• 我们自己写下的两个作废条件本周全部触发，原文一个字没删；指数两创新高，而十七只里七只是跌的。
• 全表重扫后不是三座，是三周里铺了十二座、约 400 万张，主体位置从跌 10.7% 一路走到跌 38.3%。这些全是当天就看得见的。
• 而周五的 13F 讲的是六周前的事。同一个问题两种答法：7 月 24 日挂在链上的那个价，要等到 11 月 14 日才进申报文件，中间隔 113 天。

这一期讲什么

第 1 到第 3 节是本周的三个亮点：说错的两条怎么认账、新高背后十七只里七只在跌、以及三周铺出来的十二座价目表。
第 4 节讲周五的 13F 为什么是旧闻，第 5 到第 7 节讲同一个问题在期权链上怎么当天就有答案，以及你手里有票的话这套东西怎么用。
不着急的话，读完第 3 节和第 5 节就有数了。

〔对账〕被推翻的那两条，我们一个字没删

周三通胀日，开奖前挂出的围栏兜住了，我们写了一句：市场提前把这场开奖标对了价。同一段里也挂了作废条件：如果周四生产者价格落地时同样的围栏被冲破，那这次就只是碰对，不是标对。

第二天它就被冲破了。指数盘中越过上沿 2.62 点，收盘也留在栏外。按我们自己写下的规矩，那句判断当场降级，原文一个字没改。

横截面那条同样被触发。周三我们记到十六只高开、十四只被从开盘一路卖回去，写了句这是开奖日当天的形状，作废条件是第二天盘中若普遍转正，这个回落就只属于那一天。第二天十七只里十三只盘中转正，条件成立。

四组开奖对上，两条挂出去没回来验，照记为欠账。能被推翻的判断才叫判断。

〔横截面〕指数两创新高，十七只里七只在跌

这一周标普 500 基金涨 0.40%，纳指 100 基金涨 1.11%，周三周四各创一次历史新高。同一周，我们观察的十七只里有七只是跌的。

两头拉得很开：闪迪涨 35.38%、美光涨 10.72%、AMD 涨 6.42% 在一头；博通跌 8.13%、亚马逊跌 4.31% 在另一头。首尾相差四十三个百分点。指数那 0.40%，是这些东西相互抵消之后剩下的残值。

还有一个数比涨跌幅更值得看：振幅，一周里最高价和最低价之间的距离。英特尔这一周涨了 0.84%，几乎等于没动，但它的振幅是 11.78%。也就是说它在一个将近十二个点的区间里来回走了一趟，最后回到原地附近。拿着它的人，中间见过相差一成多的两个价格。

涨跌幅是位移，振幅是这只票真实走过的路。同一只票同一周，这两个数可以差十倍。

顺带一句成本侧：一个月内的波动率从 15.46 降到 14.25，降了 1.21 个点；而一年期从 22.76 到 22.75，一周只动了 0.01。近端是情绪，远端是成本，这一周松的只有情绪。

〔纵深〕不是三座，是铺了三周的十二座

本周日更里写过三座同型结构。周末我们把全表重扫了一遍，按同样的公开判据：同一天同一个到期、三个等距的看跌价位、张数一比二比一、最小那条腿一万张以上。

结果是十二座，从 07.24 一直排到 08.13，合计约 400 万张，其中十一座在标普 500 基金上，一座在纳指 100 上。

连起来才看得见两条形状。一是同一个到期被反复搭：11.20 那个到期被搭了三次，08.03、08.04、08.12 各一次，而且每次整体上移 10 点，跟着指数往上挪。二是越铺越深：主体距当日指数的距离，从最早那座的 10.7%，一路走到最新那座的 38.3%。

已经被再次采到的五座全部验出真沉淀；其余七座建仓后没再被采到，标为不知道，不猜。

红线不变：不写这是什么结构、不猜谁在做、不推方向。我们能说的只有形状和沉淀。

但有一件事这十二座已经说清楚了：三周里，有人一次又一次把如果跌、要在哪个位置拿到赔付，写在了明面上。每一次都是当天就能看见的。

记住这句，下一节马上要用到。

〔13F〕新闻是新的，交易是旧的

上一节那十二座，最要紧的一点是：它们全都是当天就能看见的。 建仓那天成交摆在那儿，第二天早上的持仓给出答案，谁都能查。

而这一周还有另一类信息刚刚露面，性质完全相反。

周五是二季度 13F 报告的申报截止日。你这两天大概率刷到了一批某某清仓某某的标题。

13F 是季末持仓快照。二季度截到 6 月 30 日，申报期限 45 天，正好是周五。所以那些文件是这两天才出现的，里面描述的交易发生在六周以前。

它还有两件事不告诉你：不告诉你什么时候买的，也不告诉你用什么价买的。 你只能看见季末那一刻手里剩下什么。

拿一份描述 6 月底持仓的文件去解释这两天任何一只票的涨跌，是把申报日期当成了交易日期。

〔另一条路〕同一个问题，期权链当天就答了

有一个地方不用等六周：期权链。因为一份合约上写着的，正是 13F 缺的那两样东西：什么价，什么时候之前。

举个我们表里记到的例子。2026 年 7 月 24 日，长天期异动表多了一行：SpaceX 12.18 到期、行权价 115 的看跌合约，当天成交 1109 张，进场持仓 7447 张。第二个交易日隔夜持仓变成 8398 张，净增 951 张，留下了当天成交的 85.8%，按我们固定的判据属于真建仓。

