US Market Recap | 2026-09-21 (ET), Monday

Coverage window: 2026-09-21 09:30–16:00 ET regular session + 16:00–17:05 ET after-hours Quote timestamps: Index/individual-stock closing prices use the official 16:00 ET close (CNBC Quote, realTime=true); after-hours quotes and commodity futures are uniformly stamped 16:50–17:05 ET; Treasury yields are 17:00 ET readings Reconciliation baseline: the day's pre-market list reports/us/2026-09-21.md, whose pre-market quote timestamps are 08:00–08:15 ET

---

0. One-line recap of the day

1. Strong risk-on, but not for the reason the pre-market gave. The Nasdaq +2.26% to close at 27,122.09, a record closing high; the Philadelphia Semiconductor Index +4.29%, Nasdaq 100 +2.83%, S&P 500 +1.49%, Dow +0.71%, Russell 2000 just +0.52%. AMD closed at 615.52 (+9.95%), hit an intraday all-time high of 616.69, and its market cap rose above USD 1.005 trillion.
2. Strongest theme: AI application-layer catalysts spilling over into server CPUs and semiconductors. The day's real drivers were two — (a) Meta's personal AI agent Muse drew collective Wall Street bullishness (META +11.43%, volume ratio 2.32x; the Meta Connect conference opens 9/23); (b) the 9/24 Thursday Trump–Xi Jinping summit and state dinner, attended by Jensen Huang (黄仁勋), Altman (奥尔特曼), Cook (库克), Bezos (贝索斯), Pichai (皮查伊), Musk (马斯克), Nadella (纳德拉), Qualcomm's Amon, and others. The beneficiary chain was server CPUs and compute: ARM +17.16% (volume ratio 2.76x), INTC +12.14%, AMD +9.95%, QCOM +9.29%.
3. 🔴 The weightiest single judgment in the pre-market list was falsified. The pre-market wrote: "there is almost no stock-specific catalyst today that truly belongs to this day … this is a pure macro-beta day," and on that basis labeled all four of INTC, AMD, MSTR, and CIEN as "watch only." The fact is that the day had two primary, belongs-to-today catalysts, and they happened to land on exactly these four names. See Sections 2 and 3.
4. Direction all right, ranking all wrong. All 10 longs closed higher and all 8 shorts/avoids closed lower — directional hit 18/18; but the 7 labeled "watch only" averaged +6.29% on the day, while the 3 labeled "watch closely / priority deep-dive" averaged +4.03% — the Pass group beat the recommended group by 2.26pp. The short side, by contrast, was a clean win: the 8 averaged −3.55%, with MPC −5.30% and VLO −4.84% ranking No. 1 and No. 2 among S&P 500 decliners on the day.
5. Where the money went: energy was the only disaster zone (median of the S&P energy sector −1.78%, only 5/21 advancers, XLE −2.30%, XOP −2.69%); information technology median +1.94%, 63/74 advancers. WTI settled at 95.37 (−4.92%), losing the USD 100 round-number level for the first time since August. BTC 86,712.69 (+6.81%), IBIT +6.50%.
6. Next-day tone: bullish, but crowding has clearly risen, and the volatility market is front-running. VIX closed at 14.87 (+0.41%) — rising instead of falling on a day the S&P gained +1.49%; more importantly, Nasdaq volatility VXN defied the tape, jumping +5.70% to 20.39, and VIX9D +7.09% to 13.14. The index made a new high while its own volatility rose in tandem — a clear non-confirmation, corresponding to the high-density event window of 9/23 Meta Connect and the 9/24 Trump–Xi meeting + COST earnings.

⚠️ The single most important sentence of this recap: the pre-market defined the risk of the entire list as "betting on just one variable: oil → rates." That variable played out today with extreme thoroughness (WTI −4.92%), but what actually made the list money was not it at all — utilities (the purest duration/bond proxy) had a median of −0.64% with only 4/31 advancers, falling against the tape on a day the 10Y dropped 4.1bp. The attribution "falling rates drive a duration-asset repricing" was refuted on the spot by evidence from within its own sector. See Section 5.

---

1. Market overview

1.1 Indexes and volume

On the precise definition of "new high": the Nasdaq Composite's 27,122.09 is a record closing high; but the day's intraday high was 27,183.93, still below the 6/1 intraday peak of 27,190.21. That is, "record close but not a record intraday" — the two are not the same thing. The S&P 500 is still 0.67% short of its 8/13 level of 7,816.70, and the Dow 4.9% short of its 8/5 level of 54,744.33 — only the Nasdaq family made new highs, which is itself evidence of narrowness.

