After the Selloff, Is the KOSPI Really Cheap?

The KOSPI fell as low as 5,450.81 intraday on October 15. A circuit breaker halted trading for 20 minutes from 12:38 p.m. Samsung Electronics closed down 9.48% at ₩210,000 and SK hynix lost 11.06% to ₩1,262,000. Foreign investors were net sellers for an eighth consecutive session. Reports differed on the amount: approximately ₩3.6 trillion (Alpha Economy: ₩3.6341tn) versus approximately ₩4.8 trillion (NewsPim: ₩4.8196tn).⁹ Closing reports cited hotter-than-expected U.S. consumer inflation and fears of rate hikes; debt pressure among AI companies, highlighted by CoreWeave; doubts about AI investment returns; and concerns about CXMT's expansion. It is not possible to assign a precise share of the day's decline to any one factor.

On October 15, the Financial Services Commission (FSC) convened an emergency market review chaired by Chairman Lee Eog-weon. It identified growing expectations of U.S. rate increases, uncertainty around the AI infrastructure industry and rising oil prices as external risks. Market experts and related agencies attending the meeting assessed that Korean equities had, in some respects, reacted excessively in both the speed of the decline and trading flows. The FSC said it was preparing to activate contingency measures promptly if needed and urged financial institutions to refrain from indiscriminate margin lending.¹³ That judgment of an “excessive” reaction does not establish that the market is undervalued at its current P/E. Nor had any market-stabilization measures actually been deployed.

Imagine a stock expected to earn ₩100 per share over the next 12 months, trading at ₩450. Its forward P/E is 4.5×. What that price implies depends on the multiple we think those earnings would normally deserve: the assumed reference P/E.

All three scenarios produce the same price. So it is misleading to say, “A 4.5× P/E prices in a 55% earnings decline.” The precise statement is: “If the reference P/E remains 10×, a 55% reduction in forecast earnings would reconcile that multiple with today's price.” Price alone cannot separate earnings weakness from multiple compression. P/E ratios also reflect growth, risk and differences in cash distributions such as dividends.

Cutting forecasts is not the same as reporting lower profits

A downward revision may mean a company earns less than it does today—or that it earns more, just not as much as analysts had expected. The Big Two, Samsung Electronics and SK hynix, have a combined 12-month operating-profit forecast of ₩907 trillion (October 12).¹⁶ Cut that forecast in half and it becomes ₩453.5 trillion, only 3.1% below the ₩468 trillion earned or estimated over the latest four quarters (Snowshagal calculation). Here, a 50% cut to forecasts is a very different proposition from profits actually halving.

The ₩468 trillion is not entirely confirmed, either. Samsung's third-quarter operating profit of ₩107.4 trillion is a preliminary estimate issued before its books are closed¹⁷, while SK hynix's third-quarter contribution comes from a consensus of brokerage forecasts.¹⁶

Different brokerages, different P/Es—even before updating the price

Brokerage forward P/Es dated October 8–13 ranged from 4.6× to 5.5× for the same index. Providers differ (Refinitiv, FnGuide and Quantiwise), as do coverage rules—whether only ordinary shares are included, how companies without earnings forecasts are treated, and when estimates are updated. Such differences produced gaps of 0.3–0.4 P/E points even on the same day: 5.5× versus 5.1× on October 8, and 4.9× versus 4.6× on October 13.^{6, 16, 18, 19} Holding each source's earnings estimates fixed and updating only the price to October 15 gives a range of 4.2–4.6× (Snowshagal calculations). These are not newly compiled October 15 consensus P/Es.

Eugene's distribution since 2005 puts the bottom-decile P/E range at 5.3–8.5×. Both its October 13 estimate of 4.9× and the roughly 4.5× repriced for October 15 are below that historical range.¹⁶ In its October 13 note, Eugene called the situation “difficult to compare with historical precedents.”

The KOSPI's earnings forecasts are heavily concentrated in two companies. Eugene estimated that the Big Two would account for 68% of the index's Q3 operating profit and 76% of net profit attributable to shareholders¹⁹. On a 12-month forward operating-profit basis, the two account for ₩907 trillion—or 71%—of the KOSPI's ₩1,280 trillion total (October 12; Snowshagal calculation).¹⁶ These operating-profit shares should not be confused with the net-earnings denominator used to calculate forward P/E.

