Foreign Flow Watch: The Board That Rose Is the One Retail Was Selling — Week of September 28, 2026

Foreign Flow Watch is a weekly count, not a forecast. Every window in this issue is defined in trading sessions, never in calendar days, and every figure carries the date we read it.

📅 What “this week” means in this issue

This issue counts four trading sessions of flow, not five: Monday 28, Tuesday 29 and Wednesday 30 September and Thursday 1 October 2026. Friday 2 October traded, and its index close is final, but its investor-flow rows were not settled when we read them — 2,039 of 2,544 close values still disagreed with the settled price table. We do not include an unsettled session in a flow count, so the flow window stops at 1 October and the index window runs to the 2 October close.

Basis, stated before the numbers. Unless a line says otherwise, every flow figure in this issue is the 2,544 KOSPI and KOSDAQ common stocks in our coverage universe, which excludes ETFs and preferred lines. Where we add the Samsung Electronics preferred (005935) we say so and label the total “all-ticker.” Figures are our own database, built from Korea Investment & Securities’ Open API; each session’s net share count is valued at that session’s settled close and the results summed, which is an approximation and not execution prices. Index levels are the published daily closes of the KOSPI and KOSDAQ. FX is a single rate, ₩1,360.59/$1, the 1 October close, applied to every won figure in this issue so that dollar comparisons between sections are internally consistent. It is not the rate that applied on each individual session.

📈 The two boards went different directions, and the gap is the story

The KOSDAQ outperformed the KOSPI by 6.87 percentage points in five sessions. That is a wide spread for a market where the two boards usually move together, and the flow table below does not explain it in the direction you would expect.

Published daily index closes. The KOSPI’s low of the window was 6,838.04 on 30 September, 3.43% below the 23 September close; it recovered 2.42% over the following two sessions. The KOSDAQ rose in four of the five sessions, and 4.48% of its 5.78% came in a single session, 1 October (855.91 → 894.29).

Two things are worth separating here. The KOSPI did not drift down and back; it fell three sessions in a row and then turned. The KOSDAQ did not drift up either; it added 4.48% on one day. A weekly number averages both of those away, which is why this column prints the path and not only the endpoints.

💸 Net buying by investor category, four sessions and twenty

Our flow database, the 2,544 common stocks of the coverage universe, each session’s net shares valued at that session’s settled close, FX ₩1,360.59/$1. The two windows are both four sessions and both on the same 2,544-stock common-stock basis, so the columns are comparable in length and in universe — which is the only reason we print the comparison at all. ⚠️ Holding the universe constant matters more than it sounds here. The Samsung preferred was a net purchase in the earlier window and a net sale in this one, so an all-ticker comparison would have shown part of the reversal as an artefact of which tickers are counted rather than as a change in behaviour. Both columns therefore exclude it, and it is reported on its own in the concentration section below.

Both domestic categories flipped sign between the two windows. That is unusual enough to say plainly, and it is also a warning about reading any single four-session window as a stance. Retail were the market’s largest sellers in the run-up to Chuseok and its largest buyers in the week after. We are not going to call that conviction in either direction.

🔍 Was this a market-wide sale, or two stocks?

Two names, 97.6% of the foreign net selling. And then a third ticker, outside the common-stock universe, that most weekly write-ups leave out entirely.

Same database and FX. The “every other stock” row is computed directly, not by subtraction, and the two routes agree to the won: −₩7,246.7bn plus −₩176.9bn is exactly the −₩7,423.6bn universe total, and adding the preferred’s −₩600.8bn gives the −₩8,024.4bn all-ticker figure. ⚠️ The preferred line’s retail figure is not a zero — it is absent. Our feed supplies no retail value for 005935 on any of these four sessions, so we print “not provided” rather than fill it. It is a small gap in coverage terms (8 of 10,183 rows in the window, across 2 tickers) and a large one here, because 005935 is the second-biggest single-ticker foreign sale of the week and we cannot say who took the other side.

Where those two sit inside each possible universe. This second table deliberately has no flow wording in its headers, because the rows are subsets and supersets of a market total rather than market totals themselves — and our own publishing gate compares anything labelled as a market flow against the whole universe, which would make an exclusion row look wrong when it is right. We reproduced every figure here twice, by direct query and by subtraction, and they agree to the won.

Won is the primary unit in this table by design. At ₩1,360.59/$1 the common-stock universe figure is −$5.46bn, which is the dollar number used everywhere else in this issue. We deliberately do not restate the all-ticker figure in dollars — it stays in won, here and in the takeaways, because it is built on a non-standard universe and a loose dollar version of it would travel without the caveat attached. The retail column is identical in the last two rows because the preferred has no retail value to add — not because retail did nothing in it.

