A-Share Market Recap | 2026-09-04 Friday

Scope and data boundaries of this report (judge the strength of the conclusions accordingly) ① All A-share quotes come from Tencent qt.gtimg.cn on a closing-settlement basis, read at 20260904153031–20260904153515, all later than the 15:00 close. The price triple (previous close / open / close) has been reverse-computed and verified stock by stock. ② The limit-up / failed-limit-up / limit-down pools use East Money's official push2ex pools (limit-up 39 / failed limit-up 48 / limit-down 8), cross-checked with stock_market_activity_legu (limit-up 41, true limit-up 38; limit-down 9, true limit-down 6); the caliber differences between the two are discussed in §1. Advancer/decliner counts use the legu official summary, not East Money clist pagination (that endpoint returned 502 across the board today). ③ Sectors are THS's 90 industry sectors on the 9/4 closing snapshot; sector fund flows are on East Money's caliber (obtained via the push2delay cluster). The two are not the same yardstick and must not be subtracted from each other. ④ No northbound flow data (daily net buy has not been disclosed since 2024-08-19), so this report gives no northbound figures. ⑤ ⚠️ This report did not obtain the 9/4 Dragon-Tiger List (LHB) — the exchanges publish it after 18:00 and this report was generated at 15:30, so it is unobtainable at that moment. This is not a data gap, it is a scheduling limit; tomorrow morning's pre-market piece can fill it in. ⑥ Limit-up determination follows board-specific limits: main board ±10%, ChiNext/STAR ±20%, BSE ±30%. Every stock in this report is tagged with its board. ⑦ The "vs. pre-market price" column = open→close change. The pre-market list is published before the open, so the execution price for anyone following it is today's opening price, not yesterday's close; the reconciliation therefore gives both the close-based and the open-execution calibers. The gap between the two reached 0.79pp today (buy side) — do not look only at the former.

⚠️ Data-fetch and search log (this section is not sent to clients):

I. Channels that worked

• Individual stocks/indices via Tencent qt.gtimg.cn; 39 stocks + 10 indices returned in a single call, 0 timeouts, and the f[30] timestamp was verified stock by stock as being after 15:30.
• The limit-up / failed-limit-up / limit-down / previous-day-limit-up pools all via push2ex, a different cluster from clist and therefore unaffected today; all four were normal.
• stock_market_activity_legu() normal, giving official advancer/decliner counts.
• stock_board_industry_summary_ths() returned 90 rows, normal.
• Daily K-lines switched to Tencent web.ifzq.gtimg.cn/appstock/app/fqkline (akshare's stock_zh_a_hist depends on push2his, which was also down today).
• Announcements via East Money np-anotice-stock; original PDFs via pdf.dfcfw.com + local pypdf parsing (WebFetch returns garbled text for PDFs, they must be parsed locally).

II. Failed channels / today's data-fetch faults

• ⚠️ East Money push2.eastmoney.com returned 502 across the board; 82.push2/push2his gave outright RemoteDisconnected. Affected: stock_sector_fund_flow_rank (JSONDecodeError), stock_zh_a_hist, and whole-market clist pagination.
• ✅ Followed memory [[ths-fallback-when-em-clist-down]] and used the fallback route, but found one more this time: push2delay.eastmoney.com still returned 200. After the close, the delayed cluster's closing prices match the real-time cluster's (verified stock by stock against Tencent using Shandong Gold / Zijin), so sector fund flows and per-stock main-force net amounts were still obtained on East Money's caliber — no need to degrade to "no sector-level fund flow obtained". This one should be added to memory.
• ⚠️ stock_lhb_detail_em(20260904) threw TypeError: NoneType — same phenomenon as yesterday's recap, and the cause is again scheduling (published only after 18:00), not an endpoint fault. This is the third recurrence; no need to investigate again.
• Three rounds of searching for "Inspur Electronic Information 9/4 limit-down" returned only old July articles and the old interim-report piece; no same-day attribution obtained. Written up under the rule "not found ≠ does not exist".

