Foreign Flow Watch: Foreigners Sold $6.0bn and the KOSPI Ended the Week Flat — Week of September 18, 2026

Foreigners sold $6.02bn of Korean shares in five sessions this week. The KOSPI finished the week down 0.23%. Those two facts sit side by side because a single category — the one Korean data calls “other corporations” — bought $5.92bn, and bought almost exactly the same amount on the day the index fell 3.26% as on the day it rose 2.66%. This is also the first issue of this column computed after Korea started producing two closing prices per stock. We audited our own pipeline before writing a word of it, found one field already carrying the wrong one, and are telling you which close is behind every number below.

This is Foreign Flow Watch, our weekly read of where global capital actually moved in Korean equities. Every figure below is computed from settled exchange data for the 2,547 KOSPI and KOSDAQ common stocks in our published universe that carried a settled Korea Exchange price on 18 September 2026. Week of 14–18 September 2026, five trading sessions. Previous issue: Issue #9, week of 11 September. The full series lives at the Foreign Flow Watch hub.

📈 Both boards went nowhere. The path did not.

💸 Foreigners sold $6.0bn. One category bought almost all of it.

The daily breakdown is where the week stops being a flat line.

Foreigners sold on four of five sessions and bought on the fifth, which happened to be the week’s largest gain. Individuals did the reverse: they bought $2,155.1M into Monday’s 3.26% fall and sold $2,717.9M into Friday’s 2.66% rise. Those two lines are close to mirror images, and either one on its own would support a tidy story about who is right. We do not know who is right, and neither does anyone else yet.

🧱 Why Was the Corporate Bid the Same Size Every Day?

Because it is not a view on the market — it is two disclosed buyback programmes executing against a filed won budget and a daily order cap, which is a mechanical process and looks like one.

Read the last column of the daily table again. In won, the “other corporations” net buy ran ₩1,484.5bn, ₩1,628.5bn, ₩1,668.6bn, ₩1,598.5bn and ₩1,667.9bn. The largest day is 12.4% above the smallest, and the smallest day was Monday — the 3.26% crash, the session on which an opportunistic buyer with ₩1.6tn a day to deploy would have leaned in hardest. It did not. It bought slightly less than average and then bought the same amount every day afterwards. On its own that is suggestive and no more — five sessions is a short window, and two paragraphs below we widen it and show what breaks.

Korea lets you check that inference against a document rather than guess at it. Both of the country’s largest buyback programmes were filed in August, with the won amount, the window and the daily order-quantity cap printed in the filing: SK hynix’s ₩40,004,340,000,000 programme running 20 August to 19 November 2026 (rcpNo 20260819000254), and Samsung Electronics’ ₩14,999,999,992,000 programme running 24 August to 21 November 2026 (rcpNo 20260821000616). Tracking the residual in each name since its start date gives this:

Two facts make that attribution much harder to argue with than the flatness alone. First, concentration: of the week’s ₩8,048.1bn of market-wide residual buying, SK hynix accounts for ₩5,537.8bn and Samsung Electronics for ₩2,437.0bn — ₩7,974.8bn, or 99.1%, in two stocks. The third-largest, Hyundai Steel at ₩41.0bn, is one-hundred-and-thirty-fifth of SK hynix. Whatever this category is doing, it is doing it in the two companies that filed buyback programmes and almost nowhere else.

Second, and more decisively, the category switched on. Across the forty-one sessions from 22 June to 19 August 2026 the market-wide residual never once left the band −₩167.0bn to +₩235.0bn. On 20 August — the first day of SK hynix’s filed window — it printed +₩1,071.3bn. On 24 August, the first day of Samsung’s, +₩1,600.6bn, and it has stayed near that level since. A six-fold step change landing on two filed start dates is a stronger piece of evidence than any amount of day-to-day steadiness, and it is the test we should have led with.

Why this series starts on 22 June and not earlier. The residual is a four-category identity, so it cannot be computed for a session in which one of the four categories is missing. In our own data the individual-investor field is absent for every stock on 18 and 19 June 2026, and the sessions before that cover 25 stocks rather than the full universe. 22 June 2026 is the first session on which all four categories exist for the whole universe, so it is the first session this measurement is defined. We are stating that rather than quietly starting the chart wherever it flattered the argument — and we checked: all 63 sessions from 22 June to 18 September carry a complete four-category record for 2,547 to 2,555 stocks, with no gaps.