那天我们不知道是谁。一个月后，段永平公开说了这笔交易：卖了 1000 张，每股收 23.26 美元。我们量出 951，他自己报 1000，对上了。

这笔账真正的关键，不是他收了 232.6 万，是他的接货成本不是 115，是 91.74。 因为 23.26 已经先收进来了。整笔交易就是一句话：现在的价我不急着买，你付我每股 23.26，未来五个月你愿意的话我按 115 接。

12 天后财报落地，SpaceX 收 108.27、当天跌 13.61%、盘中最低到过 106.66。他没等着被行权，直接买了 10 万股。说的和做的是同一件事。

现在回头算日子：这个价 7 月 24 日就挂在链上，谁都看得见；而它要出现在申报文件里，得等三季度报告，也就是 11 月 14 日。中间隔 113 天。

图上那条线就是这件事的形状：选定的那个价是竖线，到期时没越过它，收的钱全归你；越过了，那一段就跟你没关系，线从那里开始就平了。

〔怎么用〕你手里有票的话，这套东西怎么读

量级跟你无关，方法跟你有关。搬到更多人手里真有的东西上，比如纳指 100。同一只票同一个到期日，只换行权价，08 月 14 日收盘后的真实报价是这样：

贴着现价卖，一个月收 2.19%，但大约一半机会要交货；挪到高 7.4% 的地方安全多了，只收 0.26%。安全和租金是同一根拉杆的两头，你不可能只往一边拉。

所以没有哪一档更划算。只有一个问题：你愿不愿意在那个价成交。 愿意就是那一档，不愿意，租金再高也跟你无关。这就是全部的选法，也正是前一节那笔交易在做的事。

反过来卖看跌是同一张表倒着看：0.30 对应 711（低 2.7%）收 1.24%，0.10 对应 666（低 8.9%）收 0.38%。一个细节：越往下的看跌，隐含波动率越高，从 20.5 升到 25.4。这是常见的偏斜，不是数据错了，市场对下跌的定价从来比对上涨贵。

那几个吓人的词，一句话就够。 隐含波动率，市场猜它未来会晃多厉害，像车险保费看你开车猛不猛；分位，今天这个价钱在它自己过去一年里排第几；Delta，到期时被成交的大致概率，0.30 就是大约三成机会真要交货；Theta，你收的那笔钱靠时间一天天变成你的，像冰块放在桌上。这几个词都不预测涨跌，只描述价钱和概率。

日期上，三十到四十五天是常见区间，更要紧的是看窗口里有没有它的财报。举我们表里的例子：09 月 18 日到期这个窗口，九只观察票里只有英伟达要在窗口内交卷，其余的都排在 10 月底以后。

〔收手〕四种结局，开仓那一刻就全在桌上了

提前平仓：租金赚到七八成就买回来结束，剩下那点钱要拿最后几天的风险去换。
放到作废：到期没越过行权价，合约归零，收的钱全落袋，票还在手上。
被行权：价格越过行权价，票按那个价交出去，或者按那个价接进来。
往外滚：买回这一张，同时卖一张更远更晚的，代价是钱压得更久。

四种里有三种发生在到期之前，所以这不是卖完就只能干等的事。

关于被行权要多说一句：它不是失败，是你当初写下的那个价真的到了。 真正的失败是另一种：在一个你根本不愿意成交的价上收了钱，价格真到了才发现自己不想接。

所以整件事的顺序是：先想清楚你愿意在什么价成交，再去看市场给这个价开了多少钱。 反过来做的人，是先看见租金高，再回头说服自己愿意那个价。

⚠️ 期权卖方承担被行权的义务，美式期权可提前行权。本节写的是怎么读这几个数，不是该买什么。

〔下周看点〕本周记的账，大多数在下周五开奖

• 周二 08.18：Reddit 开盘前正式进入标普 500。公告日的跳空 11.17 个百分点已经记下，生效日的跳空到时候对照，这是两段完全不同的定价。
• 周三 08.19：波动率期货换月日，曲线形状要连着滚动效应一起读。
• 周五 08.21：月度期权到期日，本周记录在案的合约绝大多数在那天开奖，逐案收卷。

这篇适合转给谁

• 手里拿着指数基金、但从来没卖过 covered call 的人：第 5 到第 7 节就是给他的，五分钟能看完，看完至少知道自己在决定什么。
• 每次刷到某某清仓某某就跟着紧张的人：第 4 节告诉他那些文件描述的是六周前的事。
• 问过你大资金到底在干嘛的人：第 3 节那十二座就是答案的一部分，而且是当天就摆在明面上的。

今天值得带走的三句话

• 申报文件告诉你他们买了什么，期权链告诉你他们愿意用什么价买。前者要等 113 天，后者当天就写着。
• 一笔交易真正的成本不是行权价，是行权价减去先收到的那笔钱。段永平那笔是 115 减 23.26 等于 91.74。
• 安全和租金是同一根拉杆的两头。所以没有哪一档更划算，只有一个问题：你愿不愿意在那个价成交。

🌍 English edition

⚡ Three-Line Summary

• Both voiding conditions we published this week fired, and we deleted nothing. The index set two records and seven of seventeen names still fell.
• A full rescan found not three builds but twelve, across three weeks, roughly four million contracts, with the payout centre walking from 10.7% below the index to 38.3% below. Every one of them was visible the day it happened.
• Friday's 13F describes trades from six weeks ago. Same question, two answers: a price posted on the chain on July 24 does not reach a filing until November 14, one hundred and thirteen days later.

What This Issue Covers

Sections 1 to 3 are the week's three highlights: how we booked the calls we got wrong, seven of seventeen falling under record highs, and the twelve price lists laid down over three weeks. Section 4 is why Friday's 13F is old news. Sections 5 to 7 are how the same question is answered on the chain the same day, and how to read it if you already hold the stock.

