US Market Recap | 2026-09-03 (ET), Thursday

Coverage window: 2026-09-03 09:30–16:00 ET regular session, including post-close earnings moves after 16:00. Time of writing: 2026-09-03 17:00–17:10 ET (one hour after the close). Price sources: open, high, low, close, volume and after-hours prices for single stocks/ETFs are taken from the stockanalysis.com quote endpoint (close basis 16:00 ET); VIX, the dollar index, Treasury yields, crude, gold and bitcoin are taken from CNBC real-time quotes; index levels and market breadth are taken from CNBC/Reuters same-day close pieces; earnings figures follow company press releases and SEC filings.

Two reading rules for this piece:

1. Every percentage change is independently reverse-computed and cross-checked ((close − previous close) / previous close compared one by one against the chg% returned by the endpoint; all 54 names agree).
2. The reconciliation table must carry a "vs pre-market price" column. Reconciling on "day %" alone systematically overstates the hit rate — 5 names were up sharply on the day yet were losses if bought at the pre-market price (SNOW and TLYS are the clearest cases).

---

0. Today in one paragraph

1. Today was a clean risk-on, and what triggered it was not data but one person. The 10:00 ET ISM services print ran hot on both the headline and the prices component (PMI 55.4, new orders 60.9, the fastest since early 2023; prices paid 72.6, above the prior 70.3). Under the transmission chain in the pre-market list, that should have pushed rate-hike pricing higher and knocked down software and gold. The actual outcome was the exact opposite: Fed Governor Waller said in a Reuters NEXT interview that if data over the next two weeks continues to show disinflation, he will support holding at 3.50–3.75% at the September 15–16 FOMC; the September hike probability fell about 12 percentage points on the day to roughly 54.6% (CNBC¹). Yields fell across the curve; the curve bull-steepened.
2. The strongest theme was software — and it is the one the pre-market call got right. IGV +3.41% vs SMH +0.39%, a spread of +3.02pp — the pre-market list set "the sign of IGV−SMH at the close" as the direct test of "whether the software de-rating has finished clearing", and that test came back affirmative today. Closing breadth among qualified samples was software 13/13 all higher (median +4.29%), wider than the 8/8 seen pre-market.
3. The pre-market list hit 13/21 (61.9%), but the three biggest misses are all the same kind of miss. NTAP (called "avoid", actual +2.55%), HPE (called "avoid", actual +5.04%), DLTH (called "avoid", actual +23.20%) — all three are "fundamental analysis correct, price conclusion inverted". Their shared pattern: gap down on (genuine) bad news, then get bought back in full during a discount-rate re-pricing tape. NTAP opened −10.4% and closed +2.55% (open→close +14.47%); HPE opened −8.2% and closed +5.04% (open→close +14.37%).
4. The most valuable judgment was "do not chase", not "buy". SNOW was up +16.55% on the day, which looks like a big win, but the $377.25 open was the second-highest print of the day, the close was $356.47, open→close −5.51%, and −5.69% versus the pre-market price — the pre-market note "watch only, do not chase the opening price, yesterday's GTLB script is less than 48 hours old" played out exactly. TLYS was more extreme: up +15.49% on the day, open→close −20.72%.
5. After the close came another "beat manufactured by a tariff refund", this time LULU. Adjusted EPS $2.92 vs consensus $1.82 looks like a huge beat, but $0.86 of it comes from a tariff refund and interest; backing that out leaves $2.06; revenue of $2.40 billion was −4% and short of consensus, comparables −9%, and the midpoint of full-year revenue guidance at $10.425 billion is far below the $11.03 billion consensus. After hours −17.88%. This is the fifth sample of the pre-market list's "retail tariff refund" theme.
6. Closing state: S&P 7,747.71 (+1.06%), Nasdaq 26,584.06 (+1.4%), Dow 53,686.11 (+624.16 points, +1.18%), Russell IWM +0.40%; VIX 14.32 (−5.79%); 10Y 4.768% (−2.6bp), 2Y 4.34% (−4.6bp); DXY 98.993 (−0.61%); gold futures $4,520.80 (+2.41%); BTC $81,473 (+5.52%).