III. Five cross-checks actively performed on method that changed the conclusion

1. ⚠️⚠️ [The single most important one in this report] "Precious metals fell because gold fell" — refuted by my own measurement. The first draft was about to attribute the −1.15% of the 8 precious-metal names to "giving back the overnight gain, gold price falling". Measured gold-api directly: spot gold $4,476.70 (read at 2026-09-04T07:37:56Z = 15:37 Beijing) vs. the pre-market piece's $4,475.70 (read at 2026-09-03T23:34Z); over the A-share session the net change in gold was +0.02%, i.e. flat, not down. In other words: the numerator (the gold price) did not move; what fell was the denominator (what people are willing to pay for it). That turned the conclusion from "the bullish case was falsified" into "the bullish case still stands, but nobody wants to hold it over the weekend ahead of nonfarm payrolls" — and the two have opposite implications for next Monday. Without the measurement, I would have written an internally consistent but directionally wrong attribution.
2. ⚠️⚠️ [Wrong in the first draft, corrected after cross-checking — China State Shipbuilding was NOT an "in-window miss"] The first draft, going off East Money's notice_date=2026-09-03 and the announcement's signature date "September 3, 2026", wrote up 中国船舶 (China State Shipbuilding, 600150)'s US$1 billion order as an "in-window miss (9/3 15:00→9/4 07:00)". Pulling eiTime overturned that: the actual publication moment was 2026-09-02 19:02:48, a full day outside the window. notice_date uses the next trading day and the signature line uses the board-meeting date; neither is the publication moment — only eiTime (consistent with the art_code timestamp) is. This is the second appearance of the same pitfall recorded in memory [[cumulative-period-figures-are-old-news]] ("East Money's notice_date is the next trading day"), and this time I stepped in it from the opposite side: last time I treated an old announcement as new, this time I treated an out-of-window one as in-window and wrote a paragraph of criticism on that basis that does not hold. ✅ The corrected conclusion is actually more valuable: the announcement came out on the evening of 9/2, the stock rose only 1.15% on 9/3 (essentially no reaction), and the pricing happened on the second trading day, 9/4 (+9.18%). This is exactly the second dimension described in memory [[scan-window-misses-multiday-themes]] — "in-window" and "already priced" are two different things, and "the stock didn't move" ≠ "the information wasn't disclosed" (see also [[trace-to-first-disclosure-date]]). §3.6 has been rewritten accordingly.
3. ⚠️⚠️ [The genuine in-window miss — one of them, and it was checked selectively] 海通发展 (Haitong Development, 603162) disclosed a private-placement plan at eiTime=2026-09-03 18:39:07 (announcement nos. 2026-089/092, proposed raise RMB 2 billion, implied issue price RMB 9.50 per share) — squarely inside the window. It hit +9.99% limit-up today, while the pre-market piece had tagged it "watch only". The damning part: that same evening the pre-market piece checked 长电科技 (JCET, 600584, eiTime=2026-09-03 19:54:02)'s placement announcement and wrote it into the §1 table, yet did not check Haitong Development's — even though the latter was published 1 hour 15 minutes earlier. So this is not "the cutoff moment didn't cover it"; it is the criterion being executed only on the names I wanted to recommend, isomorphic to memory [[checklist-skipped-on-favored-branch]]. The correct practice is to pull announcements name by name across the candidate pool, rather than searching by theme keywords.
4. ⚠️ ["Agriculture was an unforeseeable theme" — half-refuted by myself] The first draft wrote "agriculture was a complete surprise, unforeseeable pre-market". After checking the 9/3 limit-up pool (44 names, data the pre-market piece already had in hand): 亚盛集团 (Yasheng Group, 600108) was already a first-day limit-up on 9/3, and today it made its 2nd consecutive limit-up at +10.10%. But the other half must be stated too: in the 9/3 limit-up pool there was only that 1 agriculture name (9/2 had 2), a density far below the 8/31 instance (3 out of 88, including the highest consecutive-limit-up name of the day). So the conclusion is "the signal was weak but present", not "the data was in hand and went unread" — different in nature from the 9/1 recap, and the same self-criticism cannot be copied over.
5. ⚠️ [The closing-commentary publication timestamp passed the check this time] Chinanews's "A-Share Closing Commentary" was confirmed via WebFetch to have been published at 2026-09-04 15:28, and its "Shanghai Composite −0.30%, Shenzhen Component −0.79%, ChiNext −0.78%, turnover RMB 2.05 trillion, STAR 50 down more than 2%" matches this report's own data item by item — it is not the "opening snapshot masquerading as a closing recap" described in memory [[opening-snapshot-sold-as-close-recap]]. Also: this report's own arithmetic of Shanghai + Shenzhen RMB 2,030.7 billion + BSE RMB 20.6 billion = RMB 2,051.3 billion ≈ RMB 2.05 trillion closes with the commentary.

IV. What this report did not obtain / did not do (listed honestly)

• The 9/4 Dragon-Tiger List (scheduling issue, see above).
• No same-day attribution obtained for 浪潮信息 (Inspur Electronic Information, 000977)'s limit-down. Verified: the company's most recent announcement is the 8/29 interim report, and there were no new announcements on 9/3–9/4; three rounds of searching found no same-day news. The price facts are hard (−9.99%, turnover RMB 13.708 billion, main-force net outflow RMB 1.826 billion, 1 break off the limit), but the cause is unknown.
• The use of proceeds for Haitong Development's placement was not obtained (whether ship purchases / ship types / number of hulls). Only the raise size of RMB 2 billion and implied issue price of RMB 9.50 per share were taken from the "dilution of immediate returns" announcement (no. 2026-092). Therefore the chain "placement = buying ships = capacity expansion is bullish" has no evidence in this report and is not adopted. ⚠️ Note the reverse possibility as well: JCET's placement in the same batch fell −5.49% today, so "a private placement" is not in itself directionally bullish; the causal link between Haitong Development's limit-up and its placement is not confirmed here, and dry-bulk freight rates or thematic factors cannot be ruled out.
• The same-day trigger for today's agriculture theme was not independently anchored. What the searches returned were cross-week narratives such as "expectations of a hog-cycle reversal", "international grain prices bottoming out", "a triple resonance of geopolitics + extreme weather + policy control" — none of which can be anchored to a specific event on the evening of 9/3 or the morning of 9/4. Written under the rule "not found ≠ does not exist".
• BDTI / VLCC TD3C freight rates — the top priority carry-over item flagged by the pre-market piece; still not obtained here, a second consecutive report without closure.
• The real trigger for the 9/3 overnight jump in precious metals (jobless claims / ISM / Warsh) — a must-check item for the pre-market piece; still not obtained.
• Did not verify fundamentals and announcements name by name for all 39 limit-up stocks.

---

0. One-Sentence Recap

Sentiment: divergent and on the cool side, and it is a "blood-transfusion" kind of cool — the indices barely fell, but the internal structure tore apart. Shanghai Composite −0.30%, Shenzhen Component −0.79%, ChiNext −0.78%, while STAR 50 −2.10%, CSI 500 −1.32%, CSI 1000 −1.21%, and SSE 50 bucked the trend at +0.15%. 39 limit-ups, 8 limit-downs, and the failed-limit-up rate jumped from yesterday's 42.9% to 55.2% (48/87); 2,249 advancers vs 2,773 decliners (advancers were 44.8% of the 5,022 stocks that moved; plus 185 unchanged and 8 suspended). The indices' composure was held up entirely by the heavyweights, while small/mid caps and the tech chain collapsed underneath.

