US Market Recap | 2026-08-20 (ET) Thursday

Coverage window: 2026-08-20 09:30 ET open → 16:00 ET regular-session close; after-hours moves through 17:03 ET.

Quote conventions:

• Closing prices, intraday OHLC, volume and after-hours prices for indices, single stocks and ETFs are taken from the CNBC quote service, read at 17:00–17:03 ET, with last_time checked line by line to rule out stale data.
• Treasury yields follow the official US Treasury CMT series (daily_treasury_yield_curve) (struck around 15:30 ET), with CNBC's 17:00 read shown separately for comparison — the two have different read timestamps and are not mixed in this report.
• Pre-market prices are back-computed from the pre-market list's disclosed "previous close × (1 + pre-market change)", read at the 08:01–08:28 ET timestamps noted in the pre-market list.
• Earnings figures are taken from the companies' original press releases, links in the body.

---

⚠️ Read this first: today's list is not a case of "a few names were wrong" — the ranking was inverted end to end

Group the 14 bullish-side names in the pre-market list by the conclusion tier assigned at the time, then look at how they actually did today:

Pre-market conclusion tier	n	Mean same-day change	Mean change vs pre-market price
Priority deep-dive (HOOD, FUTU)	2	+1.17%	−3.92%
Watch closely (IBIT, BULL, XOM, COIN, OXY)	5	+3.89%	−1.00%
Watch only (DE, NDSN, MSTR, CRCL, GLXY)	5	+7.27%	+2.52%
Avoid (MARA, BMNR)	2	+11.05%	+4.56%

Table scrolls horizontally

All four tiers increase monotonically, and they do so on both measures. The lower the rating, the more it rose today — rank correlation is roughly −1. This is not "a few names misjudged", this is a ranker wired backwards.

And the most glaring part: the pre-market list itself got the reason right, and then failed to act on it. Its own judgment was "this is a short squeeze beta ignited by crowded shorts, not a fundamental reversal" — and that sentence was fully validated today. But the list then went on to rank the sector in descending order by "revenue and profit in the most recently disclosed quarter", so: HOOD, the only name with both revenue and profit growing, was ranked No. 1 (−0.70% today), COIN with three straight loss-making quarters was downgraded to No. 6 (+7.58%), and MARA — Q2 revenue −27%, net loss $611 million, explicitly tagged "avoid" — rose +15.54%, the best in the entire list.

The correct formulation is: when a sector's driver is judged to be "a short squeeze", fundamental quality is not the ranker, it is the inverse ranker — short crowding is. Unrealized losses on coin holdings, three straight quarterly losses, BTC-collateralized loans: these are precisely the reasons shorts are willing to be short, and therefore precisely where the squeeze has the most elasticity. The pre-market list put the answer to "who is worth holding" into the slot marked "who will go up today".

⚠️ But read this next paragraph at the same time, or you will draw the wrong inverse conclusion: today the bearish/avoid-side judgments were almost all right — WMT −9.15%, MRNA −23.55%, AAP −24.55%, COTY −9.24%, CRWD −5.60%, all 5 fell; at the sector level, cybersecurity (4/4), homebuilders (5/5) and refiners (3/3) all played out as well. The same list scored near-perfect on "what will fall" and the exact opposite of perfect on "the ranking of what will rise". These two must be booked separately; merging them into a single "hit rate" would simultaneously hide a genuine strength and a systematic defect.

---

0. One-line recap

Today was a risk-off split tape: "macro data too good → the long end reversed → stocks and bonds sold off together, but the squeezed crypto complex and the abandoned memory/optical names led the market higher against a falling index."

① Every major index fell, and the losses increased from large caps to small caps. The S&P 500 closed at 7,641.16 (−0.87%), the Nasdaq Composite at 26,067.17 (−1.00%), the Dow at 52,759.21 (−1.32%), the Russell 2000 at 2,992.43 (−1.34%). The only major index to close green was the Philadelphia Semiconductor Index SOX at 11,800.02 (+0.53%) — this is today's single most important structural fact, see §3. VIX closed at 16.01 (+7.52%), intraday high 16.14.

② The proximate cause of the decline was two macro prints that were "too good". At 08:30 ET, initial jobless claims came in at 206,000, below the 210,000 consensus (prior revised up to 212,000); released at the same time, the August Philadelphia Fed manufacturing index printed 47.4 against a consensus of just 25.0, a five-year high (July 41.4). Strong data → rate-cut expectations pulled back → the long end gave back most of Wednesday's Treasury-buyback-driven decline. Official Treasury CMT: 10-year 4.69% (Wednesday 4.65%, +4bp), 30-year 5.23% (Wednesday 5.19%, +4bp), 2-year 4.19% (unchanged) — the curve continued to bear-steepen, the front end did not budge, and the entire move higher was in duration.

③ The pre-market list's principal falsification point was not triggered, so that judgment stands. The pre-market list explicitly wrote the falsification condition for its "long-end yields rebound" call as "if the 08:30 claims print keeps rising materially, the long end could turn lower on the spot". Actual claims fell (206,000 < prior 212,000), the condition was not triggered, and the judgment stands and is reinforced. But the statement has to be precise — see the conventions table in §1: on the official Treasury closing series, the 10-year closed at 4.69% today, still 2bp below the 4.71% seen before the announcement (8/18); the 30-year at 5.23%, still 5bp below the pre-announcement 5.28%. In other words, about two-thirds (10Y) and about 40% (30Y) of the buyback effect has been given back, but as of the close it has not been fully erased. The pre-market list's "the 10-year has completed a full round trip", written at 08:19 ET off a CNBC read, held at that read timestamp but is slightly too strong on a closing basis.