Two rows there do the work. “Retail bought $1.59bn” is a true sentence about the market total and a misleading one about retail behaviour: $1.74bn of buying went into two stocks and the other 2,542 saw net selling. One stock, Samsung Electronics, accounts for more retail buying than the entire universe total. And the amber row is the part that normally disappears: the preferred is a separate ticker with its own order book, ticker 005935 ran another −₩600.8bn of foreign net buying on its own order book, and it sits outside the common-stock universe that almost every weekly flow number is built on — including ours, which is why we label the two totals differently rather than quietly mixing them. We wrote last week about which Samsung ticker you are actually buying; this is the flow version of the same problem.

🔁 What did foreigners buy instead?

Semiconductor equipment and materials — and retail sold every one of them.

Same database and FX. Boards are from our coverage universe. Six of the seven are semiconductor equipment, materials or inspection names; four of those six are KOSDAQ-listed.

Set that against the index table and a shape appears. Over the four flow sessions foreigners sold the KOSPI and bought the KOSDAQ, while retail did the reverse — the board-level figures are in the table below. The board foreigners bought is the board that rose, and the board retail bought is the board that fell. We are not claiming the flows caused the move. The KOSDAQ’s net foreign buying is two orders of magnitude smaller than its KOSPI selling and cannot lift a board 5.78% on its own, and the single-session 4.48% jump is far larger than any flow number in this issue. What we can say is narrower: the usual story told after a week like this, that retail buying held the market up, is the opposite of what the board-level tape shows.

Board assignment from our coverage universe, so these rows sum exactly to the universe totals above: −₩7.556tn plus +₩0.132tn is the −₩7.424tn foreign figure, and +₩3.017tn less ₩0.857tn is the +₩2.159tn retail figure. The Samsung preferred sits outside both boards, which is one more reason its −₩600.8bn is reported separately rather than folded in. Index columns are the published closes and cover five sessions, one more than the flow columns — shown side by side for shape, not as a matched comparison.

🧾 Why did these three see the largest buying? We could not find out

We checked, and there is nothing filed. Woori Financial Group, Jusung Engineering and HPSP — the three largest foreign net buys in the window — filed no disclosures of any kind with Korea’s electronic disclosure system between 25 September and 3 October 2026. We queried each company’s filing list directly and all three returned an empty result.

That is a real answer, not a gap we are papering over. This column screens price and flow moves first and then looks for the event, because events decided outside a company do not generate filings — a competitor’s result, an index rebalance, a foreign broker’s model change. For these three the honest statement is: the buying is in the tape, the reason is not in the filings, and we are not going to supply one.

And on Woori there is a harder fact than the buying: the stock fell while it was being bought. Woori Financial Group (316140) closed at ₩36,400 on 28 September and ₩33,250 on 1 October, −8.65% across the same four sessions in which foreigners net bought ₩625.3bn of it. Retail added ₩51.0bn and institutions were flat at −₩1.6bn. Those three categories together were net buyers of ₩674.7bn, and the price still dropped 8.65% — so the selling came from the category our three-way feed does not break out, the residual that Korea’s tape calls “other corporations.” We do not know who that was or why. It is the single clearest case this week of a flow number pointing one way and the price pointing the other, and it is the reason this column prints the price move next to the flow instead of letting the flow stand alone.

One further note, without claiming causation: Woori was the subject of our own deep dive six days before this window opened, which decoded its buyback, share swap and CET1 position. We have no evidence connecting the two and we are not implying any; we flag it because a reader who saw both might otherwise assume we had.

🧱 The corporate bid has a date on it, and it is close

Two buyback programmes have been the steadiest domestic bid under Samsung Electronics and SK hynix all quarter. Both are now near their limits, and the two are not the same instrument.

Progress is our own daily tracking of the exchange’s treasury-purchase disclosures, read on 3 October 2026. The SK hynix cancellation filing puts its planned cost at ₩40,004,340,000,000 and its on-market purchase period at 20 August to 19 November 2026, with shares issued of 730,492,365 on the filing date. The authoritative check is each company’s execution report, and neither has been filed yet — so these percentages are our running estimate, not a disclosed figure.

On the recent pace, and counting only actual trading sessions — 5 October is a substitute holiday for National Foundation Day and 9 October is Hangeul Day, so October loses two sessions — Samsung’s budget runs out around 6 October and SK hynix’s around 15 October. Both are well inside their disclosure windows, which close on 21 and 19 November respectively.

