Data as of: all figures are 2026-08-03 Korea Exchange (KRX) closing prices and same-session post-close investor-type flow data, benchmarked against the 2026-07-31 KRX close. Every ratio below can be recomputed from the raw share counts and volumes given in the tables; nothing depends on undisclosed data. See "Limits of this method" at the end.
Bottom line
The reading "foreign investors reversed and dumped Korean memory" fails at all three layers. The apparent symmetry in percentage terms is an artifact of collapsing turnover — measured in shares, foreigners sold back only half and remain net buyers across the two sessions. The investors who actually liquidated Friday's entire position were domestic institutions, and on Samsung they moved in exactly the opposite direction from foreigners. And what foreigners sold was the semiconductor leg, not Korea: they were net buyers of Korean autos and batteries in the same session. The drawdown itself also needs its base: after the fall, the two names are still +18.53% and +15.70% over the two-day window, still ahead of the KOSPI.
Price: put the drop back into its two-day window
On Friday 2026-07-31 the Korean memory complex closed near limit-up and the KOSPI rose +17.91%, its largest single day since 1985. The next session, Monday 08-03, gave part of it back.
| Name | 07/31 close | 07/31 move | 08/03 close | 08/03 move | Two-day |
|---|---|---|---|---|---|
| SK Hynix (000660.KS) | KRW 1,718,000 | +29.95% | KRW 1,567,000 | -8.79% | +18.53% |
| Samsung Electronics (005930.KS) | KRW 262,500 | +26.81% | KRW 239,500 | -8.76% | +15.70% |
| KOSPI | — | +17.91% | — | -5.35% | +11.60% |
Both directions matter:
- On the day: both underperformed the KOSPI (Hynix by 3.44pp, Samsung by 3.41pp) — the give-back was indeed concentrated in the memory leg.
- Over two days: both still outperformed the KOSPI (by 6.93pp and 4.10pp) — the rally was not erased.
Calling 08-03 a "crash" in isolation means letting day two's minus sign overwrite day one's plus sign. A mechanical give-back the session after a limit-up move, and a fresh leg down, are different events.
Layer one: the symmetry is manufactured by collapsing volume
The single most misread table is this one — foreign net flow as a share of each stock's traded volume flips from the low twenties positive to the low twenties negative, almost a mirror image.
Foreign flows: percentage versus shares
| Name | 07/31 foreign net buy | % of volume | 08/03 foreign net sell | % of volume | Sold back | Two-day net |
|---|---|---|---|---|---|---|
| SK Hynix | +2.206m shares | 21.3% | -1.107m shares | -20.4% | 50.2% | approx +1.10m |
| Samsung Electronics | +8.359m shares | 14.4% | -3.896m shares | -14.2% | 46.6% | +4.463m |
The percentages look symmetric. The share counts are not. There is exactly one reason: the denominator collapsed.
Volume: the denominator halved
| Name | 07/31 volume | 08/03 volume | Change |
|---|---|---|---|
| SK Hynix | 10.35m shares | 5.43m shares | -47.5% |
| Samsung Electronics | 57.88m shares | 27.39m shares | -52.7% |
With turnover cut roughly in half, the same percentage represents half as much stock. Hence:
- Hynix: sold back 50.2%, still a two-day net buy of about 1.10m shares
- Samsung: sold back 46.6%, still a two-day net buy of 4.463m shares
Foreigners did not reverse; they took half of it off. "The percentage went negative" and "the position went negative" are different claims, and they must never be used interchangeably on a day when volume moves this much.
Layer two: extend the window and the roles of foreigners and institutions invert completely
Change the window length and the answer reverses. The natural read is the two-session cumulative, 07/31-08/03: foreigners net buyers, institutions net sellers — from which one concludes "the actual liquidation was domestic institutions." That is arithmetically correct, but the window begins on 07/31 — the KOSPI's largest single-day gain since 1985 (+17.91%). Extend the window and every sign flips.
SK Hynix (millions of shares, positive = net buying)
| Window | Foreign | Institutional |
|---|---|---|
| 2 sessions 07/31-08/03 | +1.098 | -0.011 |
| 3 sessions 07/30-08/03 | +1.543 | -0.091 |
| 5 sessions 07/28-08/03 | -1.205 | +1.089 |
| 10 sessions 07/20-08/03 | -2.014 | +0.527 |
| 20 sessions 07/03-08/03 | -4.398 | +0.867 |
Samsung Electronics (millions of shares)
| Window | Foreign | Institutional |
|---|---|---|
| 2 sessions 07/31-08/03 | +4.463 | -1.421 |
| 3 sessions 07/30-08/03 | +5.119 | -1.659 |
| 5 sessions 07/28-08/03 | -4.986 | +3.677 |
| 10 sessions 07/20-08/03 | -7.361 | +1.021 |
| 20 sessions 07/01-08/03 | -23.346 | +3.287 |
From the fifth session onward, two stocks and two investor categories flip four signs at once, and the 10- and 20-session windows agree with the five, not the two.