④ The crypto squeeze did not fade — it accelerated on a day the index fell 0.87%. BTC $72,603 (24h +6.24%, measured 17:02 ET), roughly 1% higher again than the $71,909 the pre-market list read at 08:06 ET. The whole complex rallied, and the gains were highly positively correlated with "bad fundamentals": MARA +15.54%, RIOT +8.26%, HUT +8.00%, CLSK +7.97%, MSTR +7.81%, COIN +7.58%, GLXY +7.17%, SBET +7.20%, BMNR +6.57%, CRCL +6.45%, IBIT +6.24%.

⑤ But inside the same sector, the "quality end" gapped up and faded. HOOD opened +5.31% ($100.85) and closed −0.70% ($95.10), open→close −5.70%; FUTU opened +7.52% and closed +3.03% (open→close −4.18%); BULL opened +15.05% and closed +2.43% (open→close −10.97%). The leveraged end, by contrast, gapped up and kept going: MARA opened +5.80% and closed +15.54% (open→close +9.21%). "Same theme, same day, quality fades and junk runs" — that is the definition of a short squeeze, and it is the single-stock origin of the table at the top of this report.

⑥ Banks fell against the bear steepening, and closed at the day's lows. JPM −1.60%, GS −1.93%, BAC −2.07%, C −2.42%, WFC −2.61%, MS −3.16%, KRE −0.39%, XLF −0.92%. JPM's closing price of 351.55 = the day's low of 351.55; GS/BAC/MS also all closed within less than 0.02% of their lows. The pre-market list's sixth verification signal read "if the long end keeps rising and KRE/JPM still do not rally, the market's worry is credit rather than net interest margin, which is a negative signal" — that verification point was delivered today in its most negative form.

⑦ The only large after-hours earnings move was Ross Stores, and it is the third hit for this week's tariff-refund theme. ROST closed the regular session −2.43%, after hours $245.00 (+6.99%, 868,000 shares); Q2 EPS $2.66, of which about $0.60 / about $253 million came from IEEPA tariff refunds, against original guidance of just $1.85–1.93. See §4.

Tone into tomorrow: cautious. Tomorrow's 09:45 ET August Markit Flash PMI is the capstone of this week's macro narrative — the Philadelphia Fed has already printed a five-year high; if Flash PMI confirms in the same direction, the long end has more room to rise and today's simultaneous stock/bond selloff pattern will continue.

---

1. Market overview

1.1 Indices and volume (CNBC, read 17:00–17:03 ET)

Three pattern features:

1. The Dow's and the Russell's opening prices = the day's highs, with not one meaningful bounce all session — a textbook one-way distribution day.
2. Losses increased strictly in inverse order of market cap (SPX −0.87% < IXIC −1.00% < DJI −1.32% ≈ RUT −1.34%). The Dow's excess decline has an identifiable single-stock source: WMT fell $10.46 per share today (−9.15%), and the Dow is a price-weighted index — 【 estimate 】 on the roughly 0.163 divisor used in recent years, each $1 of share price is worth about 6.1 index points, so WMT alone contributed about −64 points, roughly 9% of the −703.84-point decline for the day. Add HD −2.85% (−$9.81), AAPL −1.75%, AMZN −2.16% and the like, and you have the bulk of the Dow's loss. ⚠️ These point figures are this report's own estimate using the publicly cited divisor; no official point-contribution table was obtained, so they are used to convey magnitude only and should not be quoted as precise.
3. The 1.40-percentage-point inverse gap between SOX and SPX is the biggest sector-level piece of information today.

1.2 Market breadth

NYSE: 693 advancers / 1,514 decliners / 40 unchanged, an advance-decline ratio of about 0.46.

⚠️ This item comes from a single source (the stockmarketwatch NYSE page); this report did not obtain a second independent source for cross-verification, so treat it as an order-of-magnitude reference rather than a precise value. But its qualitative reading is consistent with index behavior: on a day when SOX closed green, decliners still outnumbered advancers by 2.2 times — semiconductor strength was extremely narrow and did not broaden into a general rally.

1.3 Rates, volatility and the dollar

Treasuries (official US Treasury CMT series, struck around 15:30 ET) — this is the one place where this report's conventions most need spelling out:

For comparison: the CNBC quote service's 17:00 ET read was 10Y 4.706%, 30Y 5.252%, 2Y 4.187%, with a 10Y daily range of 4.633%–4.714%.

⚠️ The two sources are not in conflict; they have different read timestamps and must be cited separately: Treasury CMT is struck around 15:30 ET, CNBC is the latest quote at 17:00 ET; yields kept rising about 1.5bp into the close, and that is the entire difference between 4.69% and 4.706%. This report always uses the Treasury series for "day-over-day changes" (same source, same strike time, therefore comparable) and CNBC for "current level", with the timestamp labeled.

⚠️ This is a tightening of the pre-market list, not a reversal of it. At 08:19 ET the pre-market list wrote "the 10-year has completed a full round trip back to the pre-announcement 4.704%" — at that read timestamp, comparing before and after within the same source (CNBC), the statement holds; moreover it explicitly limited itself at the time by saying "not claiming a new high", "not claiming the 30-year's decline was fully erased", "not claiming the buyback tool has failed", and all three of those self-imposed limits proved necessary and correct today. But on a closing basis, the official Treasury series shows the 10-year still 2bp below pre-announcement and the 30-year still 5bp below — most of the buyback effect has been given back, but as of today's close it has not been fully erased. The accurate one-liner is: "a full round trip intraday, closing back at about two-thirds given back", not "fully erased".