Today's strongest theme was "big agriculture + big consumption + media", and the pre-market list did not write a single word about it. Livestock farming +4.36% (33 up, 3 down), agricultural processing +3.60% (40 up, 3 down), baijiu +2.87% (19 up, 0 down), planting & forestry +2.48%, cultural media / film & cinema +2.44%. Of the 39 limit-ups, agriculture and food accounted for 14 (35.9%), the only limit-up cluster that formed a family. The weakest was semiconductors at −2.82% (13 up, 172 down), main-force net outflow RMB 15.311 billion; the level-1 electronics sector saw a net outflow of RMB 27.224 billion.

Pre-market list hit rate overview: 22/33 overall (66.7%), but the structure is "Pass group a big win, buy side flat, theme call a complete miss". In the Pass group, 17 names produced 14 underperforming the Shanghai Composite (82.4%), mean −2.67% — the hardest block in the whole report. The buy side's 16 names averaged −0.29%, almost exactly equal to the Shanghai Composite's −0.30%: a full report's worth of work, and zero excess return. The 5 recommendation slots closed at −0.23%, and at the opening execution price, −1.52%.

⚠️ But the thing most worth remembering today is not that the theme call was wrong; it is an attribution that could have been falsified on the spot and wasn't. The pre-market piece's first branch was precious metals. Today the 8 buy-side names in that branch averaged −1.15%, with only 2 up, and the THS precious-metals sector was −1.78%, 2 up 12 down. But gold did not fall at all today — spot gold $4,476.70 (read at 2026-09-04T07:37:56Z) against the pre-market piece's $4,475.70 (read at 2026-09-03T23:34Z), a net change of +0.02% over the whole day, i.e. flat. So this is not "the bullish case was falsified", it is "the bullish case still stands, but nobody wants to carry it over the weekend ahead of tonight's payrolls." The pre-market piece wrote that sentence itself — "the whole day is a front-running window, not a confirmation window" — it got the nature of the day right, and still put 4 of its 5 recommendation slots into it. Cognition and position sizing were disconnected.

Tone for the next session (Monday 9/7): the theme must be rotated, and the first task is to do the homework on agriculture and media, not to give precious metals a second chance. The referee for precious metals blows the whistle tonight at 20:30 (after the A-share close); by Monday's open the result is already old news and has already been fully priced by Friday's US session — what A-shares face then is "already repriced", not "yet to be priced". What is genuinely unpriced is the agriculture/consumption line that only got volume today.

---

1. Market Overview

1.1 Indices and Turnover (Tencent qt.gtimg.cn, read at 15:34–15:35)

Shanghai + Shenzhen combined RMB 2,030.7 billion + BSE RMB 20.6 billion = RMB 2,051.3 billion, which closes with Chinanews's closing commentary figure of "RMB 2.05 trillion" (cross-check passed).

⚠️ The most important row in this table is "SSE 50 +0.15% vs STAR 50 −2.10%", a spread of 2.25pp. This is not a broad decline; it is a single-day migration from tech growth into heavyweights/value. The index looks like it only fell 0.3%, but if your book sits on the STAR Market, today felt like −2%.

1.2 Limit-Ups/Downs and Sentiment Temperature

Caliber note: the official limit-up pool has 39 names, while legu's summary reports "limit-up 41, true limit-up 38, ST limit-up 2"; the limit-down pool has 8 names, while legu reports "limit-down 9, true limit-down 6". The difference comes from how ST stocks and BSE names are counted, not from a data error. This report uses the official pool's 39/8 throughout and computes the failed-limit-up rate on that basis, with legu used only for cross-checking. THS's 90 sectors sum to 2,444 up / 2,914 down, directionally consistent with legu's 2,249/2,773 (sector constituents do not cover the whole market, hence the different absolute values).

Sentiment call: divergence within a pullback — neither a freeze nor euphoria. Three signals:

1. The 55.2% failed-limit-up rate is the key one — the number of limit-ups fell by only 5 while failed limit-ups rose by 15, meaning the money trying to push stocks to the limit has not shrunk, but the money willing to take the baton clearly has. The pre-market piece already flagged "9/3 weakened into the close, the money-making effect is contracting", and that call was right.
2. Limit-downs fell from 16 to 8, so panic did not spread — this is not systemic risk, it is a structural rotation.
3. The maximum streak held at 5 but changed hands: yesterday's 5-streak 国芳集团 (Guofang Group, 601086, Shanghai main board) broke its streak today (intraday high +9.91%, closed +3.60%), and the baton passed to 龙版传媒 (Longban Media, 605577, Shanghai main board) at 5 consecutive limit-ups. The height did not collapse, but the leader changed — and changed into today's strongest line, media — a textbook marker of a theme rotation.

---

2. Pre-Market List Reconciliation

Caliber: "vs. pre-market price" = open→close change (the true cost basis for anyone executing the list). The Shanghai Composite was −0.30% today, which is the benchmark for "outperform".

2.1 Recommendation Slots: 5 Names Today

Recommendation group mean: close −0.23%, vs. pre-market price (open-execution caliber) −1.52%. Beat the Shanghai Composite 2/5, closed up 2/5.

⚠️ The two calibers differ by 1.29pp, and unfavourably in the same direction — all 5 gapped up (+0.52% to +1.62%) and then all 5 closed below their opening price. This is the most textbook "gap up then slump" pattern, and the very first risk signal in §8 of the pre-market piece read: "gap up then slump: if precious metals gap up more than 2% and then fall back to flat, treat it as a failed front-run". It predicted its own failure pattern and still put the position on.