The run-rate estimate barely moved in a week: 36.7 → 36.6 sessions for SK hynix and 29.1 → 29.5 for Samsung, against filed windows of roughly 63. That leaves about 15 and about 10 more sessions of buying. With the Chuseok closure on 24–25 September and public holidays on 5 and 9 October, both land in the first half of October 2026 — six weeks inside the filed windows. That has been this column’s standing claim since Issue #8, and a second week of data did not move it.

🔁 Did Foreigners Sell SK hynix and Samsung This Week?

Yes — and they sold both stocks over twenty sessions as well, which is the stronger statement, because a single window invents narratives and two windows that agree do not.

The two companies are 48.5% of our universe’s market value and carried 73.9% of all foreign net selling over five sessions and 85.9% over twenty. Foreign selling in Korea this month has been, to a first approximation, foreign selling of two memory stocks.

SK hynix is the sharper case. Over the five sessions 14–18 September 2026, foreigners net sold $2,532.3M of SK hynix. SK hynix rose 2.48% over that window. Over the twenty sessions 24 August – 18 September 2026, foreigners net sold $7,593.7M of SK hynix. SK hynix rose 7.34% over that window. Flow and price disagree on both, in the same direction, in the second-largest company in the country.

Samsung is the contrast. Its price and its flow point the same way over twenty sessions — net selling against a 7.28% fall — and opposite ways over five, where a 0.58% gain sits on top of $1,916.3M of net selling. Same sector, same week, two different relationships between flow and price, which is why this column prints both windows and refuses to pick one. The two stocks closed the week at $1,345 (₩1,857,000) and $189 (₩261,000) respectively, both as of 18 September 2026.

One detail worth keeping: Friday reversed inside SK hynix. Foreigners bought 723,470 shares on 18 September after selling on each of the four preceding sessions (−1,215,512, −586,836, −262,655, −701,796). The weekly net of −2,043,329 shares is arithmetically correct and hides a turn on its final day. For context on how these four categories are constructed and where they mislead, see our reference piece on how foreign and institutional flows move Korean stocks.

🕓 Does Korea’s New Evening Session Change These Numbers?

Not the flow figures — we verified them against the exchange’s own investor-trend records and they match to the share — but it did change one price field inside our own database, and we found it by looking before we wrote this issue rather than after.

On 14 September 2026 the Korea Exchange opened a continuous after-market running 16:00–20:00 KST. Yesterday we published a measurement of what that did to end-of-day data: in a 60-stock sample, a widely used market-data library’s close matched the KRX settled close on 239 of 239 stock-days before launch and differed on 221 of 240 after, with the daily percentage change carrying the opposite sign on 14% of stock-days. Publishing that and then computing a flow column off the wrong close the next morning would have been the least defensible thing this column could do. So we ran three checks first.

The third row of the audit table is the uncomfortable one, and it is ours. A field we own carried the unofficial price for four days and nothing in its name said so. That is precisely the failure we described in another organisation’s data yesterday. We found it because we assumed we had it too.

The consequence for a reader is not abstract. Because a weekly change is a ratio of two closes, using the evening price at one end changes the answer — and sometimes changes its sign:

Hanwha Aerospace is the case that should worry a screen-builder: a $40.3bn defence name, one of the twelve export champions this column tracks, whose week reads as a small loss on the exchange’s numbers and a small gain on a feed’s. The settled close on 18 September was ₩1,078,000, which was also the session’s regular-hours low; the feed carries ₩1,092,000. Neither number is fabricated. Only one of them is the closing price. If you want the mechanics of the session that produces the second one, we wrote them up when the rules were published: what actually changed on 14 September, and the broader market-structure background in our Nextrade explainer.

📊 The export-champion scoreboard: 5 sessions against 20

Five of twelve disagree between the windows. A reader shown only the five-session column would say foreigners were selling Kia, Samsung Biologics, Samyang Foods and APR this week; the twenty-session column says they have been net buyers of all four over the month. A reader shown only the twenty-session column would say they were selling KAI; the week says they bought it. Neither reader is wrong about the number in front of them. Both would be wrong to call it positioning. Two names survive both windows and have the price agreeing with the flow. LIG Defense & Aerospace was bought in both and rose 9.73% on the week. Hanwha Aerospace was sold in both (−$20.1M and −$37.5M) and fell in both (−0.28% and −0.65%) — but its weekly move is the one that flips sign on a feed close, as the previous section showed, so the agreement in that row is worth exactly as much as the close convention behind it. LIG Defense & Aerospace’s 9.73% survives any convention; Hanwha’s 0.28% does not.

