Foreign Flow Watch: Retail Sold $4.2bn and 98.9% of It Was One Stock — Chuseok Week, September 23, 2026

Korea traded for three days this week and then shut for five. In those three sessions the KOSPI rose 2.71% and Korean individual investors net sold $4.20bn of stock — the kind of number that writes its own headline about retail investors de-risking before a long holiday. That headline would be wrong. Strip out two companies and the same investors were net buyers.

📅 What “this week” means in this issue

This issue covers three trading sessions, not five: Monday 21, Tuesday 22 and Wednesday 23 September 2026. The Korea Exchange closed on Thursday 24 and Friday 25 September for Chuseok, Korea’s autumn harvest holiday, and reopens on Monday 28 September. We documented that closure in advance in our 2026 Korea market calendar, checked against the exchange’s own holiday feed.

Every window in this issue is defined in trading sessions, never in calendar days, and every table says which. Where a previous issue used a five-session week, this one uses three — so a like-for-like comparison with Issue #10 is not available on the weekly column, and we do not make one. What we do instead is carry a three-session block from the previous week (16–18 September) alongside it, so the comparison is three against three, and a twenty-session window as the standing second window this column always prints.

📈 Both boards rose into the closure

Index levels as published by the Korea Exchange (data.krx.co.kr), settled closes. The three-session change measures from the 18 September close to the 23 September close. Daily percentages are the exchange’s own.

Both boards went up, and they went up on every session bar one small KOSDAQ dip. The KOSPI gained 186.69 points in three days. That is the surface. The interesting part is who was on each side of it.

💸 The four categories, three windows

Korea’s exchange splits every session’s net buying into four investor categories. Three are reported directly — foreigners, institutions and individuals. The fourth, “other corporations,” is the residual: non-financial corporate entities, which includes an issuer buying back its own shares.

Method, stated once. Share counts are the Korea Exchange’s investor-trend series. Each day’s net shares are multiplied by that day’s settled close and summed — an approximation of money traded, not execution prices. The won total is converted to dollars once, at ₩1,376.0/$, the mean of the Bank of Korea base rate (ECOS series 731Y001, item 0000001) across 21–23 September: ₩1,383.8, ₩1,384.3 and ₩1,360.0. Computing the three-session figures session by session instead moves the foreign total by 0.79% and the retail total by 0.16%. Universe: our published coverage set of KOSPI and KOSDAQ common stocks, with ETFs and preference shares excluded. Exactly 2,544 names have flow records on each of 21, 22 and 23 September; the twenty-session window spans 2,551 and the prior three-session window 2,548, because listings enter and leave. Each column is computed over the names present in its own window, and we print the count rather than asserting one universe for all three. The four categories sum to zero by construction, which we check on every window.

Read that table on its own and the story is simple: foreigners and individuals sold, institutions and companies bought, and the index went up anyway. Every window agrees. It is also, at the aggregate level, not a story at all — retail has been the largest net seller in all three windows, including the week before the holiday, so nothing about the holiday explains it.

The table becomes a story only when you take it apart.

🔍 Did Korean Retail Investors Sell Before the Holiday?

No. Retail investors net bought 1,605 of Korea’s 2,544 listed common stocks in those three sessions and net sold only 836; the entire $4.20bn net-sell aggregate is Samsung Electronics, plus SK hynix.

Same method, universe and rate as above. The ex-two column removes Samsung Electronics (005930) and SK hynix (000660) only. “Sign changes” compares the two columns for that category.

Two of the four categories change sign. Retail goes from the market’s largest net seller to a net buyer. Institutions go from a $1.24bn net buy to roughly flat. And the corporate bid — $3,573.4M across the whole market — falls to $40.7M once two names come out, which is to say it was never a market-wide corporate bid at all.

The individual-stock numbers behind that are blunt. Retail net sold $4,148.8M of Samsung Electronics and $872.0M of SK hynix in three sessions. Those two alone are 119.6% of the market-wide retail figure — more than the total, because everything else was on the other side. The breadth count says the same thing in a different unit: retail net bought 1,605 of 2,544 names and net sold 836. The remaining 103 had no retail net flow at all, and we counted them as their own line rather than adding them to either side: not one of the 103 traded a single share in any of the three sessions, so calling them bought or sold would have been an invention. They are 4.0% of the list and 0% of the money.

The daily path adds one more thing the intuitive story gets wrong. If individuals were flattening positions ahead of a five-day closure, the selling should build toward the last session. It did the opposite: retail net selling was $2,135.2M on Monday, $1,061.8M on Tuesday and $1,000.1M on Wednesday (₩2,938.1bn, ₩1,461.0bn and ₩1,376.2bn at ₩1,376.0/$) — largest on the first day, smallest on the last. We are not claiming that disproves holiday positioning; we are saying the sequence does not support it, and we would have had to ignore the sequence to publish that story.