How much does the answer change with the assumed P/E?

Even using the same P/E of roughly 4.5× (4.51×), a 5.3× reference multiple corresponds to a 15% earnings haircut, while a 12× multiple implies 62% (Snowshagal calculations). As prices fell between October 13 and 15, the implied haircut increased by 3–7 percentage points at each unchanged reference multiple.

The right reference multiple depends on industry mix and where earnings sit in the cycle. Daishin notes that cyclical shares tend to command low P/Es when profits are high and higher P/Es when profits are depressed. In its analysis, the KOSPI's cyclical group saw earnings forecasts fall 42% between late 2021 and September 2023, but market capitalization fell only 9%; the group's P/E therefore rose. Since late 2025, the relationship has reversed: forecast earnings have multiplied 4.8 times, while market capitalization has increased only 2.1 times.¹⁸

Valuation theory points to the same problem. NYU's Aswath Damodaran cautions that extrapolating boom-year profits for cyclical firms overstates value. Normalizing earnings also requires normalizing the reinvestment and financing costs needed to sustain them. Earnings can be inflated at a cyclical peak, making the P/E look unusually low; at a trough, the reverse applies.²¹ His work does not establish an appropriate P/E for the Korean market.

Under Daishin's classification, cyclical sectors—those whose earnings move sharply with business conditions—accounted for 68% of the KOSPI's market capitalization as of October 8.¹⁸ With cyclical companies carrying such weight, part of the 55% haircut implied by a 10× reference P/E may be attributable not to lower earnings but to the low multiples these businesses commonly trade on near peak profitability (Snowshagal assessment). We have not established that 10×, roughly the historical median, is the appropriate benchmark for today's index.

A · Doubts about future earningsCXMT

One name recurred in coverage of the October 13 plunge: Chinese DRAM manufacturer CXMT (ChangXin Memory Technologies). At an industry conference in Shanghai on October 10, its president, Cao Kanyu, said the company had made “key progress” on 4F² DRAM architecture, which reduces cell area, and planned to introduce server-grade DDR5 RDIMMs based on that architecture by year-end. International reports carried the remarks on October 12.¹ Server DRAM has largely been supplied by Samsung Electronics, SK hynix and Micron. The prospect of another competitor raised concerns about pricing and margins. CLSA expects CXMT's DRAM shipment share to increase from 12% this year to 14% next year.¹ Concerns about more supply from CXMT's expansion were again cited in October 15 closing reports.⁹

There is contrary evidence. Asked on October 14 whether Samsung was also developing 4F², Samsung Electronics CTO Song Jai-hyuk replied, “Yes.”³ Samsung has previously said it aims to mass-produce sub-10nm DRAM using new technology by 2028. SK hynix also disclosed a 4F² roadmap last year. CXMT has not released detailed product specifications or a precise launch date. According to TechInsights data cited in coverage, the three established memory makers are producing server products on the 1c (sixth-generation 10nm-class) process, while Chinese manufacturers remain at least two generations behind.¹

Memory-related discounting is not confined to Korea. As of October 12, four U.S. memory and storage names—Micron, Sandisk, Seagate and Western Digital—were down an average 33.5% from their peaks. GPU designers Nvidia and AMD, together with foundry TSMC, were an average 7.5% below their 52-week highs. DS Investment & Securities suggested the gap could reflect investor attention shifting toward compute rather than outright weakening memory demand.²²

B · A lower multiple, even without an earnings cutRates · AI financing

The price investors will pay for a given stream of earnings can fall even if forecasts do not change. From October 1 to 14, the U.S. 10-year Treasury yield held between 5.19% and 5.31%; the 30-year yield ranged from 5.52% to 5.67%.²³ Higher-than-expected September consumer inflation raised fears of rate hikes, but⁹the U.S. 10-year yield closed lower on October 14 than the day before,²³ and Korea's 10-year government-bond yield fell 6.8 basis points to 4.193% on October 15.¹¹ After the Korean close on October 15, U.S. September producer prices also came in well above forecasts, rising 1.3% month-on-month.^{14, 15} The question here is not whether yields rose or fell on any single day, but whether financing will remain expensive. On BlackRock's October 14 earnings call, CEO Larry Fink pointed to historically large government deficits and a generational AI infrastructure buildout that companies must finance. He expects structurally higher interest rates to persist for longer, particularly at the long end of the curve.²⁴ Broadcom CEO Hock Tan was also reported as saying at an October 12 event in Singapore that, as chip shortages ease, access to capital could become the biggest constraint on AI infrastructure expansion.²⁵