[Scoreboard] Two Calls Overturned. Nothing Deleted.

On CPI day the pre-posted fence held and we wrote that the market had priced the event correctly. In the same section we published the condition that would void it: if the same fence broke on PPI day, that was a lucky hit rather than a correct price.

It broke the next day. SPY cleared the upper edge by 2.62 points and closed outside. By our own rule the reading was downgraded that day, and the text stayed.

The cross-section call fired the same way. We had written that Wednesday's selling was an event-day shape, void if intraday turned broadly positive the next day. Thirteen of seventeen printed positive intraday.

Four groups settled as written; two published checks went unkept and are logged as debts. A call that cannot be overturned is not a call.

[Cross Section] Two Records Above, Seven Names Falling Below

SPY rose 0.40% for the week and QQQ 1.11%, with records on Wednesday and Thursday. Seven of our seventeen names still fell.

The ends are far apart: SanDisk +35.38%, Micron +10.72%, AMD +6.42% at one end; Broadcom −8.13%, Amazon −4.31% at the other. Forty-three points between them. That 0.40% is what remains after all of it cancels out.

One number matters more than the change: the weekly range. Intel finished up 0.84%, almost still, on a range of 11.78%. It walked a twelve-point corridor and came back near where it started. Anyone holding it saw two prices more than a tenth apart.

The change is displacement. The range is the distance actually travelled. For one name in one week the two can differ tenfold.

On the cost side: front volatility fell from 15.46 to 14.25, down 1.21 points, while the one-year went 22.76 to 22.75, a move of 0.01. The front is mood, the far end is cost. Only the mood eased.

[Depth] Not Three Builds. Twelve, Over Three Weeks.

We wrote about three identical structures in this week's dailies. Over the weekend we rescanned the full table on the same published test: one day, one expiry, three evenly spaced put strikes, sized one to two to one, smallest leg above ten thousand contracts.

The answer is twelve, from July 24 to August 13, roughly four million contracts, eleven on SPY and one on QQQ.

Two shapes only appear once you line them up. The November 20 expiry was built three times, on August 3, 4 and 12, each one shifted 10 points higher as the index rose. And they went deeper: the payout centre moved from 10.7% below the index to 38.3% below.

Five have been resampled and all verified. The other seven were never resampled and are marked unknown.

The limits stand: we do not name the structure, guess who did it, or infer direction. Shape and what stayed overnight are all we can speak to.

But one thing these twelve already settle: over three weeks, somebody repeatedly wrote down, in the open, the position at which they wanted to be paid if it fell. Every one of those was visible the same day.

[13F] The News Is New. The Trades Are Not.

That is the whole point of the previous section: those twelve were visible on the day. The volume sits there when it is built, and the next morning's open interest answers whether it was real. Anyone can check.

This week also produced the opposite kind of information.

Friday was the second-quarter 13F deadline, which is why your feed filled with headlines about who exited what.

A 13F is a quarter-end snapshot. Q2 closed on June 30 and the filing window is 45 days, landing on Friday. The documents are new; the trades inside them are six weeks old.

Two more things it does not tell you: when the position was opened, and at what price. You only see what was left at quarter end.

[The Other Route] The Chain Answered on the Day

One place needs no six-week wait. A contract states exactly what a 13F omits: what price, and by when.

On July 24 our long-dated screen logged a SpaceX put, December 18 expiry, strike 115: 1109 contracts traded against 7447 standing. The next settlement showed 8398, up 951, keeping 85.8% of that day's volume, which by our fixed test is a real build.

We did not know whose it was. A month later Duan Yongping said publicly that he had sold 1000 of them at 23.26 per share. Our ruler read 951. His own number was 1000.

The part most people get wrong is not the 2.326 million dollars received. It is that his cost to take delivery is not 115. It is 91.74, because the 23.26 came in first.

Twelve days later SpaceX closed at 108.27, down 13.61% on earnings day, with an intraday low of 106.66. He did not wait to be assigned. He bought 100,000 shares outright. What he said and what he did were the same thing.

That price was visible on the chain on July 24. It will not appear in a filing until November 14. One hundred and thirteen days apart.

[How to Read It] If You Already Hold the Stock

Scale is not the point; the method is. Same underlying, same expiry, only the strike changes. Real quotes on the Nasdaq 100 fund after the August 14 close: at a 0.50 chance of assignment, strike 735, 0.5% above spot, one month of rent 2.19%. At 0.30, strike 756, 3.4% above, 0.99%. At 0.19, strike 770, 5.3% above, 0.53%. At 0.10, strike 785, 7.4% above, 0.26%.

Safety and rent are two ends of one lever. You cannot pull only one.

So no strike is the better deal. There is only one question: would you actually be happy to transact at that price. If yes, that is your strike. If no, the rent does not matter.

Four exits exist, and all four are on the table the moment you open: close early once most of the rent is earned; let it expire worthless and keep the stock; be assigned at the price you wrote down; or roll further out. Three of the four happen before expiry.

Assignment is not failure. It is the price you chose arriving. The real failure is collecting money at a price you never wanted, and finding out only when it gets there.

⚠️ An options seller carries the obligation to be assigned, and US-style options can be assigned early. This section is about reading the numbers, not about what to buy.

[Next Week] Most of This Week's Files Settle on Friday

Reddit joins the S&P 500 before Tuesday's open. The announcement-day gap of 11.17 points is already on record; the effective-day gap is its counterpart, and the two are priced in completely different ways.

Wednesday is the volatility futures roll, so read the curve's shape that week together with the rolling effect.

Friday is the monthly expiry, when most of the contracts logged this week settle at once.