The distinction in the right-hand column matters more than the dates. SK hynix is retiring 24.07 million shares; that supply is gone. Samsung is buying for employee compensation, which means the same purchases that support the price now become supply later. A reader who treats the two programmes as interchangeable will draw the wrong conclusion about what happens when each one ends.

☢️ The nuclear basket, week three of four

Issue #9 found that four listed Korean nuclear-chain companies saw net foreign selling across both a three-session and a twenty-session window, at a time when the sector narrative was strongly positive. Issue #11 set the falsifier: if that flow turned net positive over both windows within two weeks, the finding was a two-week artefact.

Tickers 034020, 051600, 052690 and 083650. Same database and FX. The twenty-session window is 2 September to 1 October 2026 — counted as twenty actual sessions in the data, not as a calendar month: September lost 24 and 25 to Chuseok, so a calendar-month window would have been nineteen sessions and we would have labelled it wrongly. This is week three of the four-week observation Issue #9 opened; the window closes on 10 October 2026 and we will print the result either way.

Worth adding, because it landed inside this window: on 1 October Doosan Enerbility filed a new supply contract — and its region field reads Vietnam. The O Mon III combined-cycle EPC, counterparty Vietnam National Industry – Energy Group, ₩837,970,880,392, 4.91% of recent revenue (rcpNo 20261001800034). The United States announced its Korean investment projects on 30 September; the one Korean nuclear-chain contract filed the day after points at Vietnam.

📊 Scorecard: how Issue #11 held up

Issue #11 left five falsifiers. Two of them have resolved, and both went the way the issue said they would. We print the ones that have not resolved as unresolved rather than quietly dropping them.

🧩 Prove this issue wrong

• If Friday 2 October settles and the four-session picture reverses — if adding that one session moves the four-session foreign total materially away from the figure in the category table above, or flips retail’s sign — then our decision to stop the window at 1 October cost readers the week’s actual shape. We will re-run the identical count next issue with five sessions and print both.

Corrected 10 October 2026: the second item below originally read “the 90.4% figure”. The concentration figure in this article is 97.6% — it appears three times in the headline, the summary and the table, and 90.4% appears nowhere else and was supported by nothing. It was a typing error in the falsifier list, not a second measurement. The underlying number is unchanged.

• If the two-stock concentration breaks down on a fifth session, the 97.6% figure is a four-session artefact rather than a feature of the week. The ex-two row is the line to watch: it is the one that says retail were net sellers of everything else.
• If Samsung Electronics’ or SK hynix’s execution report lands below our running estimate, the 6 October and 15 October exhaustion dates move out and the “corporate bid has a date on it” framing weakens.
• If the KOSDAQ’s 1 October jump turns out to have a disclosed cause in one or two large constituents, then describing the board-level divergence as unexplained was premature. We looked at the three largest foreign buys and found no filings; we did not enumerate every KOSDAQ constituent, and we are saying so rather than implying we did.
• Our weakest number this issue is the single FX rate. Using ₩1,360.59/$1 across four sessions in which the rate moved between 1,350.50 and 1,360.59 makes the dollar figures internally consistent and individually approximate. Every won figure in this issue is the primary one.

📚 Lingo Check

🎯 Why it matters for K-Export Stars

This column exists because the categories in Korea’s investor tape are routinely reported as if they were opinions — “foreigners are leaving,” “retail is defending the market.” Four sessions of data say something narrower and more useful this week: foreigners sold two memory names and bought semiconductor equipment, retail did the exact opposite, and the board that rose was the one retail was selling.

None of that is a forecast. The two reversals between windows — both domestic categories flipping sign in four sessions — are the reason we keep printing the comparison window and the path, not just the week’s endpoints. A number that reverses this fast is not a stance.

Related reading: why SK hynix’s non-operating income matched its operating profit · which Samsung division earns the money · Woori Financial Group, decoded · every issue of Foreign Flow Watch.

Disclaimer: This article is for informational and educational purposes only and does not constitute investment advice, a recommendation, or an offer to buy or sell any security. Flow figures are our own computation from a broker data feed, valued at settled closes and therefore approximate; index levels and filing figures are as published, each carrying the date on which we read it. Disclosure of interest: the author’s portfolio holds Samsung Electronics, SK hynix, Mirae Asset Securities, Kia, Hanwha Ocean, Hana Financial Group, LigaChem Bio, SK Biopharmaceuticals, Hyundai Rotem, Dalba Global, Kolon TissueGene and a KOSPI 200 ETF. Corrections policy: we publish them.