2.2 Other "Watch Closely" Names

2.3 "Watch Only"

The 8 "watch only" names averaged +0.49% (vs. pre-market price −0.00%), with 5/8 beating the Shanghai Composite. ⚠️ But that positive return is entirely contributed by 海通发展 alone — excluding it, the other 7 averaged −0.86%, underperforming the Shanghai Composite by 0.56pp.

2.4 Pass List (the best-performing block in this report)

The 17 Pass names averaged −2.67%, with 14 underperforming the Shanghai Composite (82.4%) and 14 falling. This is the most valuable part of the report.

⚠️ But note the divergence inside the Pass group; it corroborates the very classification the pre-market piece wrote. The 5 names the pre-market piece explicitly labelled "direction-falsified" (白银有色 601212 loss-making, 集泰 002909 and 飞龙 002536 self-issued risk notices, 兆易 603986 and 中际旭创 300308 opposite to the overnight offshore direction) averaged −4.16% today; whereas among the names passed on grounds of "position too high / turnover overheated / no industrial logic", 海通发展 hit +9.99% limit-up and 国芳集团 rose +3.60% — the exact opposite direction. This is the second recurrence of the same conclusion recorded in memory as "chip-distribution metrics size positions, they do not set direction" — when the Pass reason is "position too high / turnover overheated", the stock can still limit-up tomorrow; only directional evidence such as "the company itself announced bad news" or "the business model runs the other way" genuinely predicts a decline. ⚠️ Note that 中际旭创 at +0.12% is itself a counterexample within those 5, showing that the reliability of the "opposite to the offshore direction" category is clearly lower than that of the "the company itself issued a risk notice" category. Next time these two should be split apart.

2.5 Hit-Rate Summary

The Shanghai Composite was −0.30%. The 16 buy-side names averaged −0.29% — a gap of 0.01pp versus the index.

2.6 One-Line Post-Mortem: What Went Wrong Today, and How to Fix the Pre-Market Piece

① The most expensive lesson: the sample source for the theme call was too narrow. §3 of the pre-market piece stated its coverage boundary explicitly — "candidates come from the union of five sources: the 9/3 gainers/losers board + the limit-up pool + the Dragon-Tiger List + overnight offshore mapping + in-window announcements; this is not a whole-market scan." That caliber is honest in itself, but it has a structural consequence: three of the five sources (gainers/losers board, limit-up pool, Dragon-Tiger List) all look at "what rose best yesterday", while today's theme was precisely the one that gave almost no signal yesterday. Agriculture had only 1 name among the 44 in the 9/3 limit-up pool (亚盛集团, first-day limit-up). How to fix it: make "the industry distribution of yesterday's limit-up pool" a mandatory question, and it must answer two things — (a) which industry does the highest consecutive-limit-up stock of the day belong to? (9/3 answer: general retail and publishing — and publishing is exactly the industry of today's 5-streak leader 龙版传媒) (b) is there any industry showing "rising first-board density but not yet a cluster"? The second question would have pointed to 亚盛集团 today.

② One in-window announcement was missed, and it was checked selectively within the same batch. Among the announcements released on the evening of 9/3, the pre-market piece checked 长电科技's placement (eiTime=19:54:02, written into §1 table item 7) but did not check 海通发展's placement (eiTime=18:39:07, no. 2026-089, proposed raise RMB 2 billion) — the latter was published 1 hour 15 minutes earlier than the former, so "the cutoff moment didn't cover it" cannot explain it. Today 海通发展 hit +9.99% limit-up, while its pre-market tag was "watch only". The root cause is the retrieval method: searching by theme keywords like "gold / silver / shipping" instead of pulling announcements name by name across the candidate pool. The irony is that 海通发展 was already in the pre-market piece's own candidate pool (§3 position 17, §7 branch two position 4); one call to the announcement endpoint would have shown it. How to fix it: for every stock in the candidate pool, pull in-window announcements one by one via the announcement endpoint, with no reliance on theme-keyword search. This is a mechanical action requiring no judgement, and it would have blocked today's biggest miss.

②b One more to write alongside it, because it is a mistake this report itself nearly made: "in-window" and "already priced" are two different things. The first draft also listed 中国船舶 (China State Shipbuilding, 600150) as an "in-window miss"; checking eiTime showed it was published at 9/2 19:02:48, a full day outside the pre-market window (East Money's notice_date says 9/3 and the announcement's signature line says September 3 — neither is the publication moment). But what is truly worth recording is its pricing rhythm: disclosed on the evening of 9/2 → the stock rose only 1.15% on 9/3, essentially no reaction → only on 9/4 did it rise +9.18% with main-force net inflow of RMB 2.038 billion (first in the market). In other words, even when a piece of news falls outside the scan window, its pricing may not be finished. Running a binary screen on "in-window / out-of-window" alone throws away this entire class of opportunity that starts on day 2. The pre-market scan should add a second dimension: besides "is there new news", also ask "which major contracts / tender wins / orders disclosed in the last 3 trading days have not yet been reflected in the share price?"

③ "The read was right but the position was wrong" — the single most important thing to remember from today. The pre-market piece said it plainly: "the whole day is a front-running window, not a confirmation window"; "on a Friday, with the added cost of carrying positions over the weekend, it is not advisable to trade as if a trend were established"; "if precious metals gap up more than 2% and then fall back to flat, treat it as a failed front-run". Those three sentences predicted today's action almost exactly, and yet 4 of the 5 recommendation slots were still precious metals. How to fix it: when a branch's own durability is self-assessed as "short (1–3 days), adjudicated tonight", that branch's cap in the recommendation slots should be 2 rather than 4. Promote "durability" from one plus/minus item on the scoring sheet to a hard constraint on slot allocation.