Three Lines Worth Keeping

• A filing tells you what they bought. The chain tells you what price they were willing to pay. The first takes 113 days; the second is there the same day.
• The real cost of that trade is not the strike. It is the strike minus what came in first: 115 less 23.26 is 91.74.
• Safety and rent are two ends of one lever. No strike is the better deal; there is only whether you would be happy to transact there.

一 · 本周回顾：挂出去的账，回来对

一句话：事前写下的区间，价值不在于它猜对了，在于它写下之后不能改。

闪迪的两条带都兜住了。 08.04 收盘时记录的区间是 1219.55 到 1588.79；08.05 收盘时市场把带子收窄并下移，记成 1209.723 到 1491.277。财报之后，08.06 收在 1258.58，两条带都装得下它，距 08.04 那条的下沿 39.03 美元，距 08.05 那条 48.86 美元。

值得记的是两条带的变化方向：随着股价从 1425 走到 1350 进入财报，市场把带子收窄了，也把整条带往下挪了。它不是一直开着同一个价在等，它每天都在改，而每次改都留了痕。

SPCX 那条更极端。 08.04 收盘给出的三日区间是 101.80 到 135.97。第二天它跌了 13.61%，日内最低 106.66、最高 117.50，全天都在那个区间里面。跌这件事本身，前一天就已经被算进价钱里了。

闪迪 1400 那一档的钱，两侧都没留下。 08.05 成交的合约里，看涨侧沉淀 7.1%、看跌侧 9.7%，两边都是九成以上当天开当天平。那批钱不是压着过夜的仓位，是当天来当天走的。

记分牌一句话：这一周押的全是范围和去向，一次方向都没押，五条全开了奖。

二 · 市场自己定的范围，也回来对了

一句话：不只我们挂的判断要回来对，市场自己用钱定的范围，同样可以对。

08.06 那天，市场给 17 只观察标的各定了一个到 08.07 到期前的波动范围。08.07 收盘开奖：

17 只里 14 只留在范围内，3 只走了出去（SPCX、微软、谷歌A）。

两头都很极端：AMD 只走了定价范围的 6%：给它定了 ±21.56，它动了 1.31；而 SPCX 走了 293%。闪迪最惊险，定 ±140.78 走了 138.29，擦着边没出去。

口径：范围是 08.06 盘前定的，上下沿锚 08.05 收盘、覆盖到 08.07 到期；开奖取 08.07 收盘。

三 · 你惦记的存储，看着一周没动，其实走过了千山万水

一句话：你惦记的那几只存储，一周下来涨跌多少，和这一周晃了多大，是两件事。

十七只观察标的，这一周 14 涨 3 跌。但把周振幅（周内最高减周内最低，除以上周五收盘）和一周净涨跌并排放，落差立刻出来：

前四名全是半导体与存储，而大盘只有 3.75%，晃得最狠的和最稳的差七倍。

闪迪是最极端的那个：周内最高到过 1446.62、最低跌到 1121.27，差 26.78%，而一周下来只动了 0.22%。逐日看是 +6.03、+10.84、−5.40、−6.81、−3.68，先冲上去，再一路还回来。

SPCX 是同样的幅度，完全不同的结果：振幅 26.44% 跟闪迪几乎一样，但它一周净涨 22.83%，晃出去了，没回来。

同样是两成六，一个是白折腾，一个是真涨。

反过来，英伟达周振幅 13.90% 而净涨 11.56%，它这一周走的每一步几乎都朝着同一个方向。

这也是为什么周线图会骗人。 它只画起点和终点，中间那 26.78 个百分点的往返，在图上一个像素都不占。

拿住闪迪的人，中间见过相差四分之一的两个价格。

这张表对拿着票的人意味着什么。 振幅不是一个抽象数字，是账户里真实出现过的两个余额：拿着闪迪的人，这一周账面见过 1446 也见过 1121，最后停在原地；拿着 SPCX 的人见过同样大的摆动，但终点在高处。两件事的差别不在幅度，在筹码有没有换到新的共识价位。闪迪从 1121 到 1446 都有人真金白银成交，说明财报之后，市场对它值多少钱的分歧宽到了四分之一个股价；分歧这么宽，来回就不会停。SPCX 的摆动是单向的，价格走到新位置就没回去，那是共识在搬家。

本节的判断：闪迪这种高振幅、零净变的形状，说明财报没有制造共识，只制造了换手。 分歧要么被下一个数据收敛，要么继续来回摩擦。

推翻条件：下周它的周振幅缩到本周一半以下、且价格站稳一侧，说明分歧已收敛，本条作废。

口径：周振幅＝周内最高减周内最低，除以上周五收盘，与当日振幅不是一回事。 高低价取自官方日线。

四 · 温度计：保险在变便宜，而便宜的是近的那一头

一句话：这个数字不是给买保险的人看的，是给卖保险的人开的价。

恐惧的标价指数这一周降了四个交易日，非农那天回了一下：08.03 读数 68.1，08.04 是 67.9，08.05 降到 63.6，08.06 是 61.5，08.07 回到 63.4。一年期波动率 VIX1Y 从 22.90 到 22.66。

我有个恐惧温度计，量的就是市场现在花多少钱给未来一年买保险。保险越贵，说明市场越怕。 63.4 分的意思是，贵到过去三年的前 37%。

但拆开看更有意思。这一周 17 只观察标的，30 天的保险价一只不落全降了，而一年期只有 15 只降。 变化的中位数，近端是 −4.60，远端只有 −0.62，差着七倍。17 只里有 16 只，近端降得比远端多。

降得最狠的是闪迪：30 天从 130.45 砸到 91.49，一周掉了 38.96 个点，而它的一年期只降 9.98。财报开完了，那个不确定性没了，短期保险立刻不值钱；但一年之后会怎样，没人因为一份财报就改口。