1.2 Market breadth (calculated in-house, n = 503)

⚠️ Breadth does not use secondary sources: the closing prices of all S&P 500 constituents were pulled and computed one by one (503/503 retrieved, 0 with stale data):

Conclusion: this is an extremely narrow ridge. The index +1.49%, while the constituent median was only +0.30%; equal-weight RSP underperformed cap-weighted SPY by 0.99pp. In other words, more than two-thirds of the index's gain came from a very small number of mega-cap tech stocks. Among the 89 names up more than 2%, information technology and communication services accounted for the vast majority.

This is exactly the criterion set by pre-market verification point ①, and today it gave the reading "no basis for broadening" — yet the market still gained 1.49% on the day, and every long on the list still rose. This contradiction is very important and is examined separately at the end of Section 2.

1.3 Rates, FX, commodities, volatility

⚠️ Treasury moves are computed in-house as last − prev. CNBC's US2Y.change_pct field once again contradicts its own last/prev (change=+0.008 yet change_pct=−0.0156%); the field is not trustworthy. The direction of lower yields was independently cross-confirmed by TLT +0.68%.

🔴 The shape of the yield curve was the exact opposite of a "rate-cut / easing repricing."

The 2Y rose 0.8bp, while the 10Y fell 4.1bp and the 30Y fell 4.0bp — a long-end decline mixing bear-flattening and bull-steepening, with the front end instead propped up. There was a primary witness on the day: Chicago Fed President Goolsbee spoke at an OMFIF event in London, with explicitly hawkish wording —

• "Supply shocks are arriving more frequently, hitting harder, and lasting longer," and they cannot be ignored by monetary policy just because they are "theoretically transitory";
• "If demand overheats, there is no ambiguity about how the Fed should respond";
• He specifically called out AI-related massive capital expenditure that could "spill out of its own lane," pushing aggregate output to a level the economy cannot absorb;
• He also warned about the stickiness of services inflation.

(Goolsbee is not a voting member this year, and did not comment on last week's FOMC outcome or his own policy leanings.)

Pre-market verification point ③ read: "If hawkish wording drives the 2Y yield back up, the duration trade will reverse quickly." The first half played out completely — Goolsbee was indeed hawkish, and the 2Y did indeed rise — but the second half did not happen: the duration trade not only failed to reverse, it was the strongest segment of the day. This shows the causal assumption of that criterion is itself flawed; see the end of Section 2 and Section 5.

Sentiment read: risk appetite is strong, but not "quietly strong." Gold −0.76%, the dollar +0.20%, BTC +6.81%, high beta outperforming across the board — textbook risk-on; yet VIX crept up on a day the index gained +1.49%, and VXN actually jumped 5.70%. The index made a new high while its own implied volatility rose, which says this wave of buying came with "chasing + hedging," not "quiet reallocation."

---

2. Reconciling the pre-market list

2.1 Long side (10 names, pre-market overall list)

⚠️ The "vs. pre-market price" column is the executable measure: the day-move is benchmarked to the 09-18 close, which includes the pre-open gap that you could not buy; "vs. pre-market price" is benchmarked to the quotes at the list's publication time (08:00–08:15 ET) and measures "whether executing on the list through the day actually made money."

Long-side total: all 10 closed higher; on the "vs. pre-market price" basis, 6 positive and 4 negative, averaging +1.39%.

2.2 Short / avoid side (8 names)

Short-side total: all 8 closed lower, averaging −3.55% on the day; on the "vs. pre-market price" basis, average short gain +2.26%. This is the cleanest side of this issue.

⚠️ But one specific piece of reasoning was falsified: the pre-market explicitly wrote "XOM/CVX: do not short, because they have barely risen since 8/20 (XOM −1.6%, CVX +1.8%) and there are no gains on the table to give back." The result: XOM −3.20%, CVX −2.79% — both fell more than OXY (−2.70%) and DVN (−2.10%). In other words, the variable "whether there are gains to give back" had no explanatory power for the ranking of declines that day; the real ranking variable was operating leverage to crude prices (refiners > integrated majors > low-leverage E&P). This reasoning was wrong, although it landed on the conservative direction — "do not short" — and caused no loss.

2.3 🔴 Biggest wrong calls: INTC (+12.14%) and AMD (+9.95%)

The pre-market call on INTC was the heaviest in tone of the entire report: "the largest money inflow, paired with the weakest same-day evidence, paired with a valuation at the 98th percentile of 5 years," with the risk deduction maxed at −15 — the only name in the entire list so treated. The core argument was: "three rounds of targeted searches plus EDGAR found no same-day catalyst; the pre-market volume of 4.59 million shares was buying a 9/16 old news item."