---

1. Market overview

Market breadth: NYSE advancers/decliners about 1.78:1, Nasdaq about 1.85:1. Breadth is healthily positive but not extreme — combined with IWM up only +0.40%, today was "large caps and specific themes leading", not an indiscriminate broad rally.

Sentiment read: risk-on, and specifically a "rate-driven" risk-on. Four pieces of evidence are mutually consistent: ① yields fell across the curve with the front end falling more than the long end (2Y −4.6bp vs 30Y −1.6bp, bull steepening), the classic shape of "policy rate expectations being revised down" rather than a safe-haven bid; ② VIX broke below 15; ③ the dollar weakened 0.61%; ④ bitcoin +5.52% and gold +2.41% rose together.

⚠️ The single most important structural observation today: the data were hawkish, the prices were dovish. ISM services 55.4 beat, new orders 60.9 were the fastest since early 2023, and prices paid at 72.6 is still rising — not one of those three numbers supports a cut. Yet yields fell. That means what was priced today was not "the economy and inflation data" but "who votes in September and what they think". This distinction determines how tomorrow's payrolls should be read (see §6).

---

2. Reconciliation of the pre-market list

Reconciliation basis: "vs pre-market price" = close ÷ pre-market price (restored from the 08:47 ET read) − 1, representing the real result of executing the pre-market list at the pre-market price. "Open→close" represents the result of executing at the opening price.

2A Long/positive side

2B Short/avoid side

2C Hit rate and review

13 correct / 5 wrong / 3 right direction but clearly underrated = hit rate 13/21 ≈ 61.9%.

• Correct (13): SNOW, GDX/GLD, NOW, XOM/CVX, TLYS, NTSK, RARE, VSXY, AVGO, semiconductor sector, ALAB/MRVL, MEI, ETD
• Wrong (5): AVAV (top pick with no alpha), DLTH, NTAP, HPE, DELL
• Underrated (3): TSLA, HOOD, CRWD

⚠️ Review: three of the five errors (DLTH / NTAP / HPE) are the same error, and it is not "the analysis was wrong" but "the analysis was right and the conclusion pointed the wrong way".

Looking back, the fundamental breakdowns of all three still hold in full — $0.44 of DLTH's $0.50 really was a tariff refund, NTAP's operating cash flow really was −25%, HPE's Cloud & AI margin really was guided down about 400bp. Not one line needs to be retracted. Yet the three closed +23.20% / +2.55% / +5.04%, two of them a complete reversal from a deep gap down.

The mechanism is this: all three share the pattern of "a name that gapped down on genuine bad news and had already been beaten down". The day's tape was driven by discount-rate re-pricing (Waller → hike probability −12pp → yields down across the curve), and in that kind of tape a gapped-down high-beta name is the most elastic long, not a short. The list ranked stocks by single-name fundamentals, while the real ranking variable that day was sensitivity to rates.

This directly contradicts a sentence the pre-market list wrote itself: in §7 it repeatedly stressed that "in a rate-hike pricing environment, rebounds in duration assets are fragile", and docked points from NOW/CRWD/SNOW on that basis. Today duration assets were not fragile at all — they were the strongest thing on the board, because the premise of "rate-hike pricing" was withdrawn by Waller around 13:00. A macro premise embedded in seven places in the argument was invalidated intraday by a single speech.

One more thing must be logged separately: AVAV, ranked first among the "5 names most worth watching today", rose only 1.25%, roughly in line with the index. And its failure pattern ("spike then fade back below $147") is a falsifiable condition the pre-market list wrote itself. The verification point was set entirely correctly, but it was never used to lower the name's ranking weight — the test was written right, yet it was not turned around to constrain the conclusion.

---

3. Theme verification

Dedicated note 1: the mechanism behind theme ② was confirmed — by a stock that was not on the list

Pre-market, AVGO's bad news was characterized as "gross margin peaking, not demand peaking", with the explicit line that "reading it as AI demand peaking is a directional error". Today:

• AVGO itself: intraday low −6.78% (briefly tripping the "−5% spillover line" set pre-market), but the close recovered to −2.74% while SMH actually rose +0.39% — the spillover did not happen.
• What really confirmed the mechanism was CIEN — a stock the pre-market list did not cover at all. CIEN reported FY26Q3 pre-market on 09-03: EPS $2.11 vs consensus $1.72, revenue $1.67 billion vs consensus $1.63 billion, revenue +37% year over year, non-GAAP EPS more than tripled year over year — a double beat with very fast growth. The result was a −10.36% close (intraday low −12.0%), because gross margin slipped about 40bp and FY27 gross margin guidance is 45–46% (Motley Fool³).