口径：两个数同源，都是恒定期限值，变化量＝08.07 减 08.03。

近端是情绪，一条新闻就能跳；远端是成本。 做市商必须给一年后报价，他不猜方向，只算承担这个风险要收多少才不亏。所以远端动得慢：一份财报改不了他对一年的看法。

对拿着正股的人，这条曲线的两端是两笔不同的账。 近端塌了 38.96 个点（闪迪），意味着给持仓买一个月的保护，周五比周一便宜了三成以上；同一个人若在卖备兑收租，能收到的租金也同步薄了三成。保险变便宜和租金变薄是同一件事的两面，取决于站在哪一边。 而远端几乎没动，意味着市场只认为这一件事出清了，没有认为这只票从此安静了。

本节的判断：这一周的降温是事件出清，不是整体撤防。 判据就在期限结构上：真正的恐惧消退应该两端一起降，而本周只有近端塌。

推翻条件：下周一年期那端补跌超过 2 个点，说明市场在整体下调长期风险的定价，本条改判。

它不告诉你下周涨跌，它告诉你，那批必须给未来一年报价、并且要为报错付钱的人，这一周把价钱调低了。

把这一周的两条线并排看，方向是反着的：指数一路往上，而给未来一年买保险反而越来越便宜。

五 · 一只票吃掉全周的三分之一

一句话：这一周异常成交的钱，三分之一压在同一只票上。

全周异常成交的权利金合计 37.4 亿美元，SPCX 一只占 12.1 亿，32.3%，上榜 303 条。第二名英伟达 5.2 亿，第三名美光 4.0 亿。一只票的分量超过第二三名加起来。

一只票占三分之一，对不同的人是两件不同的事。 拿着 SPCX 的人，持有的是全场期权定价最贵的注意力：那 12.1 亿权利金绝大多数押在下周四到期，意味着未来五个交易日，它的每一次波动都会被期权对冲放大，涨和跌都比别的票用力。拿着其他票的人也有一笔账：异动榜被一只票占掉三分之一时，剩下十六只的动静只分享三分之二的注意力，本周美光 900 那档、特斯拉相邻两档的新建，就是在这种阴影里发生的。

本节的判断：集中度本身就是信息。 37.4 亿里 32.3% 压在一只票上，说明的不是市场看好谁，是市场把波动的预算花在了哪。

推翻条件：下周 SPCX 占比回落到 15% 以下而总盘子不缩，说明这只是解禁周的特例；若总盘子跟着一起缩，说明这一周的热闹本来就是它一家的。

口径：每个交易日只取收盘那一轮，五天相加。同一天的多轮绝不相加，做市商倒手会被重复计。

我们把含义摆全。看懂之后怎么动，是每个人自己的账。

六 · 下周看点：钱怎么表态

下周三件事挤在一起：08.12 消费者物价指数、08.13 生产者物价指数、08.14 密歇根消费者信心初值。

三件事全部落在 08.14 到期那一批合约的寿命之内。也就是说，这一批合约要同时为三个数字定价，而它们只有一次结算机会。08.07 收盘时市场为那一档开的价，就是它对这三件事加起来有多大把握的报价。

08.19 是波动率期货近月到期日，从下周起近月进入最后两周。曲线形状再变，就不能只归因于情绪了，滚动效应要一起算进去。

期权异动与逐票数据 → https://chronicle.klay-wang.com/options

温度计读数与两个台账 → https://chronicle.klay-wang.com

这两页每个交易日收盘后自动更新，页上带日期，改不了也补不回去。

English edition

Subject line

The stock you keep watching went nowhere. It travelled a very long way.

Three lines

• Some names look like they went nowhere for the week. They travelled a very long way. SanDisk finished the week −0.22%, essentially back where it started, while its weekly high and low were 26.78% apart. Anyone reading only the weekly bar sees a stock that never moved.
• Every judgment we posted this week settled. Both pre-frozen ranges contained SanDisk's earnings result, and on the day SPCX fell 13.61% it stayed inside the range drawn the session before. We did not bet on direction once this week. Every bet was on range and on outcome.
• Three names closed Thursday in the same kind of place: right against a strike. Micron stood 18.53 from 900, Tesla 0.47 from 320, Nvidia 1.01 from 220. By Friday's close, Tesla and Nvidia crossed, and Micron moved the other way.

1. The week in review: the accounts we posted, settled

In one line: the value of a range written in advance is not that it was right, it is that it cannot be changed afterwards.

Both SanDisk bands held. The range recorded at the 08.04 close was 1219.55 to 1588.79. At the 08.05 close the market narrowed the band and moved it down, to 1209.723 to 1491.277. After earnings, 08.06 closed at 1258.58, inside both, 39.03 above the lower edge of the first and 48.86 above the second.

What is worth logging is the direction of the change. As the stock walked from 1425 down to 1350 into earnings, the market narrowed the band and moved the whole thing lower. It was not holding one quote and waiting. It revised daily, and every revision left a trace.

SPCX was more extreme. The three day range given at the 08.04 close was 101.80 to 135.97. The next day it fell 13.61%, with an intraday low of 106.66 and a high of 117.50, inside that range all session. The fall itself had already been priced the day before.

On SanDisk's 1400 strike, the money stayed on neither side. Of the contracts traded on 08.05, 7.1% stuck on the call side and 9.7% on the put side. Over ninety percent, both sides, opened and closed the same day. That money was not a position held overnight. It came and went within the session.

The scoreboard in one line: every bet this week was on range and outcome, never once on direction, and all five settled.

2. The range the market set for itself also came back to be settled

In one line: it is not only our own posted judgments that get settled. The range the market priced with real money can be settled too.