Others:

1.4 Sentiment read: risk-off, but not "flight to safety" — it is "repricing rates"

The evidence is three mutually corroborating negative datapoints:

1. The dollar barely moved (DXY +0.03%). If today had been geopolitics/safe-haven driven, the dollar should have strengthened alongside oil +2.70%. It did not.
2. Utilities and health care fell too (XLU −0.57%, XLV −1.87%, the worst of the day). Defensive sectors did not absorb any money — which refutes the narrative that "money rotated from offense into defense". XLV's deep decline has an identifiable single-stock source: MRNA −23.55% (profit-taking after Wednesday's +176.97%).
3. Long-duration assets (TLT −0.82%) fell alongside equities, while the front end did not budge.

So the characterization of today is: strong data → rate-cut expectations trimmed → duration repriced → stocks and bonds fell together. This is the mirror-image day of Wednesday's "buyback → yields fall → homebuilders/precious metals rally hard" — the same causal chain running the other way. The only thing not obeying that chain is precious metals (silver +3.62%), which today looks more like it is following industrial metals / a supply narrative than real rates — the pre-market list wrote at the time that "precious metals in particular are more driven by real rates and the dollar; this report has not verified that correlation", and that self-imposed limit saved it: the half-sentence bundling GLD/SLV together with homebuilders under "avoid" was wrong today, but it had already flagged the evidence grade of that half-sentence in advance.

---

2. Reconciliation against the pre-market list

2.1 Bullish-side single stocks (14 names)

The "vs pre-market price %" column is the point of this table: looking only at the same-day change systematically overstates the hit rate, because the list was published pre-market, so the execution price for anyone following it is the pre-market price, not the prior close.

Bullish-side statistics:

• Directional hits (up/down): 13 / 14 (only HOOD closed lower). But this number is meaningless — the entire crypto complex rose indiscriminately today, so the directional hit rate mainly reflects "the sector was picked correctly", not "the stocks were picked correctly".
• Ranking hits: 0. See the opening table; the four conclusion tiers run monotonically opposite to actual returns.
• Positive vs pre-market price: 8 / 14; of the 6 negatives, 3 (HOOD −3.39%, FUTU −4.44%, BULL −8.29%) happen to be the three most highly rated names in the list.

2.2 Bearish/avoid side (this is the part that was done right today)

Bearish-side statistics: direction 5 / 5 = 100%; 2 of them (AAP, MRNA) fell far beyond what was priced pre-market; 1 (CRWD) hit its verification point exactly; and 1 (COTY) had a "do not short" add-on judgment that saved 12.24 percentage points.

Sector-level avoid list:

✅ A methodological win worth recording: %ADV bucketing works. The pre-market list split retail peers into two groups by "pre-market volume as a share of average daily volume" and said "this group's decline has real absorption behind it and must not be lumped in with the previous group". Today: the real-volume group (COST %ADV 3.67%, etc.) averaged −2.18%, the thin-volume group (LOW 0.13%, TGT 0.21%, etc.) averaged −1.65%, so the real-volume group fell 0.53pp more — direction entirely correct. At the single-stock level it is even clearer: TGT, down 1.65% pre-market in the thin-volume group, closed only −0.47% today (recovering almost all of it), while AZO, the lowest %ADV name in the real-volume group, fell the most (−3.73%) — the latter is the only counterexample. The method itself — "use %ADV to judge how credible a pre-market gap is" — stands up and is worth keeping in the process.

❌ A methodological failure worth recording: the very same %ADV tool produced an inverted conclusion when applied to the AI hardware group. The pre-market list said this group's "rebound this morning had %ADV of only 0.3%–1.2%, with essentially no absorption", and on that basis called it "avoid / short watch". Today the group averaged +0.65%, beating the market by 1.52pp. The difference: the retail group's low %ADV describes "a decline that has already happened and lacks absorption", while the hardware group's low %ADV describes "a rebound that has not yet started" — the former can be extrapolated, the latter cannot. Thin pre-market volume can only falsify "this pre-market price", it cannot falsify "today's direction".

2.3 DE dissected on its own: analysis entirely right, conclusion entirely wrong — the most instructive name of the day

The pre-market list did the week's highest-quality earnings teardown on Deere, and all three core findings were confirmed by third parties today:

1. "Revenue-basis mismatch" — the list pointed out that media were comparing total revenue of $12.608 billion against an equipment-basis consensus, manufacturing about $1.7 billion of beat out of thin air; the real beat was only 0.8%–2.5%. Third-party aggregation today put Deere's consensus at $10.73B (= $10.73 billion, equipment basis), exactly matching the list's judgment.
2. "Nearly all of the year-over-year growth came from tariff refunds" — the 8-K explicitly disclosed a Q3 refund of $110 million, and the list computed that ex-refund EPS was about $4.79, up only +0.9% year over year. That computation has not been overturned today.
3. "The core large-agriculture segment had revenue −6% and operating profit −9%, and is the only segment whose full-year outlook was cut" — factual, uncontested.

And then the stock rose +6.94%, with an intraday high of +10.05%.

The analysis was not wrong; the question the analysis answered was the wrong one. The three findings answer "what valuation does the quality of this earnings report justify"; what the market traded today was "is 2026 the cycle bottom" — i.e. the CEO's own words and the order-book language the list itself had quoted. The list even flagged the correct verification point at the time: "management's specific language on the order book on the 10:00 ET call — that is the only path by which 'cycle bottom' turns from a company view into a verifiable fact." But it locked the rating in at "watch only" before the call.