This is the second independent sample of "revenue surging + gross margin falling = de-rating", and it is far cleaner than AVGO (AVGO also carries the confounder of stagnant non-AI businesses, whereas CIEN is purely "volume rising, rate falling"). The pre-market mechanism call therefore gets stronger evidentiary support from CIEN than from AVGO itself.

But it must equally be logged as a coverage failure: CIEN reported pre-market, inside the statistical window the pre-market list declared for itself (09-02 16:00 → 09-03 09:00 ET), and it was not on the list. For a theme it got right, the strongest evidence appeared outside its own sample.

Dedicated note 2: theme ④ "the stagflation trade" — the conclusion was right, but the falsifiable condition was tripped and did nothing

Pre-market, a single, explicit falsifiable condition was set for the gold leg:

"If today's 10:00 ET ISM prices-paid component (July 70.3) falls back noticeably, rate-hike pricing loosens and this leg turns weak on the spot."

What actually happened is that the opposite of the condition was satisfied — prices paid did not fall back at all, it rose to 72.6 — while the headline PMI of 55.4 beat and new orders of 60.9 were the fastest since early 2023 (ISM press release⁴, FXStreet⁵). Under the pre-market transmission chain, gold should have been stronger and software should have collapsed.

Actual result: gold +2.41% (right), software +3.41% (wildly wrong) — and both were pushed up by the same cause: falling yields.

Lesson: the verification point was bound to "which data print", but not to "who is doing the pricing".

The pre-market chain was "ISM prices → hike probability → 10Y → asset prices". The first link held today (the data ran hot); the second link broke — because Waller's remarks on the same day pushed the hike probability straight from about 66% down to about 54.6%, and the marginal impact of the data was overwritten by the policy signal. The result: the data forecast was right and every direction was wrong; and the conclusion "gold will be strong" was right by luck, via a completely different mechanism (yields down → real rates down) rather than the one written pre-market (inflation expectations up).

This is more dangerous than a plain miss: a conclusion that happened to be right via the wrong mechanism gets retained as evidence that the framework works, and gets used again next time. This piece records it explicitly: today's rise in gold does not constitute verification of the "stagflation trade" framework.

Theme verification summary: surprise themes missed pre-market

1. Prediction markets / fintech (the biggest omission): a Deutsche Bank note argued that financial KPI contracts have the strongest growth potential within prediction markets; HOOD +16.57% (51.26 million shares). Pre-market it was classed as "crypto beta, no primary news" — the classification itself was wrong.
2. Crypto complex: BTC +5.52% to $81,473, far stronger than the +1.01% read pre-market. Part of HOOD's gain comes from here.
3. A broad rebound in rate-sensitive assets: XLRE +1.19%, XLU +0.84%, XLF +1.56%, KRE +0.85%. The pre-market list laid out none of this, and it happens to be the most direct beneficiary of the day's macro shift.

---

4. Post-close earnings moves (next-day catalysts)

⚠️ Note on read timing: the after-hours changes below were read at 17:01–17:03 ET. The LULU / GWRE / DOCU conference calls all began at 17:00 ET, so these reads are essentially "after the release, at the very start of the call" prices, and do not yet include call content. The pre-market list has previously recorded the AVGO precedent of "−5~6% after the 8-K, recovered to flat after the call"; reads like these can still move substantially before the next open.

⚠️ The most important structural signal after the close tonight: the software sector bifurcated after hours. GWRE −14.72% vs DOCU +5.28%, while in the regular session they were +5.24% and +0.89% respectively. The split is not about the current quarter (both beat) but about next-quarter guidance: GWRE's Q1 FY27 revenue guidance is about 3% below consensus, while DOCU raised all three guidance items. This stands in direct contrast to the indiscriminate 13/13 software rally during the session — the intraday buying was of "the discount rate", the after-hours selling is of "guidance". That implies the breadth character of today's software rebound may be broken up the next day (see §6①).