On 08.06 the market set a move range for each of 17 watched names, running to the 08.07 expiry. At the 08.07 close:

14 of 17 stayed inside the range. 3 went outside (SPCX, Microsoft, Alphabet A).

Both tails are extreme. AMD walked just 6% of its priced range: ±21.56 was priced, 1.31 was moved. SPCX walked 293%. SanDisk came closest of all, priced ±140.78 and moving 138.29, just inside.

Method: the range was priced pre-market on 08.06, anchored to the 08.05 close and covering the 08.07 expiry. Settlement uses the 08.07 close.

3. The memory names looked still. They travelled a very long way

In one line: how much a name finished up or down for the week, and how far it swung during it, are two different things.

Across 17 watched names, 14 rose and 3 fell. But put the weekly range (weekly high minus weekly low, over the prior Friday's close) beside the net change for the week, and the gap is immediate:

The top four are all semis and memory, while the index moved just 3.75%. The widest swinger and the calmest are seven times apart.

SanDisk is the extreme case. It reached 1446.62 and fell to 1121.27, a 26.78% spread, and finished the week 0.22% from where it began. Day by day: +6.03, +10.84, −5.40, −6.81, −3.68. Up first, then all the way back.

SPCX had the same amplitude and a completely different result: a 26.44% range, almost identical, but up 22.83% net. It swung out and did not come back.

Same twenty six percent. One was churn, one was a real move.

The other way round, Nvidia ranged 13.90% and finished up 11.56%：nearly every step it took that week pointed the same way.

This is why a weekly chart misleads. It draws the start and the end. Those 26.78 points of round trip in between take up not a single pixel.

Anyone who held SanDisk saw two prices a quarter apart along the way.

What this table means for anyone holding these names. Range is not an abstract number. It is two balances that actually appeared in the account: whoever held SanDisk saw 1446 and saw 1121 and ended where they started, while whoever held SPCX saw a swing just as wide and ended higher. The difference is not amplitude. It is whether the shares changed hands into a new consensus price. SanDisk traded real money everywhere from 1121 to 1446, which means post earnings disagreement about what it is worth spans a quarter of the share price, and disagreement that wide does not go quiet. SPCX's swing was one way: price moved to a new place and stayed. That is consensus relocating.

The judgment of this section: SanDisk's high range, zero net shape says earnings manufactured turnover, not consensus. Either the next data point closes that gap, or the round trips continue. Void if: next week its range halves and price holds one side. Then the disagreement has resolved and this reading retires.

Method: weekly range equals weekly high minus weekly low, over the prior Friday's close. It is not the same as daily range. Highs and lows come from official daily bars.

4. Thermometer: insurance got cheaper, and the cheap end is the near one

In one line: this number is not for the people buying protection. It is the price quoted by the people selling it.

The Fear-Price Index fell for four sessions this week and ticked back on payrolls day: 68.1 on 08.03, 67.9 on 08.04, down to 63.6 on 08.05, 61.5 on 08.06, and back to 63.4 on 08.07. One year volatility VIX1Y went from 22.90 to 22.66.

The thermometer measures one thing: what the market currently pays to insure the next year. The dearer the insurance, the more the market is afraid. A reading of 63.4 means it is dearer than on 63% of days in the past three years.

Pulling it apart is more interesting. Across the 17 watched names this week, the 30 day price of insurance fell in every single one, while at one year only 15 fell. The median change was −4.60 at the near end against −0.62 at the far end, seven times apart. In 16 of 17 names, the near end fell further than the far end.

The steepest was SanDisk: 30 day volatility went from 130.45 to 91.49, down 38.96 points in a week, while its one year figure fell only 9.98. Earnings were out and that uncertainty was gone, so short dated cover was immediately worth less. But what happens a year from now, nobody rewrites on the strength of one quarterly report.

Method: both figures are constant maturity, same source. The change is 08.07 minus 08.03.

The near end is sentiment and moves on a headline. The far end is cost. A market maker has to quote a year out. He is not guessing direction, he is calculating what he must charge to carry that risk without losing. So the far end moves slowly. One earnings report does not change his view of a year.

If you have ever bought anything with the word "principal protected" on it, a snowball note or a structured product, the long end of this curve is what prices your contract. You have never looked at it, and it has been pricing you all along.

For anyone holding the shares, the two ends of this curve are two different bills. The near end collapsing 38.96 points (SanDisk) means one month of protection cost over a third less on Friday than on Monday, and the same holder writing covered calls collects a third less rent. Cheaper insurance and thinner rent are the same fact seen from opposite sides. The far end barely moving means the market judged that one event cleared, not that the name has gone quiet.

The judgment of this section: this week's cooling was event clearing, not a general stand down. The test lives in the term structure: a genuine retreat lowers both ends, and this week only the near end fell. Void if: the one year end catches down by more than 2 points next week. That would be the market repricing long horizon risk, and this reading changes.

It says nothing about next week. It says that the people who must quote a year ahead, and who pay when they quote wrong, lowered their price this week.

Put the week's two lines side by side and they run opposite: the index climbed all week, while insuring the next year kept getting cheaper.

5. One name took a third of the week

In one line: a third of this week's flagged money sat on a single name.

Flagged premium for the week totalled $3.74 billion. SPCX alone accounted for $1.21 billion, or 32.3%, appearing 303 times. Nvidia was second at $520 million, Micron third at $400 million. One name outweighed the second and third combined.

One name taking a third means two different things to two different holders. Holding SPCX means holding the most expensively priced attention in the market: most of that $1.21 billion expires next Thursday, so for five sessions every move it makes gets amplified by option hedging, in both directions. Holding anything else carries its own note: when one name takes a third of the flagged tape, the other sixteen share what is left of the attention, and Micron's 900 line and Tesla's adjacent strikes were built in exactly that shadow.