④ The places where the call was right but not acted on also deserve a note. §7 branch two of the pre-market piece wrote an excellent verification point: "the relative strength of 海通发展 (dry bulk) versus the three tanker names can reverse-engineer whether 9/3's driver was tankers or dry bulk". Today's answer was extremely clean — dry bulk (海通发展) +9.99% limit-up, while the three tanker names at −0.52%/+0.05%/+0.09% were all flat. The driver was dry bulk, not tankers. Yet the pre-market piece gave two recommend/watch slots to the tanker names 招商轮船 and 中远海能, and tagged the dry-bulk name 海通发展 "watch only". The quality of the verification-point design was clearly higher than the quality of the conclusion allocation — exactly consistent with the memory "verification points beat headline conclusions". Next time, write the verification point first and derive the conclusion from it, not the other way round.

⑤ Two wrongly-Passed names, with different reasons, to be recorded separately.

• 国芳集团 (601086) +3.60%: the pre-market Pass reason was "belongs to no branch with industrial logic, pure sentiment speculation". That judgement itself was not wrong — it did break its streak today (the 5-streak ended, spiking to +9.91% intraday before fading). What was wrong was equating "no industrial logic" directly with "will fall". This is a "position/logic type" Pass, which does not predict direction.
• 中际旭创 (300308) +0.12%: the pre-market Pass reason was "AVGO −2.74% overnight, direction unsupportive". Optical modules did not follow it down today (新易盛 Eoptolink 300502 +0.32%); what actually collapsed was server hardware (浪潮信息 000977 −9.99% limit-down, turnover RMB 13.708 billion) and semiconductors (−2.82%). The pre-market piece itself pointed out in §9③ that "using HDD to falsify DRAM/NAND is a sampling error", and then committed the same class of error again on optical modules: using an ASIC/networking-chip vendor (AVGO) to falsify optical modules. The same methodological error appeared twice in one report and only one instance was fixed.

---

3. Verifying Today's Leading Themes

3.1 Branch Reconciliation Table

3.2 Did the Pre-Market Piece Identify the Strongest Theme? No. Of today's three strongest, it wrote about none.

The top ten sectors (THS 90 industries) in order: livestock farming +4.36%, agricultural processing +3.60%, baijiu +2.87%, planting & forestry +2.48%, film & cinema +2.44%, cultural media +2.44%, beverage manufacturing +2.23%, food processing & manufacturing +1.91%, gaming +1.81%, travel & hotels +1.47%.

Nine of the ten are "agriculture + food & beverage + media/entertainment", and their number of appearances in the pre-market piece is zero. Today's ranking for the three branches the pre-market piece backed: precious metals 11th from the bottom (−1.78%), insurance +0.11%, software development +1.31% (11th).

3.3 ⚠️ Precious Metals: the Bullish Case Was Not Falsified — Positions Withdrew Ahead of Payrolls. This Distinction Decides How to Handle Monday

This is the passage that most needs to be spelled out, because the two explanations have opposite implications for Monday.

First the facts. Gold did not fall today:

Meanwhile, the A-share precious-metal names:

All eight gapped up, and all eight closed below their opening price. And gold was flat over that same stretch.

Conclusion: the money that sold precious-metal stocks today was not reacting to a falling gold price — it saw the same thing it saw in the morning. What it was selling was the combination of "weekend + payrolls". The August payrolls print lands tonight at 20:30 (Beijing) while A-shares close at 15:00, with a weekend in between that A-shares cannot react to; if the data comes in strong and pushes rate-hike expectations up, Monday simply gaps down with no chance to trim.

The practical implication of this distinction:

• If it were "the bullish case was falsified" (gold fell), then Monday should continue to avoid it.
• But the reality is "the bullish case still stands, the carrying cost was too high", so what Monday faces is a completely new situation with a known answer — the payrolls result will already be out and already fully priced by Friday's US gold miners (GDX/GLD). On Monday A-shares are "following an already-priced result", not "joining an unpriced bet".

⚠️ But this must be said at the same time: it is not a reason for this branch to turn around. Neither of the two medium-term judgements the pre-market piece made on primary-verified financials was overturned today, and both still stand: all five gold miners' Q2 profits fell sequentially by −17% to −39%; the SGE gold price's Q4-2025 base of RMB 949.9/gram vs the current RMB 957.50 is only +0.8% higher, so the Q4 year-on-year dividend goes to zero. "There may be a follow-through repricing on Monday" and "this line has no medium-term holding value" are true simultaneously — do not read the former as the opposite of the latter.

There is one faint but real positive signal inside the branch: 山东黄金 (600547) had main-force net inflow of RMB 240 million, the only respectable institutional buying in the branch, and it was the leader of the THS precious-metals sector today (+1.07%). The pre-market piece ranked it 2nd on the grounds of "the deepest decline among mid-to-large-cap gold miners, biggest expectation gap" — that intra-branch ordering was right; what was wrong was the weight given to the entire branch.

3.4 ⚠️ The Strongest Theme Missed Pre-Market: Big Agriculture — "Weak Signal", Not "Data in Hand and Unread"

This section is written more temperately than the 9/1 recap, because the nature really is different.

Of today's 39 limit-ups, agriculture and food accounted for 14 (35.9%), the only limit-up cluster that formed a family (the table below excludes the yellow-wine names 古越龙山 600059 and 会稽山 601579, which are counted in the consumption line to avoid double-counting with §3.1):

Was it visible pre-market? — Partly, but the signal really was weak. Checking the 9/3 limit-up pool the pre-market piece already had in hand (44 names): agriculture had only 1 name, 亚盛集团 (600108), and it was a first-day limit-up. On 9/2 there were 2. This differs from the 8/31 instance (3 agriculture names among 88 limit-ups, including the highest consecutive-limit-up name of the day), so the 9/1 recap's self-criticism of "data in hand, conclusion not drawn" cannot be applied here.