Also note: GWRE's pattern is exactly the same type as VSXY, which the list got right today — "current quarter beats + full-year guidance is fine + next-quarter guidance clearly below consensus → crash". The reverse-computation method the pre-market list applied to VSXY applies again to GWRE tonight.

---

5. Flows and sentiment

Sector rotation (ETF closes)

Rotation read: the classic "beneficiaries of falling rates" outperformed across the board. Software (duration assets), gold miners (falling real rates), financials/real estate (curve and valuation) and consumer discretionary led; energy, materials and metals & mining lagged — precisely the three "inflation / real asset" directions.

⚠️ A contrast that is easy to write backwards: gold up, energy down, metals & mining down. If today were an "inflation trade", GDX, XLE and XME should move together. In fact GDX was +3.95% while XME was −0.90% and XLE −0.74% — they diverged. This proves gold was not pushed by inflation expectations today, it was pushed by "falling real rates" — consistent with the conclusion of dedicated note 2 in §3, and it directly answers the question left over from yesterday's recap, "are steel and precious metals moving together again": no, and the gap widened.

The three questions yesterday's recap left for today (unanswered pre-market for lack of trading; now answerable)

VIX and Treasuries

• VIX 14.32 (−5.79%), previous close 15.20, broke below 15. Falling even at the tail end of earnings season and on the eve of payrolls indicates the market is pricing the risk of tomorrow's data on the low side.
• Bull steepening: 2Y −4.6bp, 10Y −2.6bp, 30Y −1.6bp. The front end fell most, the standard shape of a move driven by policy expectations (not growth expectations), matching the nature of Waller's remarks.
• Dollar −0.61%, gold +2.41%, bitcoin +5.52% all moved the same way, jointly pointing to "falling real rates" rather than "safe haven".

Overall read: risk-on, driven by policy expectations rather than by fundamental data. That distinction is the source of the fragility — what supports today's gains is one governor's conditional statement, and the condition (disinflation) gets tested by payrolls tomorrow and by CPI next Thursday.

---

6. Next-day outlook (2026-09-04, Friday)

① Theme continuity

② Tomorrow's earnings and macro calendar

⚠️ How tomorrow's payrolls should be read changed fundamentally today; this is the single most important input this piece provides for the next day.

Yesterday's and this morning's framework was "the reflex of weak data = good for stocks has been reversed" (because the Fed was discussing a hike). After Waller's remarks that reflex is partly restored, but restored asymmetrically:

• Data on the weak side → reinforces Waller's "disinflation holds → stay on hold" path → positive (yields keep falling, today's tape continues).
• Data on the strong side (especially average hourly earnings) → hike probability rebuilds above 66% → software, gold miners and real estate, today's biggest gainers, are all the most fragile, because what they rallied on is precisely the "no hike" expectation.
• The key asymmetry: today's gains have already priced in part of "no hike", but Waller's statement is conditional and the real verdict comes with the 09-11 CPI. Therefore the downside risk from a strong payrolls print exceeds the upside room from a weak one.

③ Focus list (ticker + falsifiable verification point)

④ What to avoid

1. Avoid treating today's risk-on as an established trend. The driver is one governor's conditional statement, and the condition (disinflation) is not tested until the 09-11 CPI. The three strongest directions today (software, gold miners, real estate) are essentially the same trade — a bet on no hike — and provide no diversification against each other.
2. Avoid treating "duration assets" and "gold" as two mutually hedging legs ahead of tomorrow's payrolls. Today already proved they move together (both driven by falling real rates).
3. Avoid chasing today's intraday reversal names (NTAP +14.47%, HPE +14.37% open→close). These two ran the entire distance from −11% to +5%, the odds on chasing them the next day are terrible, and not one of the fundamental negatives (cash flow divergence, the Cloud & AI margin cut) has gone away.
4. Avoid buying any post-close earnings name on a "current-quarter beat" without doing two restatements first: ① restate the tariff refund (LULU is tonight's fifth case); ② next-quarter guidance vs consensus (GWRE is tonight's cautionary case).
5. Avoid energy and tankers. Brent gave back the entire day's premium even with the geopolitical conflict not de-escalated, and XLE was one of the worst sectors; tankers, on real intraday volume, are confirmed to have no independent driver.
6. Avoid extrapolating the AVGO/CIEN gross margin problem into a direction for the semiconductor sector. SMH +0.39%, NVDA +1.78% and LRCX +1.51% all show the spillover did not happen; this theme can only be used for single-name screening.