The lesson is actionable: when a piece of research writes down that "some event that has not happened yet is the only path to verification", the conclusion should not be locked in ahead of that event. The right handling is a conditional conclusion ("upgrade to X if the call confirms order-book improvement"), not locking in a rating and then attaching a verification point. This is in sharp contrast with CRWD — CRWD's verification point ($195) was observable intraday that same day, so locking in was safe; DE's verification point was the 10:00 ET call, so locking in was front-running.

2.4 WMT dissected on its own: the bearish call was right, but the "gap-fill risk" add-on was thoroughly falsified

• Pre-market −5.86% ($107.60) → open $106.38 → close $103.84 (−9.15%), low $102.85. Another 3.50% decline vs the pre-market price.
• The verification point set pre-market: "can it reclaim $110 (half the gap) in the first 30 minutes. If yes, the market understood that 'traffic is not broken'; if no, the zero-growth second-half guidance dominates."
• Actual: the high for the entire day was $107.00 — it could not even reclaim the opening price, and finished 2.7% away from $110. The verification point gave the cleanest possible negative.

The list's main judgment (guided to zero growth in the second half + nearly 38x forward P/E, so the negative is hard) was entirely correct, and it understated the magnitude. What was falsified is the add-on line "intraday gap-fill risk is not low" — justified by "traffic is not broken + advertising +38% / membership fees +17% / e-commerce +23%, a layer the market had not read in as of this morning".

The market did not "read it in" today, and quite possibly has no intention of doing so. Worth noting: this add-on judgment is exactly the same optimistic bias that the list had worked hard in §5.1 to argue against ("do not add back the 125bp", "do not claim it actually beat"), let back in through a side door. It did not show up as "adding back comps" — it showed up as "intraday gap-fill risk". The same bias just changed location. That is more worth remembering than an isolated mistake.

Another signal: WMT was +0.19% after hours (931,000 shares), with no sign of a bounce. Among retail peers TGT was +0.30% and TJX +0.68% after hours, small repairs, but WMT itself had none.

2.5 Hit-rate summary and self-critique

A one-paragraph self-critique

What most deserves to be remembered about today's list is that its verification points were more accurate than its conclusions.

Four of the five verification points gave a correct reading the same day, and two of them pointed directly opposite to the conclusion:

• HOOD's verification point was "its relative strength versus COIN after the open. If it lags, this is still a pure beta move". HOOD −0.70% vs COIN +7.58%, lagging by 8.28 percentage points — the verification point should have killed the No. 1 name shortly after the open, yet the conclusion had it written up as "priority deep-dive".
• DE's verification point was the 10:00 ET call, yet the conclusion was locked in before the call.

So the problem is not "we got it wrong", it is "the conclusions were not constrained by our own verification points". The pre-market list spent a great deal of space on fundamental verification (the verification quality really was high — COIN's three straight loss-making quarters, MSTR's 4.6% unrealized loss, MARA's collateralized-loan feedback loop; not one of them has been overturned today), and then applied those factually correct conclusions to a tape it had itself already judged to be "a squeeze beta, not a fundamental reversal".

In one sentence: today we were not short of facts, we were short of the layer of judgment that asks "is this fact being priced today". Concretely, two things can be hard-coded into the process:

1. When the tape is judged to be "squeeze / sentiment / flow driven", fundamental quality must be explicitly labeled a "non-ranking dimension", otherwise the default ranker will use it as a weight.
2. For any name whose verification point falls after the close or on the next day (earnings calls, next-day data), the conclusion must be written conditionally and not locked in.

---

3. Theme verification

3.1 Feature: SOX +0.53% — the most important structural fact of the day

On a day when the S&P was −0.87%, the Nasdaq −1.00% and the Russell −1.34%, the Philadelphia Semiconductor Index closed +0.53% and SMH +0.31%, a 1.40-percentage-point gap to the S&P. And this sector was avoid target No. 1 in the pre-market list.

There is only one confirmable same-day catalyst: Micron announced a $10 billion, ten-year "Micron Research Labs" memory and AI research campus in Boise (company IR / GlobeNewswire, publication date 2026-08-20, verified; construction starts 2027). MU closed at $974.33 (+3.97%), with an intraday high of $977.28.

The memory chain strengthened in tandem: SNDK +2.02%, STX +2.12%, WDC +1.51% — and these three are precisely the group that led Wednesday's decline (−7.87% / −6.87%) and were put on the list's "short watch".

⚠️ This report must disclose one gap it could not close: no same-day catalyst could be found for LITE +6.24%. Lumentum was the biggest gainer on the hardware side today (close $879.28, intraday high $881.88). Every related article this report retrieved was discarded after checking publication dates one by one: "Lumentum surges 15%" corresponds to the FQ4 earnings on 8/12 (revenue $1.01 billion, doubling year over year), "indium phosphide price increases" to 8/17, "the US bans Chinese optical modules" to 8/4 — not one of them is from 8/20. The COHR share price quoted in those articles ($358) differs by 19% from today's measured close ($290.03), which directly proves they are stale. So this report states only the confirmable part: LITE was +6.24% today, moving with the memory/optical complex, but no company-level catalyst dated 8/20 was found. It may be sector beta or news this report did not cover. This is a genuine coverage gap; it is not the same as "there was no news".

How this theme should be revised:

The pre-market list's basis for "avoid AI hardware" was Wednesday's (8/19) single-day price divergence (SOX −2.12% while NOW +6.45% and CRM +5.07%), plus the low %ADV of this morning's rebound.