Catalysts that actually existed that day (primary / near-primary, findable after the fact):

🔴 This is a complete recurrence of "couldn't find the news ≠ there was no news." The pre-market had actually written a self-protection clause in the INTC section: "Please read this as 'this report could not find a catalyst,' not as 'there is definitely no catalyst' … the pre-market volume of 4.59 million shares is real, tangible money, very likely corresponding to a piece of news this report failed to capture."

This qualification was correct, but it stayed at the level of wording and never entered the decision layer — the −15 risk deduction was applied anyway, and the label remained "watch only." The right approach: when "abnormally large money + no catalyst found" appear together, the weight of that judgment should be reduced (flagged as "insufficient evidence, exclude from ranking"), rather than translating "I didn't find it" into "it doesn't exist" and heavily penalizing on that basis. The pre-market volume of 4.59 million shares was hard evidence, observable at that very moment, pointing in the opposite direction — and it was overridden by the narrative.

AMD is the same story: the pre-market's stated reason was "the least friendly position: 94.3rd percentile of the 52-week range, high gap-fill risk." The result: volume surged to 1.92x, it gapped +4.30% at the open, gained another 5.42% open-to-close, finished near an all-time high, and its market cap broke through USD 1 trillion for the first time. On a day with a real incremental catalyst, a "high position" is not resistance but a tailwind — this is the same variable seen from the other side of the "low position = large repair room" argument the pre-market used in its own NVDA section, and that day it failed in both directions (see below).

2.4 🔴 NVDA: verification point triggered, top pick downgraded

The verification point the pre-market set for NVDA, its overall No. 1, was falsifiable and stated with precision:

Measured only by the open→close relative return during 09:30–16:00 (pre-market not counted): if NVDA's open→close underperforms AMD by more than 1pp, the "large repair room" thesis fails and it should be downgraded.

Actual result: NVDA open→close +1.99%, AMD open→close +5.42%, an underperformance of 3.43pp — threshold 1pp, and the trigger magnitude was 3.4x the threshold. The verification point was clearly triggered, and the top pick was downgraded per the rule.

And there was a second, independent piece of evidence pointing to the same conclusion: NVDA's volume ratio that day was 0.96x (below its 10-day average volume) — the only long in the entire list that failed to show a volume surge; over the same period AMD 1.92x, ARM 2.76x, INTC 1.65x, META 2.32x. On a day the Philadelphia Semiconductor Index gained +4.29%, NVDA rose 2.30% on below-average volume — no incremental buyers came in.

What does this mean for the pre-market's fulcrum: NVDA's fulcrum was the valuation percentile — "TTM GAAP P/E of 28.1x at the 1.7th percentile of the past 3 years." The day's result: the valuation percentile had no ranking power whatsoever that day. The day's top four gainers — ARM (+17.16%), INTC (+12.14%), AMD (+9.95%), QCOM (+9.29%) — had valuation percentiles of extremely high, 98th percentile of 5 years, 95th percentile of 5 years, and relatively low, respectively — with zero correlation to the gains. Pre-market verification point ④ had actually pre-specified this test ("look at whether today's biggest gainers correlate with their fundamental quality; if there is no correlation at all, stock-level conclusions should all be downweighted"), and the test result was: no correlation at all.

2.5 🔴 VLO: the trade made money, but the thesis was rejected by its own criterion

This is the most memorable item of this issue. The verification point the pre-market set for VLO (short) was:

Watch the crack spread, not WTI: if the 3-2-1 rebounds above $69.28 (last Friday), the inference fails.

Actual (computed in-house from NYMEX futures, timestamp 2026-09-21 16:50 ET, formula (2×RBOB + 1×ULSD)×42 ÷ 3 − WTI):

The crack spread not only failed to keep narrowing, it climbed back to $69.52, back above the $69.28 line. Verification point triggered — by the pre-market's own written rule, the "refiner short" thesis does not hold.

The mechanism makes complete sense, and in the direction opposite to the pre-market's description: the pre-market's 08:12 reading was ULSD (−3.37%) falling faster than crude (−2.86%), so it judged "refining margins are compressing." By the close, this relationship had entirely reversed: WTI −4.92% > ULSD −3.69% > RBOB −2.10% — crude fell the fastest while products fell more slowly, so by definition the crack spread widened.

So: VLO/MPC/PSX falling hard that day had nothing to do with "refining margins being compressed." What they fell on was crude beta and positioning, not margin.

The evidence that actually fits is in the intraday price action, and it is far harder than the crack spread:

VLO and MPC both made new 52-week highs that day, then reversed to close −4.84% / −5.30%, with MPC volume at 1.36x. This is a textbook "new-high reversal day" (bearish outside reversal) — the direct fulfillment of the pre-market's positioning argument ("+18.6%/+21.0% since 8/20, closing at the top of the range, dense profit-taking supply with no one leaving") — not the fulfillment of the fundamentals argument.