⑤ Handoff notes for the next pre-market list

1. The post-close prices of GWRE, LULU and DOCU must be confirmed after their calls before writing the post-close earnings section. This piece's reads are at 17:01–17:03 ET (the very start of the calls); the next pre-market piece must re-read them once and note the difference between the two moments.
2. The macro framework must be rewritten wholesale. The premise "the Fed is discussing a hike, September hike probability 60–66%", used in seven places pre-market, has expired; the current value is about 54.6%, and the decisive variable has moved from ISM/payrolls to the 09-11 CPI.
3. Add the two lines missed today: ① prediction markets/fintech (HOOD's Deutsche Bank note — is there follow-through); ② rate-sensitive sectors (XLRE/KRE/XLF).
4. The CIEN coverage failure must be fixed: it reported pre-market, fell inside the statistical window, and was still not picked up by the scan. The next pre-market scan must explicitly enumerate a list of "companies reporting pre-market that day", not just scan the prior session's post-close releases.
5. Add a rate-sensitivity check to every "avoid / short" conclusion: the three biggest errors today (DLTH/NTAP/HPE) were all "gapped-down high-beta names bought back in a discount-rate tape". Before issuing an "avoid" conclusion, first ask: "if today turns out to be a falling-rate day, how does this name trade?"
6. Both payroll scenarios should be pre-written (see the asymmetry in §6②) rather than attributing after the fact once the data lands.

---

⚠️ Risk disclaimer: this recap is a post-close review of information and observations only and does not constitute investment advice. Data may differ in timeliness or definition; company disclosures / SEC filings prevail, and this should not be used directly as a basis for trading.

Internal notes (not sent to clients)

Data channels

The stockanalysis quote endpoint was fully available today, and the after-hours field behavior matched yesterday's record. All 54 names returned complete OHLC (o/h/l/p/cl/v), and reverse-computing cp one by one via (p−cl)/cl matched in 54/54 cases (tolerance 0.06pp), without exception. epv (after-hours volume) was null today, confirming again that the memory entry "epv is available pre-market, null after hours" is correct — this piece therefore did not apply a notional threshold to after-hours moves, and instead only annotates the read time.

⚠️ The CNBC Treasury change_pct field broke on its own for the third time today. US2Y: last=4.34%, prev=4.386%, change=−0.046 (all pointing down), while change_pct showed +0.0898% (the opposite sign). Taking it at face value would have turned today's "bull steepening with the front end falling more than the long end" into "bear flattening", and thereby inverted the entire macro characterization — which is the load-bearing point of this whole piece. Every yield in this piece was reverse-computed independently via last − previous_day_closing. The memory entry "CNBC Treasury change_pct breaks in a single field" has now scored a direct hit three times; recommend upgrading it to "never trust that field".

yfinance was not used at any point this time (skipped directly per the existing memory entry, with no pointless retries).

⭐ The most important search-side trap this time: a 2021 article almost became tonight's post-close conclusion

Searching "Guidewire GWRE Q4 earnings September 3 2026 ... after hours", TipRanks returned guidewire-delivers-upbeat-q4-results-shares-gain-7-after-hours, with a summary saying "shares gain nearly 7% after hours", strong Q4 results, and a record 17 core cloud deals in the quarter.

But measured locally: GWRE after-hours price $172.50 vs close $202.86 = −14.72%, and stockanalysis's after-hours movers list also had GWRE as the 4th-largest decliner (−14.97%). Two independent reads both point to the −15% magnitude, the exact opposite sign from the search summary.

WebFetch pulled the original article to check the publication date: 2021-09-03T12:23:28.000Z — September 3, 2021, five full years earlier.