On October 14, those concerns showed up in one company's securities. Shares of U.S. data-center lessor CoreWeave dropped 15.5%. The yield on its 9.625%-coupon bonds due 2032 rose to 14.8%, as the price fell below 78 cents on the dollar, according to a report citing Bloomberg data. CEO Michael Intrator told employees that markets can misread a company and behave irrationally; he maintained that its backlog more than covered all obligations.⁷ This is a signal from a single company and a single bond maturity—not a measure of borrowing costs or default risk across the AI industry.

Financing pressure can compress multiples, but it also links back to Path A. If funding constraints force data-center operators to slow capital spending, expectations for GPU, HBM and server-DRAM orders and prices may change. No such shift has yet been established. Higher interest rates can raise the return investors demand on future earnings, but rates also embed growth expectations. The “normal” P/E does not fall mechanically with yields.

C · Forced selling before fundamentals changeMarket flows

Direct observations. On October 13, forced sales against unsettled stock-purchase obligations reached ₩59.35 billion, the highest since July 31 (₩122.01bn), according to Korea Financial Investment Association figures reported by SBS Biz.²⁶ Foreign investors remained net sellers for an eighth session on October 15, with estimates ranging from roughly ₩3.6 trillion to ₩4.8 trillion, depending on the reported tally.⁹ These figures do not capture all derivatives liquidations. Nor can they establish how much of the index's decline was caused by forced selling.

Positioning before the crash. BofA's October Global Fund Manager Survey (161 respondents overseeing $453bn; fielded October 2–8) found the largest overweight in technology since August 2020 and a cash allocation of 3.9%. “Long global semiconductors” was named the most crowded trade by 45% of respondents.²⁷ Because the survey preceded the selloff, it does not show how much of that positioning was subsequently unwound.

Brokerage interpretation and the regulator. Yuanta Securities suggested that reductions in total-return-swap (TRS) and contract-for-difference (CFD) exposures may have contributed to the October 12–13 slump. It expected selling pressure to ease quickly once those positions had been sharply reduced.²⁸ That easing had not been demonstrated by October 15. Nor is there direct evidence separating those liquidations from the reported foreign net-selling totals. Since September 2023, Korean CFD transactions have been classified by the underlying investor—retail, institutional or foreign. Routing an order through an overseas broker alone does not make it a foreign investor's sale.²⁹ FSC Chairman Lee Eog-weon's October 15 warning that investments relying on excessive debt and margin financing are risky is also relevant to this channel.¹³

Still unverified. The actual size of liquidations, who initiated the sales, and their contribution to the index decline. Position unwinds can drive prices down without establishing that the underlying businesses are worth less.

1 · Discounting peak-cycle earningsDaishin Securities

Daishin decomposed an 8.95-percentage-point rise in the KOSPI's earnings yield (E/P, the inverse of P/E), from 9.3% at end-2025 to 18.2% on October 8. It allocated 1.21 points to sector-weight changes, 1.44 to interest rates, 6.81 to doubts about cyclical earnings and −0.52 to other factors.¹⁸ In that model, the cyclical-earnings discount accounts for approximately 76% of the increase in earnings yield over the period (Snowshagal calculation). That is a model classification, not an observed attribution of the last few days' losses.

Daishin also estimated that the P/E would have been 8.4×, rather than 5.5×, if sector weights and interest rates were the only influences. Treating 8.4× as the benchmark translates into a 35% implied earnings haircut for the KOSPI: 1 − 5.5 ÷ 8.4. Keep that October 8 benchmark fixed and update only the price to October 15, and the implied cut is about 45% (conditional Snowshagal calculation). Cyclical sectors—memory, equipment, chemicals, steel and shipbuilding, among others—traded at 4.5× earnings, or 0.44 times the 10.3× P/E of noncyclicals, versus an average relative multiple of 0.99 since 2015. Daishin interpreted this as the market valuing only 44% of current cyclical earnings as sustainable (October 8 basis).¹⁸ Those numbers do not isolate differences in growth, risk or cash distributions—other factors that also determine multiples.