The lesson of this item: after finalizing, the pre-market had specially added a whole section of first-hand crack-spread data collection (EIA 20-year series, 99.78th percentile, implied $39 threshold, P/B at the 99.9th percentile of five years), and quality review even listed it as "the most valuable item." Yet what actually worked that day was the plainest positioning argument; the elaborate crack-spread framework not only failed to work, it produced a reading in the opposite direction. The profit masked the failure of the thesis — if verification points are not back-tested one by one, this would be recorded as "a victory of fundamental analysis," thereby reinforcing a wrong method next time.

2.6 Final reconciliation of the 5 watch-list names

2.7 Hit rate and review

Directional hit rate: 18/18 = 100%. —— But that number carries no information and must be voided. On the day, the S&P gained +1.49%, the information technology sector had 63/74 advancers, and the energy sector had only 5/21. On a day of such uniform direction, "all longs up, all shorts down" is mostly beta, not judgment.

Switch to a relative-benchmark basis (the informative version):

🔴 Review No. 1: the Pass group beat the recommended group by 2.26pp that day; the recommended group's relative advantage exists only on the "vs. pre-market price" basis. Both numbers are true at the same time, and the difference is the gap: the 7 passed names generally carried larger overnight gaps, while the recommended group had smaller gaps. That is: the pre-market's screening logic ("big gap = don't chase") works on the "executable return" basis (recommended group +1.83% vs. Pass group +1.20%), but fails on the "was this stock worth holding today" basis (+4.03% vs. +6.29%). These are two orthogonal conclusions; they cannot substitute for each other, and one must not use the winning basis to cover up the losing one.

🔴 Review No. 2: of the three macro verification points, two gave misleading signals, and the only effective one was not the highest priority.

Four criteria, three of which drove you out of correct positions. The common root cause: they were all measuring the "macro environment," while what actually determined P&L that day was "whether there was company-level new information." The pre-market itself wrote "this report's verification points were deliberately designed to test different variables, not six copies of the same variable" — but in fact, breadth, oil, and rates were still three copies of the same layer of variable (macro beta), and not a single criterion in the list tested "whether there is a stock-specific catalyst belonging to today." And that was precisely the only dimension with explanatory power.

🔴 Review No. 3: what was rejected was not just the conclusion but the entire attribution chain. The pre-market's causal chain was "oil falls → inflation expectations and term premium decline → yields fall → duration assets reprice + risk appetite returns," and it repeatedly emphasized that "the 20 names are 20x correlated exposure to a single bet." Today's within-sector data directly rejected the middle of this chain (see Section 5): utilities, as the purest duration proxy, had a median of −0.64% with only 4/31 advancers on a day the 10Y fell 4.1bp. If the discount rate were really the driver, XLU could not be the second-worst sector in the entire market.

---

3. Theme verification today

3.1 Did the pre-market call the strongest theme correctly? — No

The pre-market's "top 3 themes today" were: 1. Crypto-linked equities / 2. Semiconductors / 3. AI power and neocloud.

The actual ranking (by sector ETFs and individual-stock medians): 1. Server CPUs / semiconductors (.SOX +4.29%) / 2. AI application layer (XLC +3.90%) / 3. Optical communication (ALAB +12.36% leading), while crypto equities were actually "decent gains but internal divergence and intraday fading," and AI power was a clear laggard group.

One and a half of the three themes were judged correctly. Specifically:

• Ranking crypto No. 1 was wrong, and wrong in a structural way: the pre-market itself had seen through "MSTR/COIN/HOOD are not three independent pieces of evidence but three copies of the single BTC event," yet still ranked this copy group No. 1. Today's result adds a finer layer — on a day BTC +6.81%, the performance of crypto "equities" depended on how close they are to the coin: pure coin-holding vehicles IBIT +6.50% and MSTR +9.47% kept up; while COIN (open→close −2.02%), HOOD (−1.40%), and CRCL (−3.67%), which monetize trading volume, were systematically sold intraday; miners MARA only +0.30% and RIOT +1.81% barely participated. In the pre-market's intra-theme ranking, COIN and HOOD were placed No. 1 and No. 2, and IBIT No. 3 — exactly the reverse order.
• Ranking AI power No. 3 was wrong: in that theme, the pre-market named "VRT, GEV" as the representative names (and had already excluded IREN). VRT +0.59%, GEV +0.63% — both underperformed the S&P by 0.9pp and were the two weakest names in the entire list. What actually rose were OKLO +5.74% and SMR +6.29%, which within that theme the pre-market had judged "avoid — concept stocks without scaled revenue." Within the same theme, the recommended names lagged and the avoided names led.

