Semis Are Falling While the S&P 500 Rises: How Do You Tell a Single-Sector Unwind From the Start of a Broad Selloff?

Bottom Line

Semiconductors are falling while the S&P 500 is rising. So far this is money leaving one sector, not money leaving the market. The evidence is not the size of the drawdown but three structural indicators: whether the decline gradient is monotonic, whether index implied volatility is being pushed up, and whether hedging demand sits on the index leg or the single-name leg. All three currently point to a single-cluster momentum unwind, not broad de-risking. But Korea has produced the first genuine change in character: foreign investors flipped from net buyers to heavy net sellers.

Asking the Right Question

When memory chip stocks fall 7-9% in a session, the instinctive question is "is this the start of a crash?" That question has no answer, because it collapses two entirely different phenomena into one:

• Single-cluster momentum unwind. A crowded theme is being liquidated. Money rotates from one bucket to another and the broad index is unaffected. Signature: concentrated losses, index volatility unchanged.
• Broad de-risking. Investors are cutting total exposure and everything is sold at once. Signature: rising correlation, index implied volatility pushed higher, safe havens bid.

The correct response to each is the opposite of the other. The first should prompt "has the thesis for this bucket changed?" The second should prompt "is my total exposure too high?" So the first step is always to classify, before discussing direction.

Test 1: Is the Gradient Monotonic, and Does the Broad Index Follow?

US session 2026-08-19, intraday (as of 22:32 Beijing time, roughly one hour into the session — not settled, not to be cited as a closing value):

"Intraday range position" = (last − session low) / (session high − session low). This column is more informative than the percentage change:

Every semiconductor name sits in the bottom third of its intraday range (11%-35%), while SPY sits at 82%. In plain terms, semis gapped up, peaked at the open, and were sold all session, while the broad index hugged its highs. Two completely different pools of money doing two completely different things, inside the same trading session.

The prior session (2026-08-18, settled) had an even cleaner gradient: SPY −0.68% < QQQ −1.69% < the semi cluster −2.34% to −4.27% < memory names MU −7.02% / SNDK −9.01%. Strictly monotonic, no bucket out of order. On the same day, Hong Kong internet names did not follow at all — Alibaba (09988) rose 3.68%.

Monotonic gradient plus a non-participating broad index is the first fingerprint of a single-cluster unwind. Broad de-risking looks the opposite: losses flatten out, because what is being sold is not a theme but risk itself.

Test 2: High Single-Name Vol With Low Index Vol

This is the most overlooked and most discriminating test. Options data as of the 08-18 close:

Single-name implied vol at 70-100, index implied vol at 11-19. That combination is itself the answer.

Index volatility is a function of constituent volatility and the correlation between constituents. If every constituent moved violently in the same direction (correlation approaching 1), index volatility would necessarily be dragged up. Single names are moving violently while the index barely moves, which admits only one explanation: those violent moves are cancelling each other out at the index level. Someone is selling memory while someone else buys something else. Money is rotating within the sector, not leaving it.

This is also the best early-warning indicator available. The moment index implied correlation turns up — visible as index implied vol beginning to rise alongside single-name implied vol — the character switches from single-cluster unwind to broad de-risking, and that signal typically leads price.

Test 3: Index Leg or Single-Name Leg?

Put/call ratios as of the 08-18 close:

The index leg is at 3.13 while every single-name leg is below 1. The people buying protection are buying ETF puts, while positioning in individual names still skews bullish. That means the hedgers and the dip-buyers are not the same people — the classic signature of an early-stage momentum unwind rather than a late-stage one. In the late stage, stress transmits to single names and their put/call ratios rise above 1 together.

A Better Metric: Not How Much It Fell, But How Many Times Its Own Normal

A 9% single-day drop sounds catastrophic, but whether −9% is abnormal for SNDK depends entirely on how much SNDK normally moves in a day. Dividing each 08-18 decline by the daily move implied by its own 20-day realized volatility gives a multiple of "its own normal":

All nine fall between 0.79x and 1.60x. Not a single name escaped its own normal range. That −9.01% in SNDK converts to 0.93x — smaller than its own typical daily move.

The conclusion is counterintuitive but important: nothing abnormal happened to any individual name in that session. The entire chain took one step, each scaled to its own volatility, and for memory names one step simply is 9%. When this is the case, hunting for a news catalyst will usually surface a causal story invented after the fact.

Who Is Actually Selling? Korea Publishes the Answer

US markets only show price, but the Korean exchange publishes net buying and selling by investor type, which answers "who is selling" directly. Korea, 2026-08-19, settled close:

Three verifiable facts:

1. Foreign investors reversed. Both names saw foreign net buying in the prior session and heavy net selling in this one, and foreign ownership percentages genuinely declined. The earlier reading — "foreigners are buying, the selling is domestic institutions" — is now formally retired.

2. But this is a decline on falling volume. On the day SK Hynix fell 9.75%, volume was 18% lower than the prior up session; Samsung was 7% lower. Falling harder on less turnover is not a stampede for the exit — it is the bids stepping away. Panic liquidation comes with expanding volume; a thinning bid comes with contracting volume. The two resolve very differently.