⚠️ The real point of failure lies elsewhere, and it is more insidious: "the agriculture ebb is confirmed", written in the 9/2 recap, was treated as a permanent conclusion. The chain of facts runs: agriculture erupted on 9/1 (planting & forestry +7.51%, first in the market) → collapsed on 9/2 (−4.41%, last of the 90 sectors, 0 names in the limit-up pool) → the 9/2 recap accordingly wrote "the agriculture ebb is confirmed" → on 9/3 亚盛集团 quietly took a first board (a signal, unnoticed) → on 9/4 a full second wave. A theme restarting two days after ebbing is the norm for A-share themes, and the label "ebb confirmed" meant neither of the following two pre-market pieces looked back at it even once.

How to fix it (down to the action): for any branch that has shown a "sector daily gain > 5%" within the last 5 trading days, even if it has been judged to have ebbed, the pre-market piece must keep one line of "re-check", with exactly one check item — does it still have a first board in yesterday's limit-up pool? Today that line would have produced "亚盛集团 first board on 9/3", and that is the first footprint of a second wave.

Drivers: what the searches returned were cross-week narratives, not same-day events. The three on the institutional side are "expectations of a hog-cycle reversal (H1 net loss attributable to parent for the hog-farming sector was RMB 20.942 billion, three consecutive quarters of deep losses, and declining productive biological assets indicating active capacity reduction)", "international grain prices bottoming and rebounding", and "a triple resonance of geopolitical conflict + extreme weather + policy control". ⚠️ These are all medium-term narratives; this report could not anchor any specific trigger event on the evening of 9/3 or the morning of 9/4. Per the iron rules, written as "no same-day catalyst obtained", not "no catalyst".

3.5 ⚠️ The Second Miss: Media — and the Highest-Streak Leader Is Right Here

Cultural media +2.44% (78 up, 7 down), film & cinema +2.44% (19 up, 1 down), gaming +1.81%; the media sector's main-force net inflow of RMB 6.174 billion was first in the market on East Money's caliber (advertising & marketing +RMB 3.862 billion, marketing agencies +RMB 3.772 billion).

More to the point: today's highest streak of 5, 龙版传媒 (605577, Shanghai main board, publishing), sits on this line, and it was already at 4 boards on 9/3. In the 9/3 limit-up pool the pre-market piece had in hand, the publishing industry had exactly 1 name — this one — and it had the second-highest streak that day. This differs from agriculture: it was clearly visible in the pre-market data; nobody just looked at the industry distribution of the limit-up pool.

The media line's limit-up / near-limit-up ladder today: 龙版传媒 (605577) 5 boards, 中国出版 (China Publishing, 601949) first board, 欢瑞世纪 (Huanrui Century, 000892) +10.04% (20 breaks off the limit, turnover rate 43.9%), 易点天下 (Yidian Tianxia, 301171) +11.12% (RMB 6.314 bn), 天娱数科 (Tianyu Digital Science, 002354) +8.75% (21 breaks off the limit).

⚠️ But mind the quality: this line's break-off counts are extremely high (欢瑞 20 times, 天娱 21 times), making it a high-turnover speculative move rather than an institutional one. Net inflows are large, but the chip distribution is unstable.

3.6 ⚠️ The Third Miss: Shipbuilding — the News Was Outside the Window, but the Pricing Was Not Finished

中国船舶 (600150, Shanghai main board) +9.18%, turnover RMB 12.393 billion, main-force net inflow RMB 2.038 billion (first among all stocks), broke off the limit 7 times intraday without sealing.

The catalyst is clear, but it was not inside the pre-market window: the announcement was published at 2026-09-02 19:02:48 (eiTime, art_code AN202609021828932369/70), whereas the pre-market window was 9/3 15:00 → 9/4 07:00. ⚠️ East Money's notice_date says 2026-09-03 and the announcement's signature line says "September 3, 2026" — neither is the publication moment — the former is the next trading day, the latter the board-meeting date; only eiTime is. Content: controlled subsidiary Guangzhou Shipyard International, together with CSSC Trading, signed with a shipowner on September 1, 2026 newbuilding contracts for 10 units of 8,200-car LNG dual-fuel PCTCs, with a total contract value exceeding US$1 billion, payable in US dollars, with delivery expected in 2029–2031. On the same day there was also a separate announcement on a subsidiary's arbitration outcome.

⚠️ So the lesson here is not "an in-window announcement was missed", but the harder-to-guard-against kind: the news fell outside the window while the pricing had not finished. The timeline: disclosed 9/2 19:02 → 9/3 the share price +1.15% (essentially no reaction) → 9/4 +9.18% with main-force net inflow first in the market. If the pre-market process only runs the one binary screen of "is there new news in the window", this opportunity is lost on both 9/3 and 9/4 — on 9/3 because it was in-window but the price hadn't moved (looking like "the market doesn't buy it"), and on 9/4 because it had left the window (looking like "old news"). The fix is very concrete: besides scanning for in-window news, add one more question — which major contracts / tender wins / orders disclosed in the last 3 trading days have moved the share price by less than 2%? That question would have hit 中国船舶 squarely today.

⚠️ The caliber the announcement states about itself must be given at the same time, or this catalyst will be overestimated: the original text says explicitly that "because the contract performance period is long, the company will recognise revenue in accordance with the revenue standard, and this will not have a material impact on the company's current-period profit". With delivery in 2029–2031, this is a positive for order visibility, not for current-period earnings. Today's +9.18% reflects the order signal, and the 7 breaks off the limit show the market is hesitant about that time span too.