3.2 Surprising themes the pre-market missed

⚠️ The WBD/PSKY pair deserves a separate note: on the same M&A positive, the target WBD closed +10.79% (volume ratio 11.64x, the highest in the entire market), while the acquirer PSKY went all the way from +12.5% intraday to a red close of −2.94%. The secondary-source headline at the time was "WBD up 7%, Paramount Skydance up 5%" — that was a pre-market/early-session reading; by the close, PSKY was down. Following the headline would have recorded the direction backwards — this is the same class of problem as this report's handling of CIEN's thin pre-market quote.

---

4. Post-earnings after-hours moves

No large-cap earnings after the bell today (9/21, Monday). Per the schedule, the day's earnings were only Ennis (EBF, pre-market, closed +0.80%) and Abivax (ABVX, after-hours, closed −1.51%), neither with any substantive market impact. The large-cap after-hours zone was essentially frozen: NVDA −0.20%, AMD −0.09%, META −0.03%, INTC −0.02%, ARM −0.25% (timestamp 17:03–17:04 ET).

The only substantive after-hours catalyst: VICR raises guidance

Read:

• This is not an earnings report but an after-hours guidance raise — in nature a company disclosure with relatively high certainty.
• ⚠️ Volume warning: after-hours turnover of 65,231 shares, a notional amount of roughly USD 16.30 million. After-hours quotes are cumulative prints, not a single point in time, and secondary sources gave three different numbers at different moments — +6%, +11.88%, +18.39% — which are not contradictory but three different timestamps. This report gives its own self-collected, timestamped reading. It must be re-pulled before the next open.
• Thematically it lands on "AI power / supply," the weakest line today (VRT +0.59%, GEV +0.63%). If a "royalty-driven gross-margin story" is accepted by the market, it could become the catch-up trigger for that sector the next day — one of the few next-day catalysts that "belong to the company itself."

---

5. Money flows and sentiment

5.1 Sector rotation (in-house medians + ETF, dual basis)

Other tools: SMH +4.02%, ARKK +2.82%, IGV +2.66%, XBI +0.96%, OIH −0.52%, TLT +0.68%, GLD −0.70%.

5.2 🔴 The attribution "falling rates drive a duration repricing" was rejected by within-sector evidence

The pre-market rested the logic of the entire list on one chain: oil falls → inflation expectations and term premium decline → yields fall → duration assets reprice. Today's data broke this chain in the middle:

On a day the 10Y fell 4.1bp, the three purest duration-proxy sectors all fell or barely moved, while high-beta semiconductors gained +4.29%. This is not the shape of a duration repricing; it is the shape of risk appetite concentrating on a single narrative (AI).

A more accurate attribution for the day would be:

1. Primary cause (attributable): Meta Muse + the Trump–Xi meeting reignited the "AI demand and US-China chip access" narrative, and money concentrated in the previously most-lagging segment of the AI compute chain (server CPUs: ARM/INTC/AMD/QCOM) rather than NVDA, whose story was already fully expressed.
2. Secondary cause (real but overestimated): oil −4.92% did push down long-end yields and inflation expectations, but its role mainly showed up as "not dragging the market down" rather than "providing thrust" — otherwise XLU would not have fallen.
3. The energy sector's −2.30% is the direct result of oil — fully consistent with the pre-market's judgment, and the only segment of the day's attribution chain that closes completely.

The pre-market's warning ("the 20 names are just 20x exposure to a single bet; if oil reverses they all pull back together") therefore needs to be corrected: the list's true common exposure was not oil but the AI narrative. This distinction is material for risk management — an oil reversal and an AI-narrative setback are two completely different trigger conditions, and the latter (e.g., a 9/24 meeting outcome below expectations) has a much higher probability this week.

5.3 Volatility: three non-confirmation signals

In a healthy return of risk appetite, rising indexes should come with falling volatility. Today all three volatility gauges rose, and the Nasdaq's volatility gain (+5.70%) was 14x the S&P's (+0.41%).

The pre-market had already noticed this pattern and offered a preliminary read ("more like short covering and duration repricing than a substantive improvement in risk appetite"). The direction was right, but the more accurate explanation for the day is a third one: the options market is pricing a known, high-density event window (9/23 Meta Connect, 9/24 Trump–Xi meeting + COST earnings, quarter-end rebalancing). Rising volatility is not negating today's gains; it is saying "the distribution for the next few days is wide."