The correct statement is therefore the reverse of what the two-session window implies: foreigners have been the persistent net sellers through this episode and domestic institutions the persistent net buyers. The two-session reading that made institutions look like the sellers is a rebound artifact. Institutions were genuine net sellers on 08-03 alone (Samsung -5.040m, Hynix -0.288m shares), but that follows heavy accumulation on 07/29 (Samsung +5.972m shares); it is a give-back, not the start of a liquidation.
This is the same error the first layer of this piece criticises, in a different variable. Layer one swapped the unit (percentage of volume to share count); this layer swaps the window length. Any claim about who is buying and who is selling must state both the unit and the window — omit either and the opposite result can be reproduced.
Layer three: they sold the semiconductor leg, not Korea
If foreigners were exiting Korea, the selling would be spread across sectors. The actual distribution is highly concentrated.
Same-session foreign net flow, as a share of each stock's traded volume
| Net bought | % | Net sold | % |
|---|---|---|---|
| Kia | +11.6% | Hanmi Semiconductor | -20.8% |
| Hyundai Motor | +11.0% | SK Hynix | -20.4% |
| Samsung SDI | +5.9% | Samsung Electronics | -14.2% |
| LG Electronics | +0.7% | LG Energy Solution | -11.0% |
The sell column is entirely semiconductors and batteries. The buy column is entirely autos and finished goods.
This is intra-market rotation, not a country exit. The two imply opposite forward paths: in the first the money is still in Korea waiting to rotate back; in the second it has already left the country.
A reusable two-step check
This failure mode is not specific to Korea. It applies to any market that publishes investor-type participation percentages.
- When a percentage flips sign, check the denominator first. Pull the stock's traded volume for both sessions. If volume moved more than roughly ±30%, cross-day comparison of the percentage is already invalid — convert back to shares or notional value before comparing.
- Decompose "who sold" down to investor type; do not stop at "foreigners." Foreign, institutional and retail are three independent legs and can move in opposite directions simultaneously. Reporting only one of them can invert the entire conclusion.
- (Bonus) Put the single-day move back in its window. If the prior session was an extreme up move, the next day's minus sign is most likely the back half of the same event rather than a new one. The two-day cumulative move, plus its position relative to the index, is closer to the truth than either day alone.
Falsifiable follow-ups
This conclusion can be overturned by the next print, so the tracking points come stated up front:
- Whether foreigners keep selling to flat, or stop here. If they subsequently sell the remaining two-day net buy, the "only took half off" characterization has to be rewritten. If they stop, this piece holds.
- Whether the institutional liquidation spreads from Samsung to Hynix. The Hynix institutional book at -0.011m shares over two days is essentially flat, making it the most sensitive observation point.
- Whether the auto and battery buying persists. If the auto-leg net buying disappears the following session, "sector rotation" must be downgraded to "one-day noise."
Limits of this method (read these with the conclusion)
- This piece uses only settled KRX data and contains no same-day U.S. equity readings. Korean prices here are closing values while U.S. sessions would be an opening snapshot; subtracting one from the other across markets produces false signals.
- "% of volume" is defined as that investor category's same-day net buy/sell share count divided by the stock's total traded share volume, with positive meaning net buying. Every percentage in this piece can be recomputed from the share counts and volumes shown.
- Share counts are reported to one decimal place in units of 10,000 shares, so small rounding residues exist. The Hynix two-day net derived from 2.206m minus 1.107m is stated as "approx +1.10m" because the unrounded source value is about +1.098m.
- Window disclosure (added in the correction): any cumulative net flow changes sign with the length of the window. This piece reports five windows (2/3/5/10/20 sessions); readers should treat the five-session and longer windows as the conclusion. The two-session window begins on a historic rebound day and is shown only to demonstrate that short windows mislead.
- This piece contains no portfolio holdings, position sizes or trade actions. It is public market data and arithmetic performed on it.
- Not investment advice.
Sources
- Prices and volumes: Korea Exchange (KRX) closing data for 2026-07-31 and 2026-08-03
- Investor-type flows: KRX same-session post-close net buy/sell statistics by investor category (foreign / institutional)
- Index: official KOSPI closing change