The judgment of this section: concentration is itself information. 32.3% of $3.74 billion on one name says nothing about who the market likes. It says where the market spent its volatility budget. Void if: next week SPCX's share falls below 15% while the total holds, in which case this was an unlock week exception; if the total shrinks with it, the week's noise was one name's all along.

Method: one closing round per session, summed across five sessions. Rounds within the same day are never added together, or a market maker's turnover gets counted twice.

We lay the meaning out in full. What anyone does with it is their own account.

6. Next week: how the money declares itself

Three events land together next week: CPI on 08.12, PPI on 08.13, and the preliminary Michigan consumer sentiment reading on 08.14.

All three fall inside the life of the contracts expiring 08.14. That batch has to price three numbers at once, and it settles only once. Whatever the market quoted for that strike at the 08.07 close is its quote on how much confidence it has about all three together.

08.19 is the front month volatility futures expiry, so from next week the front month enters its final fortnight. If the curve changes shape after that, it cannot be attributed to sentiment alone. The roll has to be counted in too.

Options activity, name by name → https://chronicle.klay-wang.com/options

The thermometer and both ledgers → https://chronicle.klay-wang.com

Both pages update automatically after each close, carry their date, and cannot be edited afterwards.

一、七月的最后一天，是这个月的缩影

07.31 收盘，标普 ETF 报 747.03，全月 +0.03%。一个月，指数回到了它出发的地方。

而这一天的盘中是这样的：开盘五分钟，它冲到距自己伽马翻转位 0.31 点的地方被拒；接下来四十分钟跌掉 8.6 点，连破昨收、看跌密集峰 743、看跌墙 740；午后又全部收回，收在翻转位上方。当天早上那份宏观数据落地时，它只动了 0.6 点。

收盘价什么也没说，中间发生的一切才是这个月的内容。指数只告诉你平均发生了什么，它从不告诉你发生在谁身上。

二、七月记分牌：先报错的

我们记录的这些交易日里，逐条立案、逐条开奖。

错的两条，错在同一个地方：财报夜押谁动得大：我们押了谷歌，动得大的是特斯拉；押两家都不会走出各自的波动围栏：结果双双跌穿下沿，围栏根本不够宽。两条都是在猜幅度和名次。

对的三条，也对在同一个地方：英伟达那笔深度价内大单是移仓不是新押注（次日两条腿的隔夜持仓一起动，坐实）；横跨三只票的五条腿是同一笔程序单（五腿隔夜齐增）；财报周当周到期的保险价开盘必然塌（塌了 55 到 60 个波动点）。三条都是在读资金留下的结构。

还有两条我们答不出来。七月最后一天，英特尔那组四腿和美光那七档看涨阶梯到期，而它们的隔夜持仓查不到了：当日到期合约只有一个采集窗口，我们的轮次没覆盖上。这两条的结论是无法判定，不是没成立。教训已经立档：埋在当日到期合约上的开奖，动作必须排进当天的轮次。

模式干净得有点刺眼：猜价格往哪走、猜谁动得大，0 对 2；看懂这笔钱在干什么，3 对 0。我们不猜牌，我们看筹码。

记录的价值全在于事前落库、事后不改；一旦允许自己回头修，整本账就一文不值。记分牌上的每一个错，都是在给对的那几条背书。

三、一个一动不动的指数，和它底下的 71 个百分点

全月 17 只标的，7 只上涨、10 只下跌。上面是微软 +24.58%、亚马逊 +13.95%、苹果 +6.76%；下面是闪迪 −46.57%、SPCX −36.57%、英特尔 −35.40%、美光 −28.70%、特斯拉 −26.01%；中间是标普 ETF +0.03%、纳指 ETF −6.57%、英伟达 +0.33%。

最高与最低差 71.2 个百分点，而指数本身一个月只走了 0.03%：这不是指数没事、个股有事，是两边的极端恰好抵掉了。一边是财报直接改写市值的（微软一天多出四千五百亿美元，有记录以来最大的单日增量）；一边是把上半年涨幅整段还回去的。

存储那两只的背景值得单独说一句，因为它很容易被误读：闪迪在过去 52 周里从 40 涨到 2354，58 倍。所以这个月的 −46.57%，是从天上下来，不是被错杀。我们只记录它走过的位置，不评价它该在哪。

一个 +0.03% 的月份里，有人赚了四分之一，有人少了将近一半。指数只告诉你平均发生了什么。

四、三个镜头看七月

价格：这是墙与位置的一个月。 07.17 月度到期日，看涨与看跌的密集价位塌缩成一根针：谷歌的现价、两道墙、翻转位、最大痛点全部挤在 1.4 点之内。07.20 重置，墙位整个重排。07.29 联储决议。财报周把伽马地图连夜重画：纳指 ETF 的翻转位一夜下移 33.5 点，美光下移 96.5 点。到 07.31，苹果收在自家看跌密集价位下方，亚马逊冲破自家看涨密集价位 16 点。整月里价格与它自己那道墙的关系反复兑现，不是因为墙有魔力，而是因为那些价位上真的堆着钱，堆得够多，价格路过时就会被拽一下。墙会移，而且每个到期日之后必须整个重读。