5.4 risk-on / risk-off characterization

Characterization: strong risk-on, but with extremely high concentration and rising hedging demand.

• Supporting risk-on: BTC +6.81%, gold −0.76%, ARKK +2.82%, high beta outperforming across the board, the Nasdaq at a record close.
• Limiting its meaning: equal-weight RSP only +0.56%, Russell 2000 only +0.52%, constituent median +0.30%; of the 503 S&P constituents, only 89 rose more than 2% while the index gained 1.49%; defensive sectors (utilities, consumer staples) fell in tandem with energy, showing that the money was not new inflow but more a switch from defensives and energy into AI — a reallocation, not additional buying.
• The dollar's modest +0.20% rise sits in slight tension with risk-on, but the magnitude is at the noise level and does not constitute an independent signal.

---

6. Next-day outlook (Tuesday, 2026-09-22)

① Theme continuity

② Tomorrow's earnings and macro calendar

⚠️ One calendar error that must be corrected (verified against primary sources): the pre-market wrote "Accenture (埃森哲) ACN, Nike (耐克) NKE … are on this week's calendar." Both are wrong:

In fact, only two earnings events this week are worth attention, and both are on 9/24 (Thursday):

➡️ Direct implication for the next day (Tuesday 9/22): this is a "day with no calendar." No data, no earnings; the dense slate of Fed speakers this week (10 appearances across the week) is the only exogenous headline source. On a day without new information, the most common follow-up pattern for a big rally driven by a single narrative like today's is a "digestion day that opens high and drifts lower," not "a second day of continued volume-driven breakout."

③ Key names to watch (ticker + falsifiable verification points)

⚠️ This section deliberately includes criteria that "test company-level rather than macro-level" factors — this is today's biggest methodological lesson (see Review No. 2 in 2.7).

④ Directions to avoid

1. Shrinking-volume rallies in software / SaaS (DDOG 0.82x, ZS 0.59x, MDB 0.75x, OKTA 0.90x). Gains of 5–9% on volume below the 10-day average are the classic signature of short covering, not new-money accumulation. This group lacks follow-through for the next day.
2. The "monetizes trading volume" layer of the crypto chain (COIN, HOOD, CRCL). Today, in the best possible environment of BTC +6.81%, all three were negative open-to-close (−2.02%, −1.40%, −3.67%). The coin-price tailwind can no longer be transmitted to their P&L narratives — a signal more important than the price itself.
3. Chasing ARM (see above) and chasing INTC. INTC has risen +12.14% in a row, accumulating +32.8% since 9/3; the day's catalyst (the AUO packaging partnership) is far too small in magnitude to support a single-day market-cap increase of USD 64 billion — a considerable part of today's gain is theme beta and short covering, not the AUO news itself.
4. Still do not recommend shorting refiners, even though it was the best trade today. The reason is unchanged, and today brought no improvement: the average crack spread so far in Q3 is 43% above Q2, and the Q3 earnings in late October will very likely set new highs, while VLO's earnings date is confirmed as 10/22 pre-market. Today once again proved that the profits in this direction come from positioning, not fundamentals, and the positioning has already released a considerable portion after a single-day −4.84%.
5. Do not take today's "all directions right" as validation of the list's methodology. An 18/18 directional hit rate carries zero information on a day of extremely high within-sector uniformity (see 2.7).

⑤ Input tips for the next day's pre-market list

1. 🔴 Highest priority: make "same-day stock-specific catalyst" an independent, must-answer search item, and allow "not found" as a reason to downweight rather than to deduct points. Today's error: three rounds of searches failed to find INTC's catalyst → judged "no catalyst" → risk deduction maxed at −15 → it ended +12.14%. The correct handling: "abnormal pre-market volume + catalyst not found" should trigger "insufficient evidence, exclude from ranking," not "insufficient evidence, therefore heavy penalty." The pre-market had actually written this qualification sentence, but it never entered the scoring.
2. Tomorrow's pre-market must check these three things before writing any conclusions: (a) the exact agenda of Meta Connect 9/23 and the verifiable details of Muse; (b) the publicly available expectation framework for the 9/24 meeting on chip access / export controls (do not use "according to reports"-grade sources); (c) whether VICR's after-hours gain holds up under pre-market volume.
3. Verification-point design must include at least one criterion that "tests stock-level information." Today, three of the four criteria misled, because they all measured the macro. It is recommended to permanently add one: "Among the day's top 10 gainers, how many can be traced to a same-day primary catalyst?" — this ratio itself is a direct reading of "whether today was a beta day or an alpha day."
4. The volume ratio must enter the ranking and be viewed side by side with the gain. Today's single most predictive variable was neither the valuation percentile nor position, but the volume ratio: ARM 2.76x / META 2.32x / AMD 1.92x all surged and strengthened on the day, NVDA 0.96x was the only shrinking-volume long and was indeed the weakest; the software group's 0.59–0.90x shrinking-volume rallies should be discounted.
5. Intra-theme ranking should be by "distance from the catalyst source," not by "fundamental quality." Today's correct order within the crypto theme was IBIT > MSTR > COIN/HOOD (distance from the coin), while the pre-market ranked by fundamental quality as COIN > HOOD > IBIT — exactly the reverse order.
6. Do not reuse the crack-spread framework to support the refiner direction without first explaining why it gave an opposite reading today. The alternative arguments available are positioning (new-high reversal, cumulative gains since 8/20, P/B percentile).
7. The calendar must be checked item by item against primary sources. This issue's pre-market put both ACN and NKE in "this week"; both were wrong (actually 10/1), and the source was a secondary calendar. Company IR / press releases are the only trustworthy sources.