---

4. Limit-Up Ladder / Streak Heights

4.1 Heights and Promotions

Ladder assessment: the height held at 5 boards but the structure is fragile. The 2-board tier has only 6 names, and among them 海通发展 broke off the limit 9 times and 百大集团 only sealed late in the session; the genuinely clean ones are just three small caps — 爱仕达, 新炬网络 and 恒盛能源. Yesterday's 5-board leader 国芳集团 breaking its streak, plus today's 5-board leader having a seal-to-turnover ratio of only 0.030, means the money taking the baton is thinning out.

4.2 Strongest Seals (seal-to-turnover caliber, excluding one-price limit boards)

⚠️ 楚天龙 (003040) is an outlier worth looking at on its own today: the communications-equipment sector was −1.64% (21 up, 67 down), yet it was the sector's top gainer and sealed at the limit on a one-price board with a RMB 543 million sealing order (the largest of the day), with a limit-up statistic of 11/7. A single strong seal inside a broadly weakening track is usually a stock-level event rather than sector-driven. This report did not obtain its catalyst; listed as a must-check for tomorrow morning.

4.3 Limit-Downs and Failed Limit-Ups

8 limit-downs, of which the most notable is 浪潮信息 (000977, Shenzhen main board) at −9.99% on turnover of RMB 13.708 billion — the largest turnover among all limit-down names in the market today, and far larger than the other seven combined.

On 浪潮信息: this report did not obtain a same-day attribution, stated honestly. Verified facts: the company's most recent announcement is the 8/29 interim report, and there were no new announcements on 9/3–9/4; three rounds of searching returned no same-day news. The known price context is — a limit-up on 8/31 (+9.99%) to RMB 85.18, three days of range-trading between RMB 85 and 86 on 9/1–9/3, and a straight limit-down to RMB 77.72 today, i.e. a single-day break after high-level consolidation. Main-force net outflow RMB 1.826 billion. Per the iron rules, this report writes "catalyst not obtained", not "no news" — the memory entry "not found ≠ does not exist" was validated repeatedly in exactly this kind of situation.

48 failed limit-ups is a recent high, and they include some big names: 中国船舶 (600150) broke off the limit 7 times and closed +9.18%, 景旺电子 (Kinwong Electronic, 603228) broke off 3 times and closed +8.96% (RMB 7.414 bn), 四方精创 (300468) broke off 13 times and closed +14.58%, 天娱数科 (002354) broke off 21 times and closed +8.75%. These heavily-traded names that "touched the limit but could not hold it" are the truest gauge of today's sentiment: people are willing to push it up there, but nobody is willing to take it off their hands up there.

---

5. Fund Flows

5.1 Northbound Flows

No data. Since 2024-08-19 the exchanges have stopped disclosing daily net buying by northbound funds, so this report gives no northbound figures. (⚠️ If you see a statement elsewhere along the lines of "northbound net inflow of RMB X billion", check its caliber — the commonly cited figures around RMB 42 billion are a quota ceiling, not an inflow amount.)

5.2 Sector Main-Force Net Inflows / Outflows (East Money caliber)

⚠️ Caliber reminder: East Money's sectors span level-1/2/3 tiers, so listing them together double-counts (e.g. "electronics" contains "semiconductors", and "semiconductors" contains "digital chip design"). Read by tier; do not sum down the column.

Top net inflows:

Top net outflows:

⚠️ Nonferrous metals saw a net outflow of RMB 6.859 billion — the most direct flow-side rejection of the pre-market piece's first branch. Combined with §3.3: gold did not fall, but RMB 6.859 billion of the money that prices it walked away.

Sector net inflows on the THS caliber (a different caliber; do not subtract against the table above): baijiu +RMB 3.291 bn, gaming +RMB 2.685 bn, software development +RMB 1.953 bn, livestock farming +RMB 1.765 bn, cultural media +RMB 1.719 bn; semiconductors −RMB 15.311 bn, components −RMB 5.889 bn, optoelectronics −RMB 5.326 bn, industrial metals −RMB 5.066 bn, communications equipment −RMB 4.819 bn. The two calibers agree completely on direction — money moved out of tech hardware and into agriculture, consumption and media.

5.3 Per-Stock Main-Force Net (East Money caliber)

One counterintuitive but important detail: 招商轮船 (601872) had a main-force net inflow of RMB 155 million yet closed down 0.52%, and 中金黄金 with main force at +RMB 94 million also closed down. Main force buying while the price falls means the sellers are dispersed retail and hot-money selling pressure. This pattern is usually not the end of a trend but an exchange of chips — it can be listed as a Monday watch item, but one day of data is not enough to conclude from (see the experience entry "single-day fund-flow extremes mean-revert").

5.4 Dragon-Tiger List

⚠️ This report did not obtain the 9/4 Dragon-Tiger List. The exchanges publish it after 18:00 on the day, and this report was generated at 15:30, so it is unobtainable at that moment. This is a scheduling limit rather than a data gap, and tomorrow morning's pre-market piece can fill it in completely.

---

6. Next-Session Outlook (next trading day: Monday 2026-09-07)

⚠️ Today is Friday and the next trading day is 2026-09-07 (Monday). In between sit tonight's US August nonfarm payrolls at 20:30 and an entire Friday US session, plus a full weekend. Every judgement in this section assumes that by Monday's open, payrolls is already known information.