Both pieces of evidence are "single-day". The first is a single day's price divergence, the second the volume over a pre-market window. Using one day's rotation to predict the next day's rotation blew up in our face today — and it did so on a down day, which makes the signal stronger: money was cutting the broad market while still adding to this sector.

The more accurate reading is: Wednesday was the outlier, and today was the reversion. The pre-market list had in fact already written the key sentence that would have seen through it — "note: yields fell that same day — if this were 'duration deleveraging', software could not have rallied hard in the opposite direction, so the rate attribution does not hold here". It falsified the wrong explanation (rates), but never applied equally strong skepticism to the alternative explanation ("rotation") itself, instead accepting it outright and setting direction on that basis.

Actionable revision: a one-day rotation can only be an observation item until it is confirmed on day two; it cannot be a reason to avoid or short. Today's evidence: ahead of NVDA's 8/26 earnings, money is not leaving AI hardware — it is adding to memory and optical, the two links closest to "AI capacity".

3.2 Feature: the energy theme's verification point gave a "yes and no" reading

The verification point the pre-market list set for XOM was: "after the open, does the gap between energy stocks and crude converge or widen; if it widens, the whole of theme 5 is void."

• Pre-market gap: crude +3.31% vs XOM +1.66% = 1.65pp
• Closing gap: WTI +2.70% vs XOM +0.84% = 1.86pp → the gap widened by 0.21pp

Taken literally, theme 5 should be void. But that reading is wrong, because it used the wrong vehicle:

The theme holds — upstream energy as a whole (COP/EOG/DVN/FANG/USO averaging +2.27%) beat the market by 3.14pp, the second-strongest direction of the day. What is void is not the theme, it is the choice of XOM as the flagship.

And the pre-market list ranked XOM first in energy, OXY second, and put "COP/EOG/DVN/FANG on watch only (pre-market volume of just a few hundred to ten thousand shares, insufficient evidence)" — today those four "insufficient evidence" names averaged +2.15%, while the "sufficient evidence" XOM managed only +0.84%. This is the third appearance of the same pattern: pre-market activity levels were used as a ranking basis for "will it rise today", when what they actually measure is only "is this pre-market quote credible".

XOM's lag has a structural cause, and the list itself had written it down: it is an integrated company, so its refining segment gets squeezed by high crude prices (today's MPC −0.70%, VLO −1.37%, PSX −0.95% are the proof). The list wrote "refiners are on the losing side" in the theme table, but never carried that back through to the segment structure of the integrated flagship XOM itself. Pure upstream names (COP, DVN, EOG) do not carry that drag, which is why they outperformed. To express "crude is rising", pick pure upstream or USO, not the integrated flagship.

3.3 Unexpected themes missed pre-market

---

4. After-hours earnings moves

4.1 Ross Stores (ROST) — the only large after-hours move today, and the third hit for the tariff-refund theme

• Regular session close $228.99 (−2.43%) → after hours $245.00 (+6.99%, 868,000 shares, 17:03 ET)
• Source: company press release (PR Newswire, 2026-08-20¹)

Guidance: Q3 comps +6–7%, EPS $1.75–1.83; Q4 comps +4–5%, EPS $2.17–2.26; full-year EPS $8.61–8.77 (the company explicitly states this includes about $0.60 of tariff benefit); 115 new stores for the year.

⚠️ This report wants to draw an explicit distinction between this and WMT / DE, because it is the week's first case of "contaminated, but the underlying business really is strong".

• Excluding the $0.60 refund, EPS is about $2.06, still about 7%–11% above the original $1.85–1.93 guidance. That is a different animal from WMT (about $0.75–0.76 ex-refund, only a small beat vs consensus) and DE (only about +0.9% year over year ex-refund).
• Comps +10% with the CEO explicitly attributing it to traffic growth (both new and returning customers) — that item cannot be explained by refunds at all.
• The company proactively flagged the refund amount in its guidance — honesty on a par with Deere, better than Walmart.

But two deceleration signals must be stated at the same time: ① comps guidance steps down quarter by quarter, from +10% this quarter to +6–7% in Q3 and +4–5% in Q4; ② the full-year $8.61–8.77 includes about $0.60 of refunds, so ex-refund it is about $8.01–8.17 — and this single quarter consumed the entire full-year refund amount, meaning there is no refund available in Q3/Q4 and the base has already been raised. This is structurally the same as the pre-market list's "Q3/Q4 base already raised" judgment on WMT/DE.

What it means for tomorrow: the +6.99% after-hours move is pricing "beat + real traffic", not pricing the guidance. What to watch after tomorrow's open is whether it can hold the after-hours price — ROST's regular session itself closed −2.43%, so the after-hours price of $245.00 amounts to +4.39% versus the prior close, a gap that needs intraday volume to confirm it.

4.2 Other after-hours action

The broad market was nearly flat after hours: SPY +0.05%, QQQ +0.02%, DIA −0.01%, IWM −0.07%. There was no sign of the intraday decline extending after the close, and no bounce either.

⚠️ This report's coverage of the after-hours movers list has an explicit boundary: checking a third-party after-hours gainers/losers list, apart from ROST, the top five movers in both directions tonight are all micro-caps (HOWL, SUGP, PFSA, FBGL, IPST / JZ, FLUX, YOUL, ZSTK), the only one of any size being OSIS (−12.88%). This report did not verify these micro-caps and does not include them as next-day catalysts.