🔁 Who Bought the Samsung Shares Retail Sold?

Foreigners took $1,881.6M of them, institutions $1,165.2M and the corporate residual $1,101.9M — and Samsung closed the three sessions up 9.39%.

Won figures rather than dollars in this table, because they are the units in which the four categories sum to zero on each session — the figures printed here are rounded to ₩0.1bn, so a column can show a 0.1 remainder that is rounding and not a gap. Closes are KRX settled closes; the 9.39% measures from the ₩261,000 close of 18 September 2026. At ₩1,376.0/$ the three-session totals are foreigners +$1,881.6M, institutions +$1,165.2M, individuals −$4,148.8M, other corporations +$1,101.9M.

Samsung’s market capitalisation rose from ₩1,525.9tn to ₩1,669.1tn across those three sessions on 5,846,278,608 common shares (Source: KRX daily market data, 18 and 23 September 2026). Trading volume ran 18.2–22.3 million shares a day against 11.8 million on 17 September.

📊 The second window disagrees, and that is the point

Here is where a single-window column would mislead you. Over three sessions foreigners bought $1,881.6M of Samsung Electronics. Over twenty sessions — the same window ending on the same day — they are net sellers of $1,551.0M.

Windows: 3d = 21–23 September 2026; 20d = 27 August–23 September 2026. Price changes measure from the settled close before each window to the 23 September settled close. Same method and rate as above. “Windows agree” compares the sign of the two foreign columns only.

Samsung’s reversal is close to exact in share terms. Foreigners are net sellers of 9,257,897 shares over the twenty sessions and net buyers of 9,254,028 shares in the last three of them — which means they sold roughly 18.5 million shares over the first seventeen sessions and bought back half of that in three days. That is a reversal, not accumulation, and anyone who read only the weekly column would have the direction of the month exactly backwards.

It is not an isolated case. Of the 2,438 stocks with non-zero foreign flow in both windows, 712 — 29.2% — have opposite signs in the three-session and twenty-session columns. In a three-session week that proportion should be read as a warning about the short window, not as a finding about rotation.

SK hynix is the opposite case and the more informative one: net sold in both windows, up in both windows. Foreigners have sold 4,957,255 shares over twenty sessions while the stock gained 10.31%. We flagged the same divergence in Issue #10 and it has not closed. If you want the background on why foreign flow and price routinely disagree in Korea, our explainer on how foreign and institutional flows work in Korean stocks sets out the mechanics.

🧱 The corporate bid was two buybacks, and it has an expiry date

The “other corporations” residual bought $3,573.4M this week. Two companies were effectively all of it.

Residual computed per stock as −(foreign + institutional + individual) net shares × that day’s settled close, then summed. “Third-largest” is the largest residual net buy excluding Samsung Electronics and SK hynix on that session. Converting the won column above at ₩1,376.0/$, the market-wide residual is $1,194.0M, $1,174.9M and $1,204.7M on the three sessions, of which Samsung plus hynix are $1,179.4M, $1,154.4M and $1,199.1M; the third-largest names are $7.5M, $7.9M and $5.2M. Those dollar figures are converted from the rounded won shown, so the last digit can sit $0.1M either side of the unrounded computation; the share column is computed from the unrounded values.

Both companies have filed treasury-share acquisition decisions on DART that name the amount and the window. SK hynix filed for ₩40,004,340,000,000 running 20 August to 19 November 2026 (receipt no. 20260819000254). Samsung Electronics filed for ₩14,999,999,992,000 running 24 August to 21 November 2026 (receipt no. 20260821000616).

Residual accumulated from each programme’s filed start date through 23 September 2026 — 25 sessions for SK hynix, 23 for Samsung. “Sessions left” divides the unspent amount by that company’s mean residual over 21–23 September (₩1,115.0bn and ₩505.4bn per session). Won figures only; these are compared against won-denominated filings, so no conversion applies.

Korea’s October calendar has two closures — Monday 5 October (substitute for National Foundation Day, which falls on a Saturday) and Friday 9 October (Hangeul Day). Counting only the sessions that exist, 6.6 more sessions from 23 September reaches roughly 6 October for Samsung, and 11.3 sessions reaches roughly 14 October for SK hynix. Both are comfortably inside their filed windows, which run to 21 and 19 November. We made a similar estimate in Issue #8; this week’s numbers are consistent with it.

🕓 Can a Five-Day Holiday Break a Market Data Feed?

