session and forecasts nothing about what follows.
- Korean 7/30 returns are implied, back-solved from the two-day cumulative and the 7/31 close rather than read directly.
- Hanmi Semiconductor is an HBM thermo-compression bonder equipment supplier, not a DRAM manufacturer. Including it in the "memory complex" is a thematic grouping, not a same-tier supply-chain grouping.
FAQ
Q: Was the +30% move fake?
A: No. The prices are real. But about one third of it came from a broad Korean market rally that day; only roughly +17% (multiplicative) is attributable to the memory theme itself.
Q: Why use the multiplicative form instead of simple subtraction?
A: Simple subtraction (18.45pp) distorts when the baseline move is large. The multiplicative form, 1.2798 ÷ 1.0953 − 1 = +16.84%, measures how much more the name earned relative to the market, and is more comparable across markets and time windows.
Q: Where did the −11.52pp number come from?
A: A data pipeline subtracted Korea's 7/31 close from the U.S. 7/30 close. The legs sit one trading day apart, so the output is a time offset, not a price spread. This was the second recurrence of the same fault.
Q: Is a 4.67pp aligned spread large?
A: No. Four names inside a +21.23% to +25.90% band is normal dispersion for assets moving in the same direction at the same time.
Q: Does this only apply to memory stocks?
A: No. It applies to any "market A name vs market B name" move comparison, and it matters most when the two trading sessions do not overlap — Korean or Japanese equities versus U.S. equities, for example.