---

Data collection and verification log for this issue (internal)

Data channel status

• The CNBC quote API was fully available this time and served as the primary channel for all prices in this report. The three batches of requests covered 666 symbols in total — 666/666 succeeded, 0 missing (46 macro/ETF + 117 individual stocks + 503 S&P constituents). The batch endpoint was called at 12 per batch + 4 retries + 0.4s interval; no rate limiting was triggered.
• yfinance was not used (this machine has historically been rate-limited by Yahoo). All data in this report comes from CNBC + primary IR / press releases + WebSearch/WebFetch.
• After-hours quotes come from CNBC's ExtendedMktQuote field (exthrs=1), which is available after the close (consistent with the asymmetry versus the pre-market session); last_time timestamps were recorded entry by entry.

Bad fields / bad data identified and worked around (still to be checked next time)

1. CNBC US2Y.change_pct broke again: change=+0.008 (consistent with last−prev) but change_pct=−0.0156% (its sign contradicts its own data; the correct value is about +0.17%). Fully consistent with past records — this field can be confirmed as chronically unusable; always compute from last/prev. The 10Y/30Y change_pct was correct this time but was still computed on the same basis.
2. BTC='s previous_day_closing rolls intraday: it gave 86,974.52 / −0.2949%, while BTC.CM= gave 81,184.96 / +6.81% and @BTC.1 gave 81,330.00 / +7.03%. BTC= is a rolling 24h basis, not a "previous trading day close" basis; using it for daily reconciliation would write +6.81% as −0.29%. BTC.CM= was adopted. This is the same family of pitfall as "FX previous closes roll intraday."
3. SQ returned $0.18 with last_time stuck at 2026-09-18 (Block has been renamed XYZ), and ANSS returned all-empty fields (acquired by Synopsys). Both were caught and removed by the requested-vs-returned count check and entered no statistics. Batch pulls must validate requested==got and check the last_time prefix name by name; otherwise such dead symbols silently enter the sample.
4. .DXY was not frozen this time (it has open/high/low, last_time 17:00), behaving differently from the pre-market session; it was accepted.

Secondary-source errors intercepted (5 items, 3 of which would change conclusions)

1. 🔴 Breadth data misattributed (would change the characterization of the entire report): searching "NYSE advancers decliners September 21 2026" returned an article titled Stock Market News for Sep 21, 2026 containing "decliners leading 1.78:1, Nasdaq 1.42:1, 92 new highs vs. 346 new lows." All of these are 9/18 (Friday) data — the same article states "VIX fell 4.08% to 14.81" (14.81 being exactly the 9/18 close) and "Friday volume of 25.29 billion shares." Copying it would have written the day's breadth as "decliners dominant," exactly the opposite of the actual 292 advancers / 209 decliners. → Replaced with computing all 503 S&P 500 constituents one by one in-house. This is the second hit on the record "Zacks / fixed-column 'Sep N' pages actually describe day N−1."
2. 🔴 Yahoo's intraday live blog had self-contradictory WTI numbers: it claimed "WTI crude fell 4.23% to $92.02." Self-check: 100.30 × (1−4.23%) = 96.08 ≠ 92.02 — the set of numbers is internally inconsistent. CNBC @CL.1 gives 100.30 − 4.93 = 95.37, self-consistent. → Adopted 95.37 (−4.92%), with direction cross-confirmed by USO −3.68% / BNO −2.81%. The rule "when two change numbers conflict, back-compute the base first" directly took effect this time.
3. 🔴 ACN / NKE earnings dates (correction already made in the body): the pre-market claimed both were "on this week's calendar." Accenture official press release: October 1 (Thursday) pre-market, 8:00 a.m. EDT conference call; NIKE IR 8/28 announcement: October 1 at about 1:15 p.m. PT. Both are 10/1, not this week. → Corrected in a table in §6② of the body.
4. ARM search returned a stale, wrong-currency quote: one site claimed ARM "rose 4.04% on 9/21 to close at CA$275.61." 275.61 is the USD closing price of 9/18 — wrong on both the date and the currency. → All quotes used are self-collected (322.90, +17.16%).
5. The WBD/PSKY headline was a pre-market reading: the secondary headline read "WBD up 7%, Paramount Skydance up 5%." Actual closes: WBD +10.79%, PSKY −2.94% (intraday had been +12.5%). Writing from the headline would have recorded the acquirer's direction backwards. §3.2 of the body states this on a closing basis with the intraday high noted.
6. VICR's after-hours gain had three coexisting readings (+6% / +11.88% / +18.39%), from different moments. → Adopted the self-collected, timestamped reading (17:02 ET, +11.88%, 65,231 shares), with an explicit note in the body that it must be re-pulled the next day.