Yes — and it broke one of ours, in a way that is invisible unless you check. Korean broker feeds commonly publish a provisional price during the session and restate it to the settled close overnight. That restatement is tied to the next trading day, not the next calendar day. Over a five-day closure, a price field that is normally stale for hours stays stale for days.

We check this every issue because we found it in Issue #10, and this week it produced the clearest result yet. Our database carries two price columns: a broker-feed close attached to the flow records, and the Korea Exchange's settled close in a separate price table. We compared them for every stock on each of the last four sessions.

Measured 26 September 2026, three calendar days after the 23 September session. The comparison is exact equality of the two price fields for the same stock on the same date. The denominator here is 2,545, not the 2,544 used in the flow tables: this test runs over every flow record for the date with no universe filter, so it includes Samsung Electronics preferred (005935), which the flow tables exclude. That one line is the entire difference. Positive control: the same test finds 2,544 of 2,545 matches on 22 September, which is how we know the comparison itself works and the 23 September result is not a broken query.

Four-fifths of the 23 September price field is still carrying an intraday value, three days after the session ended, because the restatement waits for 28 September. On an ordinary Wednesday this would have cleared by Thursday morning. Every won and dollar figure in this issue uses the Korea Exchange settled close, from the separate price table, and none of them uses the feed column. Share counts and net-share figures come from the flow records, which are unaffected — it is only the attached price that lags.

This matters beyond our own plumbing. If you hold Korean stocks through any vendor feed, the closing prices you see for a session immediately before a long closure may be provisional for the length of the closure, and nothing in a field named close will tell you so. We wrote up the broader version of this problem when Korea’s evening session launched in Korea Now Has Two Closing Prices; the holiday adds a second, independent way for the same field to be wrong.

🏛️ Two director filings, one reconcilable price

Issue #10 found that the date column in a Korean officers’ and major shareholders’ ownership report is the settlement date, not the trade date — the filing form’s own footnote says “계약결제일” (contract settlement date) — so with T+2 settlement you subtract two sessions before matching the stated price to a chart. Both Samsung and SK hynix filed one of these reports on 22 September 2026. We re-ran the test on both.

Sources: DART receipt nos. 20260922000524 (Samsung) and 20260922000361 (SK hynix), both filed 22 September 2026. “Only session it fits” tests the stated price against each session’s KRX published low and high over 14–23 September 2026 and reports the sessions where low ≤ price ≤ high. In that ten-session window each price fits exactly one session.

Positive control for the test itself, using no filing data: fed SK hynix's own 16 September settled close of ₩1,759,000 (Source: KRX), the same routine returns 16 September among its matches — so a null result below is a real null and not a broken query.

Samsung’s filing replicates the finding exactly. Heo Eun-nyeong’s 200 shares at ₩255,000 are dated 21 September, a session whose published range was ₩263,500–₩275,000 — that price was not available. Two sessions earlier, on 17 September, the range was ₩251,500–₩259,000, and ₩255,000 sits inside it. No other session in the surrounding ten works. She is the fourth Samsung independent director to file an on-market purchase inside six days, after the three Issue #10 covered, which were filed on 17 and 18 September 2026. The amounts remain trivial — 200 shares is about $37,000 — and we would not draw a signal from the size.

SK hynix’s filing does not reconcile, and we cannot resolve it. Jeong In-chul’s 142 shares at ₩1,682,000 are dated 16 September. Under the settlement-date convention the trade would be 14 September, whose published low was ₩1,686,000 — ₩4,000, or 0.24%, above the stated price. Reading the date as a trade date does not help: 16 September’s low was also ₩1,686,000. The only nearby session where ₩1,682,000 was available is 15 September, which is neither convention.

☢️ The nuclear basket, week two of four

Issue #9 found that all four companies in Korea’s nuclear export chain were net sold by foreigners in both windows while every one of them rose. We set a falsifier: if the basket’s foreign flow turns net positive over both windows within four weeks, that finding was an artefact of one fortnight. This is week two.

Same windows, method and rate as the tables above. Basket total sums the dollar columns only; price columns are not additive and are marked as such. The 20-session window here (27 August–23 September) is not the window Issue #10 used (21 August–18 September), so the 20-session price column is not comparable with that issue’s figures — only the direction of the flow test is.

The falsifier is not triggered. All four are still net sold in both windows, with the basket selling −$165.7M over three sessions and −$572.9M over twenty. The price half has now also turned: all four fell over the three sessions, and three of the four are down over twenty. That is a change from Issue #9, where the juxtaposition was “rose, and foreigners sold all of it.” The juxtaposition that made the headline no longer holds; the flow claim it rested on is unchanged. Two weeks of the falsifier remain. The company-by-company background is in our nuclear export chain filing map and, for the largest name, our Doosan Enerbility valuation.