The four-way isomorphism of this trap is extremely rare: the same company, the same calendar day (September 3), the same fiscal-quarter slot (FY Q4), the same phrase "record 17 core cloud deals" — the only difference is the year. And it contaminated more than once: an earlier search for "after hours earnings movers" returned SNOW +20% / WOOF +9% / HPE −4%, all of which were the 09-02 post-close moves (those companies had reported the previous evening), and the same search also carried a line saying "the indices were basically flat all day, in a narrow range" — when in fact the Dow rose 624 points.

Two rules to lock in:

1. The memory entry "recurring column headlines hide the publication date" needs to be expanded: the previous carriers were fixed media columns and SEC filings; this time it is the "fixed earnings-season slot" — a company reports the same fiscal quarter at the same point each year, and the headline, the wording, even the phrasing of operating highlights overlap heavily. The test remains the only one there is: read the publication date, not the headline.
2. For questions as time-sensitive as "after-hours moves", price and volume come first, search second. Today it was measured prices that falsified the search summary, not the other way around — and this is the second occurrence within the same day, after the 09-03 pre-market use of measured pre-market prices to falsify the VSXY "shares surge" summary. Recommend writing "for after-hours/pre-market moves, always take a primary quote first and use search only to fill in the why" into the hard rules.

Other search-side problems caught this time

1. "Investors weighed upbeat employment data; indices traded in a narrow range" (stockmarketwatch) — in complete conflict with what was measured (SPY +1.05%, Dow +624 points), and no employment data was released that day at all (jobless claims were yesterday's print, payrolls are tomorrow). Template-style generated content; discarded in its entirety.
2. Multiple versions of the payrolls consensus exist: Dow Jones +53,000, market consensus around +58,000, some surveys +50,000 to +55,000; unemployment rate at 4.1% and 4.2% both circulate. This piece writes them side by side as "+53,000 (Dow Jones), range +50,000 to +58,000", rather than dressing a single value up as a settled consensus.
3. Four versions of CIEN's decline are circulating: −9.05% (pre-market), −10.36% (close), −11.2% and −14%. This piece uses the self-measured close basis of −10.36% (previous close 354.16 → close 317.46, reverse computation agrees); the other versions are different moments or different sources, and were not mixed together.

⭐ Methodology: what today exposed is "conclusions not being constrained by one's own verification points"

While reconciling, this piece found a recurring pattern worth logging on its own:

• AVAV: pre-market it wrote itself that the failure pattern would be "spike then fade back below $147". That is exactly what happened (open $151.95 = the high of the day → drop to $140.62). Yet it still ranked first among the "5 names most worth watching today".
• GDX: pre-market it wrote itself the falsifiable condition "if ISM prices paid falls back, this turns weak". In fact prices paid rose to 72.6 (the opposite), and GDX rose 3.95% — the opposite of the condition held and so did the conclusion, which means the condition had no discriminating power from the start.
• DELL: pre-market it wrote itself "yesterday we wrongly judged it as do-not-chase and missed +12.82%", and then still issued "watch only", missing another +4.91% today.

What they share: the verification point / lesson was written down correctly, but was never turned around to modify the conclusion or the ranking weight. This belongs to the same family as the memory entry "verification points outrank headline conclusions" but points the other way: that one says "the verification point is more reliable than the conclusion", this one says "writing a verification point and then not using it to constrain the conclusion is the same as not writing it". Three independent samples appeared in a single piece; recommend locking this in as its own memory entry.

Another one: the three biggest errors are the same error

DLTH / NTAP / HPE are all "fundamental breakdown correct, price conclusion inverted", and all three gapped down on genuine bad news and were then bought back in full (NTAP open→close +14.47%, HPE +14.37%). The list ranked stocks by single-name fundamentals, while the real ranking variable that day was rate sensitivity / beta.

Same type as the memory entry "direction and holding value are two orthogonal conclusions", but this time there is an actionable patch: before issuing an "avoid / short" conclusion, run a scenario check — "if today is a falling-rate day, how does this gapped-down name trade?" Gapped down + high beta + already beaten down is the most elastic long on a risk-on day. This check is extremely cheap and would have avoided at least two of today's three errors.