---

5. Flows and sentiment

5.1 Sector rotation (SPDR sector ETFs, benchmarked to SPY −0.84%)

Two readings:

1. There were only two strong directions today: semiconductors and energy. Both sit in the pre-market list's avoid/downgrade column.
2. "Defensives did not receive any money" is the most important negative datapoint of the day. XLU −0.57% and XLV −1.87% both fell. This is not "money rotating from offense into defense", it is "duration assets being repriced indiscriminately". It shares a root with a piece of evidence the pre-market list once used: software (short-duration sensitive) and hardware (physical demand) parted ways today, which again shows rates are not the whole explanation — what is really being sold is everything with a high valuation and no near-term cash-flow catalyst.

5.2 Banks: the most notable contrarian signal of the day

The pre-market list's sixth verification signal: "if the long end keeps rising and KRE / JPM still do not rally, the market's worry is credit rather than net interest margin, which is a negative signal."

Today: the long end did rise (10Y +4bp, 30Y +4bp), the curve did steepen (2Y flat), and all six large banks fell, nearly all closing at the day's lows. By the list's own rule, this is a negative signal, and it is the cleanest reading among all of today's signals — it simultaneously rules out the optimistic "margin improvement" explanation.

Where restraint is needed: the list's original text says "the worry is credit". This report does not make that assertion, because there is no credit event today that can be pointed to. Only two things are certain: ① bear steepening did not buy bank upside, so that mapping does not hold today; ② all six closing at the day's lows is a distribution pattern, not an absorption pattern. "Credit worry" is one of several possible explanations (others include investment-banking revenue expectations and general deleveraging), and this report has not ruled any of them out.

At the same time, this also corrects the pre-market list's write-up of KRE/JPM as "unpriced beneficiaries within the long-end yield rebound theme" — that mapping was rejected today.

5.3 Volatility: VIX jumped for a second straight day, on a decline that was not large

• VIX 16.01 (+7.52%), versus 14.89 yesterday (already +7.39% at the pre-market timestamp). Cumulatively about 15% higher over two trading sessions.
• The S&P fell only 0.87% while VIX jumped 7.52% — the pre-market list flagged this divergence yesterday ("a small decline with a volatility jump usually means money is buying protection against a known event"), and today it persists, and is still widening even after the index actually fell.
• Known events still to be priced: tomorrow's 09:45 ET Flash PMI, NVDA after the close on 8/26, Jackson Hole on 8/27–29.

But 16.01 is still a very low absolute level. This is not panic; it is hedging costs going from extremely low back to somewhat low. Combined with "six banks closing at their lows" and "defensives falling too", it looks more like the early stage of systematic deleveraging than a rotation into safety.

5.4 risk-on / risk-off characterization

Risk-off overall, with two clear exception pockets:

• Risk-off evidence: all four major indices fell with small caps falling most; NYSE decliners 2.2 times advancers; VIX +7.52%; banks closing at their lows; defensives falling too; the Dow and the Russell putting in their highs at the open.
• Exception pocket 1 (speculative risk-on): crypto rallied across the board, and the junkier the bigger the gain — MARA +15.54% led IBIT +6.24% by 9.3pp. This is a purely technical squeeze move and does not represent a recovery in risk appetite.
• Exception pocket 2 (fundamental risk-on): semiconductors and energy closed green against the tape, each with an identifiable catalyst (Micron's $10 billion research campus / WTI +2.70%). This pocket is more worth tracking than the first, because it has a real industrial catalyst and it happened on a down day.

---

6. Outlook for the next session (2026-08-21, Friday)

① Theme persistence

② Tomorrow's earnings and macro calendar

Macro (Eastern Time):

⚠️ This report's coverage of tomorrow's calendar has a gap; clients please note: the econoday calendar lists only Treasury and equity settlement items for tomorrow, with no macro data detail; the 8/21 release date for Flash PMI was verified via S&P Global's official page. This report could not obtain the consensus figure for tomorrow's Flash PMI, so no forecast value is given. Other second-tier data that may be scheduled for tomorrow could not be verified one by one.

Earnings: third-party calendars show about 18 companies reporting tomorrow, but this report could not verify the specific names or which large companies are among them, so no list is given. Judging by magnitude (55 on Thursday → 18 on Friday), tomorrow is not a heavy earnings day; the aftershocks of ROST after the close and the 09:45 Flash PMI are the main variables.

Key dates next week (confirmed):

• 8/26 (Wednesday) after the close: NVDA FQ2'27 (management previously guided to about $91 billion of revenue) — also the adjudication point for the theme "is AI hardware being repriced"
• 8/26: CRM, CRWD — CRWD is the closeout point for this week's short watch
• 8/27–29: the Jackson Hole global central banking symposium — against a backdrop of "strong data → trimmed rate-cut expectations", the weight on Powell's language is materially raised

③ Focus list (Ticker + verification point)

This section is rewritten under the rule derived from today's self-critique: write the verification point first, then the conclusion; for any name whose verification point falls after the close or on the next day, write a conditional conclusion and do not lock in a rating.