🧾 Scorecard: how Issue #10 held up

The fourth row is the one worth pausing on. We published a rule last week, applied it this week to two fresh filings, and it worked on one and not the other. A rule that holds one time in two is not a rule, it is a hypothesis, and we would rather say that than quietly report only the case that worked.

🧩 Prove this issue wrong

• If Samsung Electronics’ or SK hynix’s treasury-share execution report shows materially less than ₩11.67tn or ₩27.40tn bought through 23 September 2026, our completion percentages of 77.82% and 68.50% are upper bounds we leaned on too hard, and the session-count estimates go with them. Still the authoritative check; still not filed.
• If the “other corporations” residual is still running near ₩1.6tn a day after both programmes are exhausted — on this week’s pace, from roughly mid-October — then attributing that steadiness to the buybacks is wrong and the category is measuring something larger. This is the cleanest falsifier this column has: it has a date and needs no filing.
• If a later SK hynix filing prints an after-hours execution, or the company’s next ownership report uses a date that reconciles on T+2, we will know which of the three readings of the ₩4,000 gap was right. Until then we cannot resolve it: it is an unexplained gap, and we are not filling it with whichever explanation we find most interesting.
• If the feed’s close column for 23 September is still mismatched after the 28 September session settles, our description of a one-trading-session restatement lag is wrong and the contamination is structural. We will re-run the identical test next issue and print it either way.
• If the nuclear basket’s foreign flow turns net positive over both windows in the next two weeks, Issue #9’s central finding was a two-week artefact.
• Our weakest number this week is the 20-session dollar column, and here is how weak. The Bank of Korea base rate ran from ₩1,384.6 on 27 August to ₩1,360.0 on 23 September, with a ₩1,337.9 low on 10 September, and we convert the whole issue at a single ₩1,376.0/$. Recomputing session by session moves the three-session figures by 0.79% (foreign) and 0.16% (retail), the twenty-session market-wide figures by 1.86% and 0.57%, and Samsung Electronics’ 20-session foreign figure by 3.81% — the largest, because its flow changes sign inside the window. Share counts and won figures are exact; every dollar figure is an approximation produced by that conversion, and the single ₩1,376.0/$ convention is a simplification we are naming rather than hiding.

📚 Lingo Check

🎯 Why it matters for K-Export Stars

Three things in this issue are the reason this column exists.

First, aggregates hide their own composition. A wire service can accurately report “Korean retail investors net sold 5.8 trillion won ahead of the Chuseok holiday” and leave every reader with a false picture. The correction requires no special access — only the willingness to subtract two stocks and recount. That is the whole method: the exchange publishes the data, and almost nobody decomposes it.

Second, the two largest positions in most foreign Korea portfolios are being supported by a bid with a known expiry. Samsung Electronics and SK hynix have between them bought roughly ₩39tn of their own stock since late August, filed the size and the dates on DART, and are on our estimate 68% and 78% through. When that bid stops in October, the largest daily buyer in the market stops with it. Nothing about that is a prediction of price; it is a change in the flow arithmetic that a holder should know is coming. Our guide to Samsung and SK hynix prices in USD covers how to follow both names from outside Korea.

Third, a holiday is a data problem before it is a market problem. Four-fifths of one of our price fields was wrong for three days for a reason that has nothing to do with markets and everything to do with when a vendor runs its overnight job. We caught it because we test it; we publish the test because a reader relying on the same class of feed has the same exposure and no way to see it.

This is Issue #11 of Foreign Flow Watch. Previous issues, and the series introduction, are on the Foreign Flow Watch hub; the issue immediately before this one is Issue #10, week of 18 September.

Disclosure of interest. The author holds positions in Samsung Electronics (005930) and SK hynix (000660), two of the stocks this issue discusses, and in other Korean equities not named above. The author holds no position in any of the four nuclear-chain companies — Doosan Enerbility, KEPCO KPS, KEPCO E&C or BHI — or in Samsung Electro-Mechanics, SK Square, Hanwha Aerospace, Hyundai Steel or KT&G. No position was opened or closed in any stock named above in connection with this article.

Disclaimer: This article is for informational and educational purposes only and is not investment advice, a recommendation, or an offer to buy or sell any security. Figures are drawn from Korea Exchange investor-trend and price data and from DART filings as cited, and are accurate as of 23 September 2026 unless otherwise dated; market data changes and derived estimates such as buyback completion are our own calculations from a residual, not company statements. Do your own research and consider consulting a licensed financial adviser before making investment decisions. If we get a number wrong we correct it publicly — see our corrections policy.