☢️ The nuclear basket, one week on

Issue #9’s lead finding was that all four companies in Korea’s nuclear export chain were net sold by foreigners in both windows while every one of them rose, one by 41%. We set a falsifier: if the basket’s foreign flow turns net positive over both windows within four weeks while the prices hold, then “sold in both windows” was an artefact of one fortnight’s tape. Week one of four:

The flow half of the test is unchanged: all four net sold in both windows, and the five-session selling more than doubled, from −$78.9M to −$190.5M. The price half broke in the other direction — every one of the four fell this week, KEPCO E&C by 11.97%. So the juxtaposition that made Issue #9’s headline (“rose, and foreigners sold all of it”) no longer describes the current week, while the flow claim it rested on got stronger. Both halves are in the table. Over twenty sessions the four are still up between 6.28% and 46.23%, which is what a one-week fall of this size looks like inside a rally. Three weeks of the falsifier remain.

🏛️ Three Samsung independent directors bought — and the dates in the filing are not the trade dates

Foreigners sold $1,916.3M of Samsung Electronics this week. In the same window, three of the company’s own independent directors filed reports showing on-market purchases. The amounts are trivial. The fact that you can see them at all, with prices attached, three days later, is not.

Three of the four line items name a price the stock did not trade at on the date beside it. That is not an error in the filing — it is the form working as designed, and the form says so in its own footnote: “where shares are traded on the securities market, the date is the settlement date” (증권시장에서 주식등을 매매한 경우에는 그 계약결제일). Korean equities settle on T+2. Shift each date back two sessions and every price lands inside that day’s range: Shin’s ₩268,000 sits in 10 September’s ₩263,500–₩270,500; Lee’s ₩273,500 in 9 September’s ₩267,500–₩275,000; Cho’s ₩259,250 in 11 September’s ₩256,500–₩261,500. Her second line, ₩250,250 against 16 September, is consistent with both readings, so we leave it ambiguous rather than round it into the story.

🇯🇵 SK hynix answered the exchange on Friday, and the answer was “undetermined”

SK hynix filed exactly one disclosure between 12 and 19 September 2026: an inquiry-disclosure response, undetermined, on 18 September (rcpNo 20260918800583). It re-answers a Korea Exchange demand issued at 11:35 on 21 August 2026 about a press report headlined “SK hynix to build a semiconductor plant in Japan.” The word re-answers matters: the filing’s own related-disclosures field records that the company replied the same afternoon, at 16:03 on 21 August, with an “undetermined” answer. Friday’s document is the three-month re-disclosure that answer obliged it to make. The full text, translated: “To strengthen the competitiveness of our memory business we are reviewing various options including the establishment of additional production bases, but nothing has been determined to date. We will re-disclose at the point at which specifics are determined, or within three months.” The filing names a re-disclosure date of 17 December 2026 and is signed by CFO Kim Woo-hyun.

Friday was also the session on which SK hynix rose 6.42% and foreigners turned net buyers of it for the first time that week. Both of those are figures from settled exchange data; neither appears in the filing, and we are not connecting them to it. The filing is the same answer the company gave four weeks earlier, on a schedule set in August — there is no new information in it to connect anything to.

What the filing does is convert a rumour into a calendar entry. There are now two of them: the company used the same form on 9 September to say nothing about reported plans for its Chongqing packaging stake (rcpNo 20260909800559), itself a re-disclosure of an August answer, with the next date set at 8 December. Two dated appointments, 8 and 17 December, on questions a company elsewhere could simply decline to answer. For a foreign investor the useful output is not the answer; it is the date.

🧾 Scorecard: how Issue #9 held up

Issue #9 left five testable conditions. Two resolved this week and neither resolved our way; three remain open on schedule. We use failed below in two different senses and flag which is which: the residual test failed against us — we made a claim and the evidence did not support it — while the 620,000-share test failed as a test, meaning the signal it was watching reversed, which says nothing about whether we were right.

What we are doing about the test that failed four times out of four. Issue #9 said of it, in its own words, that it was “still badly written — this is the second consecutive rewrite” and that it “fires every week by construction and tells us almost nothing.” We then rewrote it a third time, with a size floor, and it produced four clean candidates. All four failed. When a test has been rewritten three times and returns the same verdict on the fourth pass, the defect is not the wording. It is the design: we were asking a disclosure regime to confirm a number it was never built to confirm.