④ What to avoid

1. Chasing the leveraged end of the crypto squeeze on day three (MARA / RIOT / CLSK / HUT / MSTR / BMNR / SBET). The reason is not weak fundamentals (that is the ranker proven wrong today), it is that the squeeze fuel is being consumed: $2.7 billion was already liquidated on Wednesday, a record since 2021, and the stock of shorts available for forced liquidation has fallen; and the quality end has already started gapping up and fading today (HOOD −0.70%, BULL open→close −10.97%), which usually precedes a top in the leveraged end. What rose most today is precisely what carries the most drawdown risk tomorrow.
2. WMT and retailers whose guidance has been cut. WMT closed −9.15%, low $102.85, with no bounce whatsoever after hours; neither the zero-growth second-half guidance nor the nearly 38x forward P/E was falsified today. The "intraday gap fill" thesis has been thoroughly falsified today — stop expecting it.
3. Homebuilders and long-duration rate-sensitive names (DHI/LEN/TOL/PHM/XHB). 5/5 fell today with a mean of −2.43%; under the strong-data chain of Philadelphia Fed 47.4 + claims 206,000, the premise that would support them rising (falling yields) has not held for two days running.
4. MRNA-related bounces within health care. −23.55% today (two days: +176.97% → −23.55%), and phase III has still not released any quantitative efficacy data — that evidence gap existed pre-market and still exists today. Dangerous in both directions.
5. The "rotation beneficiary" narrative in software. IGV −0.88% today, with everything except NOW (+2.00%) falling; this narrative's only basis was Wednesday's single-day divergence, and it has been overturned by contrary evidence.
6. Using integrated oil majors as a vehicle for crude (XOM). This is not a bearish view on XOM; it is that the squeeze on its refining segment makes it an inefficient way to express rising crude (+0.84% today vs COP +3.30%).

⑤ Input notes for tomorrow's pre-market list

This is what this report most needs to hand to tomorrow morning's list — all three problems exposed today are process problems, not judgment problems:

1. ⭐ The ranker must match the driver, and it must be declared explicitly. Today's four conclusion tiers ran monotonically opposite to actual returns, and the root cause is: the list judged the driver to be "a short squeeze", yet used "fundamental quality" as the ranker. After stating theme strength, tomorrow morning's list must explicitly write one line: "the ranking dimension for this theme is X" — flow/sentiment-driven themes rank by short crowding and elasticity; only fundamentally driven themes rank by financial quality. The default ranker will automatically use fundamentals, and that must be actively overridden.

2. ⭐ Conclusions must be constrained by their own verification points; where the verification point falls after the close or on the next day, do not lock in a rating. Four of today's 5 verification points gave a correct same-day reading, and HOOD's pointed directly opposite to the conclusion. The rule:

   • Verification point observable intraday the same day (e.g. CRWD's $195) → a rating may be locked in
   • Verification point after the close or on the next day (e.g. DE's 10:00 call) → only a conditional conclusion is allowed

3. ⭐ %ADV can only falsify "is this pre-market price credible", not "is today's direction right". The same tool gave one right and one wrong answer today: the retail group (describing whether a completed decline has absorption) ✅, the AI hardware group and the energy group (describing a rebound that has not yet started) ❌ ❌. When using it tomorrow morning, it is essential to distinguish whether it is measuring "a gap that has already happened" or "a move that has not happened".

Three more specific inputs:

4. Memory/optical/semiconductors need to move from the avoid column to the focus column, with a fresh fundamental and catalyst check tomorrow morning (the Micron campus, positioning ahead of NVDA on 8/26). The evidence strength of this revision is high — it happened on a down day.
5. The tariff-refund theme is now in its 4th day, but the reading must change: the market punishes guidance, not refunds. When screening retail/industrial earnings tomorrow morning, the weight should be on guidance vs consensus, with refunds used only to judge whether the headline number is usable.
6. Two known coverage gaps need to be closed tomorrow morning: ① no same-day catalyst was found for LITE's +6.24% today (all search results were stale once publication dates were checked); ② WOLF −9.42% is a genuine miss, being outside the pre-market theme pool. The pre-market list's candidate pool is hand-constructed by theme (about 170 names), not a whole-market ranking — moves outside the pool will be systematically missed, and this structural gap remains unsolved.

---

Data-gathering and execution log (internal)

Data-gathering failures

• yfinance returned a blanket YFRateLimitError for a third consecutive trading day; even a minimal SPY sample test failed, confirming this is still an IP-level ban rather than a leftover process. Not one number in this report comes from yfinance.
• All quotes in this report were switched to the CNBC quote service (quote.cnbc.com/quote-html-webservice/restQuote, with exthrs=1), pulling 133 symbols in 7 batches, read at 17:00–17:03 ET. After-hours prices come from the ExtendedMktQuote field, all of type POST_MKT, with last_timedate checked to be after 16:00.
  • Consistent with the cnbc-extended-hours-asymmetry memory: for the after-hours slot, taking the after-hours price and the prior close is reliable (it is the pre-market slot that has to switch to stockanalysis).
• WebFetch returned 403 for both tradingkey.com and philadelphiafed.org. The Philadelphia Fed data was sourced instead from a haver.com article (publication date verified as 2026-08-20) and cross-checked against independent WebSearch results (47.4 / prior 41.4 / consensus 25.0).
• FRED's fredgraph.csv endpoint (ICSA, GACDFSA066MSFRBPHI) returned an empty response to curl, so neither series was obtained. Claims and the Philadelphia Fed were switched to WebSearch + haver as a two-source combination.
• A direct connection to treasury.gov's daily_treasury_yield_curve CSV succeeded, yielding four complete curves for 8/17–8/20 — this is the highest-quality data in this report, and the only rate source that permits a same-source, same-strike-time before/after comparison.
• The CNBC symbol SPLK returned an empty record (delisted after the Cisco acquisition, not an endpoint failure). SK returned an exthrs type of PRE_MKT with an anomalous timestamp; it was discarded and is not cited.
• A WebFetch of earningswhispers.com's Friday calendar page returned only the navigation shell with no actual entries; econoday's homepage listed only settlement items. Tomorrow's earnings list therefore could not be obtained, as stated honestly in the body.