3. Zero-sum back-out identifies the buyer. Foreign + institutional + other = 0. Backing out the residual (predominantly retail) gives net buying of 1,510,628 shares in SK Hynix — 35.9% of the day's volume — and 7,100,814 shares in Samsung, 31.3%. Foreign investors and domestic institutions left together; individuals absorbed the supply.

Taken together: leveraged money and foreign money are retreating, but by declining to bid rather than by rushing the exit.

The Options Market Flagged AVGO Three Sessions Early

One observation worth recording. Subtracting 20-day realized volatility from implied volatility gives the volatility risk premium (VRP). As of the 08-18 close:

Almost everything is deeply negative, meaning options are priced cheaper than the underlying actually moves — the market is selling insurance. AVGO was the only positive reading, and it was positive for the third consecutive session, meaning someone was willing to pay more for protection on Broadcom than its own realized movement justified.

In the following session (08-19 intraday), AVGO is the worst performer on the board at −5.39%, sitting at the 11th percentile of its intraday range.

To be explicit: this is a single observation, n=1. It is not statistical evidence and it is not a repeatable strategy. But it does illustrate a principle — when one name's option pricing persistently diverges from its own realized volatility, that divergence deserves a separate look, because whoever is paying for it usually knows something.

What Would Falsify This Read

Writing down falsification conditions in advance is worth more than explaining moves after the fact. If any of the following occurs, the "single-cluster unwind" call should be revised:

1. SMH put/call falls from 3.13 and single-name put/call ratios rise above 1. Both conditions must hold simultaneously; one alone does not count. This would mark hedging demand migrating from the index leg to single names — stress beginning to transmit.
2. Index implied volatility starts rising alongside single-name implied vol (i.e. implied correlation turns up). This is the earliest available signal of the switch from internal rotation to wholesale exit.
3. Korean volume switches from contracting to expanding on down days. Currently contracting (thinning bid); expansion would upgrade it to an accelerating offer.
4. The decline gradient flattens — the broad index starts falling in line with semiconductors, instead of rising as it is now.

None of the four has triggered.

Sell-Side Consensus Did Not Move At All

Worth noting separately: after a session like 08-18, sell-side analysts did not adjust their models (figures benchmarked to the 08-18 close):

On mean targets that implies +66.0% upside for MU and +32.5% for SNDK. Note carefully: that expansion in implied upside is entirely a function of price falling, not of analysts raising targets. Rating distributions and price targets did not move at all. Reading "upside has expanded" as bullish is the single most common misinterpretation of this data.

Data Conventions and Known Traps

In cross-market work, convention errors are more fatal than model errors. The rules observed here:

• Three markets, three different sessions — no side-by-side ranking. Korea 08-19 settled, US 08-19 intraday, US 08-18 settled are three independent blocks. Ranking is permitted within a block, never across blocks, and no cross-block ratios.
• Intraday data is always labelled unsettled. The US 08-19 section is as of 22:32 Beijing time, about one hour into the session, and is not cited as a final value.
• Previous-close values must be cross-checked against settled data. Market data APIs can return a previous close that lags by one session, which computes the change across two sessions and can invert the sign. All ten previous closes here were verified individually against 08-18 settled values and matched exactly.
• Check the sample size behind implied volatility percentiles. The sample used here spans only 73-102 trading days — less than a year — so phrasing such as "lowest of the year" or "one-year low" is not used.
• For custodial holding data (CCASS-type), strip newly-listed participants before computing net change. A participant renaming is flagged as a "new" seat and its entire holding is booked as an increase; citing the net-change field directly can produce a conclusion with the wrong sign.

Frequently Asked Questions

Q: If semis are falling this hard, why is the S&P still up?
A: Because what is being sold is the semiconductor theme, not risk itself. Money rotated into other buckets rather than leaving the market. The evidence is that index implied volatility was not pushed up (SPY IV at 11.78). Under genuine de-risking, index volatility moves first.

Q: Are foreign investors fleeing?
A: In Korea, in the 2026-08-19 session, yes — both heavyweight names saw heavy foreign net selling and foreign ownership genuinely declined. But note this happened on contracting volume, which is a withdrawn bid rather than a stampede. And "foreigners are fleeing" is a claim that requires decomposing by investor type and passing a zero-sum check; it cannot be inferred from price alone.

Q: Is a 9% single-day drop abnormal?
A: Only relative to that stock's own normal daily move. SNDK's −9.01% converts to 0.93x its own typical daily move — smaller than usual. Judging abnormality by absolute percentage is the most common source of misdiagnosis.

Q: How would I know when the character changes?
A: Watch index implied correlation. High single-name vol with low index vol means internal rotation; once index implied vol starts rising too, that is the first signal of broad de-risking, and it usually leads price.

Q: Is this the time to buy the dip?
A: This analysis classifies the character of the move only. It makes no directional call and is not investment advice. The purpose of classification is to determine which question applies — a single-cluster unwind warrants "has the long-term thesis for this bucket changed?", while broad de-risking warrants "is my total exposure too high?" The answers do not transfer between the two.

Sources: US market data (Tiger), Korea Exchange investor-type flow data, options chain snapshots, sell-side consensus database. US 08-19 figures are intraday and unsettled, as of 22:32 Beijing time on 2026-08-19; Korea 08-19 and US/Hong Kong 08-18 are settled closing values. This is public-layer market structure analysis containing no position or transaction information, and is not investment advice.