Korea’s filing rules are triggered by who holds what — crossing 5%, a 1% change by an existing large holder, a largest-shareholder change, a treasury-share decision. Our residual measures who traded on the exchange this session, in a category that lumps every corporate share dealing together. Those two things overlap sometimes and coincide exactly almost never. Issue #9’s own POSCO example made the point without our noticing: a filing did appear, and it was roughly twice the size of our residual.

Our running record of grading ourselves: Issue #4 scored Issue #3 at two failures of three; Issue #5 scored Issue #4 at three holds and one open; Issue #6 scored Issue #5 at two failures of three; Issue #7 scored Issue #6 at three tests that could not run; Issue #8 recorded one of its own tests as badly written; Issue #9 recorded the same test failing after a rewrite. This week it failed four times out of four and we withdrew it. We print the bad weeks at the same size as the good ones, and this is a bad week.

🧩 Prove this issue wrong

• If the “other corporations” daily bid keeps running near ₩1.6tn after both buyback windows are exhausted — on our estimate, from roughly the middle of October — then attributing that steadiness to the two programmes is wrong, and the category is measuring something much larger that we have mislabelled. This is the cleanest falsifier this column has, because it has a date and requires no filing.
• If either execution report shows materially less than ₩24.06tn (SK hynix) or ₩10.16tn (Samsung) bought through 18 September, our residual-based completion percentages of 60.14% and 67.71% are upper bounds we leaned on too hard. Still the authoritative check; still not yet available.
• If the nuclear basket’s foreign flow turns net positive over both windows in the next three weeks, Issue #9’s central finding was a two-week artefact. Week one did the opposite: the selling doubled.
• If the broker feed’s close column in our own database still carries the evening price for sessions older than one day when we check again next Saturday, then our description of it as a one-session restatement lag is wrong and the contamination is deeper than four days. We will re-run the same three checks and print the result either way.
• If the pre-announced Lee Jae-yong transfer of Samsung Electronics shares settles on 12 October and produces any visible trace in the four-category flow data, then our claim that off-exchange transfers are invisible to this dataset is wrong, and the residual is measuring something broader than we think. Checkable on 12 and 13 October.

Our weakest number this week is the 20-session dollar column, for the same reason as last week and in the opposite direction. The won weakened 3.15% across the five sessions alone, from ₩1,338.2 on 11 September to ₩1,380.3 on 18 September, reversing part of the 5.4% strengthening Issue #9 measured across its 20-session window, 14 August to 11 September. Across the full 20-session window the rate ran from ₩1,383.6 to ₩1,380.3 — nearly flat end to end, but with a ₩1,337.9 low in between. A 20-session dollar figure computed at a single end-date rate therefore differs from one computed session by session by around 2% for a stock whose flow holds one sign, and by more where the sign flips inside the window. Every dollar figure in this issue uses session rates and every table says so. Share counts and won figures are exact.

📚 Lingo Check

🎯 Why it matters for K-Export Stars

A single KOSPI print told you Korea was flat this week. Underneath it, foreigners removed $6.02bn, retail sold another $530.6M, and a single corporate category put back $5.92bn at a pace that did not vary by more than 12% from the worst day to the best. Two memory stocks were 73.9% of the foreign selling over five sessions and 85.9% over twenty, and one of them rose in both windows while being sold in both. Meanwhile three of Samsung’s independent directors bought 1,358 shares between them on dates that are not the dates printed in their filings, and SK hynix told the exchange it has nothing to confirm about a Japanese fab until 17 December.

None of that requires a view on memory pricing. It requires reading the filings, matching them to settled exchange data, and printing the disagreements instead of the tidy version — including the disagreement we found inside our own database this week. The reason this column can do that is structural: Korea makes issuers disclose a buyback’s won budget and window before they buy, makes insiders report what they bought and at what price, and makes companies answer the exchange on the record when a rumour moves a stock. Western coverage rarely reads those documents because they are published in Korean. For how the flow categories themselves are built, start with how foreign and institutional flows move Korean stocks; for the two stocks at the centre of this issue in dollar terms, Samsung and SK hynix stock prices in USD; for the reactor chain, our nuclear export chain map and Doosan Enerbility valuation; and for what the evening session actually changed, Korea’s two closing prices.

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 or in Hanwha Aerospace, HMM, Hyosung, Korean Re, Hyundai Department Store or Kolon Industries. 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 as of 18 September 2026 and go stale quickly; prices, flows and programme-completion estimates in particular. Do your own research and consult a licensed adviser before investing.