Cross-check against the memory bank (old traps hit this time)

• recap-needs-ohlc-not-just-close: this report's core finding depends entirely on OHLC. HOOD was only −0.70% on the day; looking at the close alone would have called it "basically flat, neither here nor there". Only with the opening price of 100.85 does it become visible that it opened +5.31% and then collapsed in a straight line, open→close −5.70%. The same goes for BULL (open→close −10.97%) and COTY (open→close +7.84%, which rescued the "do not short" judgment).
• recap-vs-premarket-price-column: necessity confirmed again. The vs-pre-market-price convention corrects MSTR from "+7.81%, wrong call" to "−0.87%, a wash at the execution level", and BMNR from "+6.57%, wrong call" to "−0.19%, avoid was correct"; in the other direction it magnifies HOOD from "−0.70%, a small miss" to "−3.39%, a real loss". 6 of the 14 bullish names were negative versus the pre-market price, while on a same-day basis only 1 was negative.
• phantom-earnings-from-search-summary + recurring-column-headlines-hide-date: hit twice this time. ① All the search results for LITE were stale articles from 8/4, 8/12, 8/14 and 8/17; the COHR share price of $358 quoted in them differs by 19% from today's measured close of $290.03, so the date was falsified by back-computing from price, and the whole block was removed with the gap stated in the body. ② One source on MU wrote that it "broke $1,000 and then fell back to around $950", contradicting the CNBC measurements (prior close 937.11 / high 977.28 / close 974.33); another source's $956.74 (+2.10%) back-computes to a prior close of exactly 937.11, proving it is merely an earlier intraday snapshot, so all third-party prices were discarded in favor of self-collected data.
• stale-source-silent-lag / cboe-stale-symbol-silent-trap: CNBC's 17:00 read for US10Y was 4.706%, identical to the pre-market list's 08:19 read, and the first reaction was to suspect a frozen quote. The open/high/low on the same record (4.645 / 4.714 / 4.633) prove the value did move intraday, ruling out staleness; the day's direction (+4bp) was then independently confirmed with the official treasury.gov series. A numerical coincidence does not mean the data is stale, but it must actually be falsified before use.
• cross-market-data-needs-read-timestamp: Treasury CMT (struck around 15:30) and CNBC (17:00) differ by about 1.5bp, split out explicitly in §1.3 and not mixed. The pre-market list's "10Y completed a full round trip" and this report's "still 2bp lower at the close" are not contradictory; they are two read timestamps — and if that is not spelled out, it reads as the two reports contradicting each other.
• financial-quality-is-not-a-short-term-ranker: this memory replayed itself today in the US slot in its most extreme form (the A-share instance was 盛美上海 (ACM Research Shanghai) with ex-non-recurring profit −15.31% yet up +12.16% and best of the day; today it is MARA with revenue −27% and a net loss of $611 million yet +15.54%, best in the entire list). Four conclusion tiers running monotonically backwards, rank correlation −1, is far stronger than the single counterexample in the original memory. Recommend upgrading that memory into a cross-market general rule and adding the actionable patch "the ranking dimension must be declared explicitly".
• verification-point-beats-headline-conclusion: today is the strongest confirmation this memory has had — 4 of 5 verification points correct, while the conclusion tiers scored 0. HOOD's verification point (relative strength vs COIN) pointed directly opposite to the conclusion.
• no-news-found-is-not-no-news: LITE was handled per this memory — no catalyst found ≠ no catalyst exists, so the body states it as a coverage gap rather than "no news".

Data this report could not obtain / gaps that exist (all stated in the body)

1. Market breadth is single-source (stockmarketwatch, 693/1,514/40); no second independent source was obtained for cross-verification, nor were Nasdaq breadth and new-high/new-low data.
2. The same-day catalyst for LITE's +6.24% today — everything retrieved was stale and was discarded.
3. Tomorrow's (8/21) earnings list and the consensus figure for Flash PMI.
4. The list of second-tier macro data for tomorrow beyond Flash PMI could not be verified item by item.
5. Today's (8/20) analyst rating changes — the same gap as in the pre-market list; likewise not covered here.
6. After-hours micro-cap movers (HOWL/SUGP/JZ/OSIS etc.) were not verified, and are not included as next-day catalysts.
7. The Dow point attribution (WMT ≈ −67 points) is this report's own estimate on a divisor basis, with no official point-contribution table obtained; used to convey magnitude only.

Methodology notes

• This report did not call any sub-agents (fundamentals-analyst / risk-auditor). This runs counter to the advice in the us-report-fundamentals-agents-worth-waiting memory, because the environment instructions for this run explicitly require not invoking the Agent tool unless the user asks. Compensating measures: every key number was either dual-sourced or self-checked by back-computation (rates: treasury.gov + CNBC, two sources; Philadelphia Fed: haver + WebSearch, two sources; ROST: the original company press release; group statistics: computed by script rather than by hand, see work/recap0820/recon.py). But it must be acknowledged that this report lacks an independent adversarial quality check — especially for the "monotonically inverted" conclusion in §2, whose arithmetic has been re-checked but which no second party has tried to refute.
• Group means and the "vs pre-market price" figures were all computed by script (work/recap0820/recon.py); pre-market prices were back-computed from the pre-market list's disclosed "previous close × (1 + pre-market change)", and pre-market changes were always taken from the values and read timestamps labeled in §3 of the pre-market list's full single-stock table, without mixing in the 08:27–08:28 re-read round in the §0 table (the two rounds differ slightly for a few names such as IBIT; the convention has been unified).

---

⚠️ 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 in the basis of measurement; company disclosures / SEC filings prevail, and this should not be used directly as a basis for trading.