2 · Regressing index levels on operating profitEugene · Illustrative scenario

Eugene regressed the KOSPI's annual average level in 2000–2025 against each year's aggregate KOSPI operating profit. The fitted equation is “Index = 9.0008 × operating profit (₩tn) + 697.05,” with an R-squared of 0.87.¹⁶ Eugene reported that cutting only the Big Two's 12-month forward operating-profit forecasts—Samsung ₩527tn and SK hynix ₩380tn as of October 12—by 60%, while leaving the other companies' ₩373tn unchanged, would give ₩730tn of KOSPI operating profit and an index level of 7,270. A 70% cut gave ₩650tn and 6,550. Taking the lower edge of its ±1,000-point scenario band (6,270 and 5,550), Eugene characterized KOSPI levels of 5,500–6,200 as consistent with a 60–70% cut in Big Two earnings. Recalculating from the stated inputs produces somewhat different figures.

Putting the October 15 close of 5,554.22 into the same equation produces very different implied earnings, depending on where the index is assumed to sit relative to the regression line (Snowshagal calculations).

3 · Market-implied ROEHanwha Investment & Securities · Illustrative

Using valuations as of October 13, Hanwha estimated a “market-implied ROE” in the 12% range, below both the 12-month forward ROE of 25.5% and trailing ROE of 15.7%. It interpreted the price as discounting lower profitability.³⁰ Hanwha did not publish the full reverse-calculation formula, so we could not independently reproduce the model.

Samsung's trailing four-quarter operating profit fell from ₩63tn in Q3 2018 to ₩28tn in Q4 2019, and from ₩58tn in Q2 2022 to ₩6.6tn in Q4 2023. SK hynix fell from ₩21tn in Q3 2018 to ₩2.1tn in Q1 2020, and from ₩15tn in Q2 2022 to a loss of ₩10tn in Q3 2023.¹⁶ Using the displayed amounts, the declines work out to 56%, 89% and 90%, respectively (Snowshagal calculations).

Direct comparison would be misleading. The historical figures are realized peaks and troughs for each company; today's figures are hypothetical outcomes for both combined. In Eugene's published scenarios, the Big Two's earnings after 60% and 70% cuts (₩362tn and ₩270tn, respectively) are 23% and 42% below their ₩468tn trailing four-quarter total, which includes preliminary and estimated Q3 figures. The October 15 regression center-line estimate (₩167tn) is 64% below that base; its lower-band and upper-band assumptions (₩278tn and ₩56tn) imply 41% and 88% declines (all Snowshagal calculations). The published scenarios' 23% and 42% gaps are smaller than the historical 56–90% declines; the center-line and upper-band figures fall within that range. But all of these comparisons mix company-specific realized cycle extremes with combined-company hypothetical figures. They cannot establish the likely depth of the next downturn.

Conversely, a sharp selloff in the bonds of a few AI companies does not establish that the entire chip upcycle is over. TSMC's and ASML's latest disclosures are evidence that semiconductor and equipment demand remains strong. Assume a 10× reference P/E, and today's mid-4× valuation can be reconciled with a 55% cut to projected earnings. But the equation does not reveal what earnings decline the market actually expects.

The questions now are less about how large the forecasts look on paper than how long those earnings can last, how much cash they produce, whether AI investment remains financeable, and whether the discount investors apply to earnings narrows. A low multiple can persist even if earnings estimates hold steady. The table below therefore looks first at price changes alongside like-for-like forward earnings revisions; the P/E is the outcome of that comparison.

Compare the next update against these same-source Eugene figures. When a new P/E is available, make sure its price and earnings dates line up. Switching data sources can change the reported multiple by 0.3–0.4× even for the same day (Snowshagal calculation). For calendar-year 2027, Eugene's October 12 KOSPI operating-profit consensus was ₩1,340tn (FnGuide), versus ₩1,344tn in Kiwoom's October 14 report.^{6, 16}