Numbers computed in-house rather than quoted in this report (reproducible)

• S&P 500 breadth (292/209/2), sector medians and advancer shares: computed one by one from the last and previous_day_closing of all 503 constituents; the GICS sector mapping was taken from datasets/s-and-p-500-companies.
• 3-2-1 crack spread ($69.28 → $69.52): (2×RBOB + 1×ULSD)×42 ÷ 3 − WTI, timestamp 16:50 ET.
• All "vs. pre-market price" columns: close ÷ (09-18 close × (1 + pre-market move)) − 1, with the pre-market move taken from the original text of that day's pre-market list.
• Treasury bp changes: (last − prev) × 100.

Still unresolved, priority next time (4 items)

1. AKAM +12.33% (No. 1 S&P gainer of the day): a primary same-day catalyst could not be located. Searches only reached old ratings such as the 2026-05 BoA upgrade and the late-2026-08 Piper Sandler upgrade. Volume ratio 1.37x, not abnormally heavy. Per this report's standard it was written as "not found," not "does not exist" — but this is exactly the same pattern as today's INTC error. Next time, names "in the top 3 gainers with unclear catalysts" should get a dedicated EDGAR + company-newsroom check, rather than stopping at generic searches.
2. For INTC's AUO Micro LED partnership, no primary announcement from Intel or 友达 (AUO) could be obtained. Sources were secondary accounts such as Blockonomi / Investing.com. This is one of the main bases for this report's attribution on INTC; the evidence grade is relatively low, and the body uses qualified wording. Next time, check AUO's Taiwan Market Observation Post System announcements.
3. Piper Sandler's ARM initiation ($320) and Wells Fargo's META price-target change (640→796): neither the instant-alert original URL nor the exact date could be obtained. Per this issue's pre-market IREN lesson ("any rating change today must have its URL date verified"), strictly speaking these two reach only "near-primary" grade. They were used in the body but not as the sole fulcrum.
4. The first publication time of the "9/24 meeting" news needs to be pinned down: Semafor 9/18 and Seoul Economic Daily 9/20 both carried reports, i.e., it spanned the weekend. This report classifies it as "new information for Monday," but strictly speaking it belongs to "pricing-progress day 1–2" rather than "launch day" — a distinction that affects the continuity judgment for that theme the next day. Next time, "pricing progress" should be a fixed column in the theme table.

Methodological summary of this issue (worth codifying) The costliest lesson of this issue is not that some number was wrong, but this: the pre-market translated "I could not find it in my searches" into "it does not exist," and on that basis applied its heaviest risk deduction (INTC −15), after which the name rose +12.14% that day. And the pre-market had itself written the correct qualification ("please read this as 'this report could not find a catalyst,' not as 'there is definitely no catalyst'") — the qualification sentence entered the body but never entered the scoring model.

Second: three of the four verification points gave misleading signals, because they were all measuring the same thing at the macro level. Criterion diversity cannot be judged only by "different variable names" (breadth / oil / rates); what matters is whether they would all flip at once on the same piece of news. Today they indeed all flipped at once.

Third (the most insidious): the VLO short made 4.84%, but its verification point was clearly triggered and its mechanism was falsified. Looking only at P&L, this would be recorded as "the elaborate crack-spread analysis succeeded" and thus be reinforced next time — what actually worked was the plainest positioning argument. Profits can mask the failure of a method, so verification points must be back-tested one by one, even when the trade was profitable.

---

⚠️ Risk disclaimer: This recap is only a post-market collation of information and observations and does not constitute investment advice. Data may differ in timeliness or caliber; please rely on company disclosures / SEC filings, and do not use this directly as a basis for trading.