① Theme Durability Assessment

② Monday's Focus (branch + stock + verification point)

③ Monday Avoid List

1. ⛔ Semiconductors / memory / consumer electronics / components — semiconductors 13 up 172 down, net outflow RMB 15.311 bn; level-1 electronics −RMB 27.224 bn. Multiple consecutive days of one-way outflows, and today accelerated rather than converged.
2. ⛔ Servers / compute hardware (newly added) — 浪潮信息 (000977) broke down from a high with a limit-down, RMB 13.7 bn turnover, main force −RMB 1.826 bn. ⚠️ This report did not obtain its attribution, which is all the more reason to avoid it: a large-cap hitting limit-down with no visible news means the sellers know more than we do.
3. ⛔ Yesterday's elevated names and today's high-turnover failed limit-ups — 集泰股份 (002909) limit-down, 飞龙股份 (002536) −7.57%, 思泉新材 (301489) −6.70%; the liquid-cooling/thermal-management ladder has completed a round of clearing. Today's newly added high-risk list is the names with more than 20 breaks off the limit and turnover rates >30%: 欢瑞世纪 (000892, turnover 43.9%), 天娱数科 (002354, turnover 38.1%), 敦煌种业 (600354, turnover 27.9%).
4. ⛔ Lone-straggler agriculture laggards — among today's 15 agriculture limit-ups, those with a seal-order/turnover ratio below 0.02 are 敦煌种业 (0.001), 安记食品 (0.002), 新五丰 (0.003) and 古越龙山 (0.000); these are the ones that only barely sealed at the close after ten to twenty-odd breaks off the limit. The theme can be traded, but not through these names.
5. ⚠️ Do not bottom-fish just because "precious metals fell for a day" — see §3.3: today's decline was pre-event de-risking, and neither of the two medium-term problems — Q2 profits down sequentially across the board, and the Q4 year-on-year gold-price dividend going to zero — has been resolved. A follow-through rebound ≠ something you can hold.

④ Inputs for Monday's Pre-Market List (this section is the reason this report exists)

1. ⚠️⚠️ [Top priority, mechanical action] For every stock in the candidate pool, pull in-window announcements one by one via the announcement endpoint, with no further reliance on theme-keyword search. Today's biggest miss was 海通发展 (+9.99% limit-up), whose placement plan at eiTime=2026-09-03 18:39:07 was inside the window, while that same evening the pre-market piece checked 长电科技 (19:54:02), published 1 hour 15 minutes later in the same batch. This is a double standard in executing the criterion, not a capability problem. 1b. ⚠️⚠️ [Equal priority, and harder to guard against] Add a second question: which major contracts / tender wins / orders disclosed in the last 3 trading days have moved the share price by less than 2%? 中国船舶's announcement (+9.18%, main-force net inflow first in the market) was outside the window (9/2 19:02), it rose only 1.15% on 9/3, and the pricing was pushed to day 2. A purely binary "in-window / out-of-window" screen loses this kind of opportunity on both days. One hard data-fetch rule to go with it: for the publication moment of an announcement, only eiTime counts; East Money's notice_date (next trading day) and the announcement's signature date (board-meeting date) are not the publication moment — the first draft of this report wrote an unfounded criticism based on the latter two.
2. ⚠️⚠️ [Mandatory question] Promote "the industry distribution of yesterday's limit-up pool" into two mandatory pre-market questions: (a) which industry does the highest consecutive-limit-up stock belong to? (the 9/3 answer was "general retail, publishing", and publishing is exactly where today's strongest leader sits); (b) is any industry showing "rising first-board density but not yet a cluster"? (the 9/3 answer would point to 亚盛集团 / planting).
3. ⚠️ [New rule] "Ebb confirmed" is not a permanent label. For any branch that has shown a "sector daily gain >5%" within the last 5 trading days, even if it has been judged to have ebbed, the pre-market piece must still keep one line of re-check, with exactly one check item: does it still have a first board in yesterday's limit-up pool?
4. ⚠️ [Hard slot constraint] When a branch's "durability" is self-assessed as "short (1–3 days), adjudicated that same night", that branch's cap within the 5 recommendation slots is 2. Today it was 4, and the pre-market piece itself had spelled the risk out in full. Promote "durability" from one plus/minus item on the scoring sheet to a hard constraint on slot allocation.
5. ⚠️ [Method] Write the verification point first, then derive the conclusion from it. Today's best single judgement was the pre-market piece's §7 line that "the relative strength of 海通发展 versus the three tanker names can reverse-engineer whether the driver is tankers or dry bulk" — it produced a clean answer (dry bulk), and yet the positions allocated on the basis of the conclusion were entirely on the wrong side.
6. Must-check items (carried over from the pre-market piece, unclosed for consecutive sessions): ① BDTI / VLCC TD3C freight rates and their historical percentiles (not obtained for the second consecutive report; until obtained, tankers may not occupy a recommendation slot); ② the attribution for 浪潮信息 (000977)'s limit-down; ③ the stock-level catalyst for 楚天龙 (003040)'s one-price limit board (an outlier inside a communications-equipment sector at −1.64%, with the largest sealing order of the day at RMB 543 mn); ④ today's 9/4 Dragon-Tiger List (published after 18:00, can be filled in tomorrow morning); ⑤ the use of proceeds for 海通发展's RMB 2 billion placement.
7. Caliber reminder: tonight's 20:30 payrolls is the only macro variable, and A-shares cannot react to it until Monday. Friday's US session will price it in first, so tomorrow morning's pre-market piece must map from Friday's US close (not Thursday's) — precisely the same pitfall of "the 9/2 session vs the 9/3 session" that this Wednesday's pre-market piece stepped in.

---

⚠️ Risk warning: this recap is a post-close information review and observation only, and does not constitute investment advice. Data may differ in timeliness or caliber; please refer to exchange/company disclosures, and do not use this directly as a basis for trading.