资金：我们记下的，和后来发生的。 07.20 清晨有人用一万二千美元买了三千张四美分的深度崩盘保险，一笔打完、全天再无加仓；07.22 美光 800 线的一亿美元大单；07.23 英伟达的深度价内移仓，次日坐实；07.30 英特尔的四腿结构与标普 09.04 到期的两侧新单：最后这笔在 07.31 开出了七月最干净的一个对照：同一天、同一个到期日，看跌那侧六成的量留仓过夜，看涨那侧只留下 8.7%。 而那个到期日，钉在八月非农那一天。四美分的保险后来没有用上；移仓的判断被数据坐实；两笔到期日的开奖我们查不到，就写查不到。能被检验，是这本账唯一的护城河。

恐惧：从 73.5 到 88.4，再到 68.5。 为未来一年买保险的价钱，七月从 73.5 出发，07.10 落到 70.6，07.17 到期日冲到 82.9，联储决议那天冲到全月最高 88.4，随后两个交易日回吐 19.9，最后一个交易日收在 68.5：全月最低。一个月吵完，这份怕收在了比月初更便宜的地方；而冲高和回落都发生在事件的当天和第二天。保险贵起来快、便宜回去也快，说明七月买的怕是对着具体日子买的，不是对着未来一年。

五、四道围栏的结算：三出一内，和一个必须说的反转

财报前，市场自己用真金白银给四家标出了一周的波动宽度，我们在事前原样印出、锚定 07.27 收盘、事后一字未改。07.31 收盘结算：微软 +19.43%（栏 ±6.17%，出栏）、亚马逊 +17.37%（±5.89%，出栏）、苹果 −8.31%（±3.35%，出栏）、Meta −6.26%（±6.98%，栏内）。

三只走出了市场事前为它们画的宽度：这一轮财报的幅度被系统性低估了。

但真正该记住的是 Meta：07.30 我们印过它已经出栏（当时 −9.23%），而结算日它回到了栏内。 那个中途读数没有错，它只是不是结论。这和我们这个月在联储决议日做的六点采样是同一件事：同一天的保险价，你取哪个时点，可以得出三个互相矛盾但都真实的结论。所以主序列锚收盘，不是因为收盘更准，而是因为它是唯一一个不需要我们挑选的点。

中途读数只告诉你那一刻，从不告诉你结局。

六、月末最后两天，债市那边动了

去杠杆解决的只是资金的问题；反弹能走多远，要看的是另一张桌子。而七月最后两天，那张桌子上有一个可查的变化：30 年期与 10 年期美债的收益率利差，从月内最低的 47 个基点走阔到 55 个基点，其中联储决议当天 +3 个基点、第二天再 +3 个基点；月末最后一天小幅收窄到 53 个基点，但两条腿双双再上台阶（10 年期 4.74%、30 年期 5.28%，均为七月最高）。

拆开两条腿看：07.29 到 07.31，10 年期从 4.62% 升到 4.74%，30 年期从 5.14% 升到 5.28%：两条腿都在跌价，而长端跌得更狠，利差因此从 47 个基点的月内低点走阔。这个形状不是在交易短期还会不会加息，它更像在给长期通胀、政策路径与财政可信度重新定价。

市场上有个说法叫债市义警：债券市场没有散户，全是机构的钱；当这批钱对政策不满意，它们不发声明，它们用价格表态。某种意义上，债市正在替联储做一部分收紧金融条件的工作：而这件事从来不问股票同不同意。

这一段是七月给八月留下的问号，不是答案。它可以被检验：七月最后三个交易日，30 年期已经连续三天收在 5.1% 上方、最后一天到 5.28%。如果这个位置继续抬升，这条逆风就还在账上；如果退回 5.0% 下方，这一段就翻篇了。08.12 的通胀数据和八月的长债标售，是接下来两个可查的时点。

而七月苦了一整个月，最后一周终于给了一段像样的反弹：这句话和上面那段并不矛盾，它们只是两张不同的桌子。

七、我们说过的那些话，后来怎么样了

不是为了邀功，是为了让你能检验：一本账值不值钱，取决于它敢不敢让你回头翻。

07.27 我们写过一句市场追的是英伟达的信用故事，不是存储的供给故事。 那天它只是当日的一个观察。到月末：英伟达全月 +0.33%，而存储那两只是 −28.70% 和 −46.57%。两个故事被定成了完全不同的价，这个分野走完了整个七月。

07.25 我们写过存储跌了三成，上车的票反而更贵了。 那句话的重点不是方向，是一个当时可测的事实：正股在跌，而它的保险价在涨。到月末，存储又跌了一大截，而它们的长端保险价仍然是全宇宙最高的一档（闪迪一年期读数 106.51，是标普的五倍多）。跌不便宜，是七月存储最稳定的一个特征。

07.23 我们写过谷歌 CFO 还没念出数字，存储股的期权已经把剧本演完了。 后来资本开支确实成了整月的主线之一。

也有没兑现的：07.20 那张四美分的崩盘保险，到七月结束一次都没用上。 我们当时只写了有人付了这个价，没说它会赢：这也是为什么现在不用改口。

07.22 我们记的美光 800 线一亿美元大单，07.31 到期，收盘 823.03，远低于保本线 996。 这一笔的结局是亏的，我们照记。

八、三个还没有答案的问题（依据 → 推理 → 推翻条件）

关于大盘：一个不动的指数，是平静还是拉锯？
依据：七月标普 ETF +0.03%，而 17 只成分里最高与最低差 71.2 个百分点；同期为未来一年买保险的价钱从 73.5 到 88.4 再到 68.5。推理：如果是平静，指数不动、成分也不该散成这样；如果是拉锯，就应该看到成分剧烈重排而指数被抵消：七月是后者。所以七月市场很稳这句话，在指数层面成立，在成分层面不成立。推翻条件：八月如果成分差距收窄到 30 个百分点以内而指数继续横着，那才是真的平静。

关于科技内部：是轮动，还是分化？
先说我们不能说的：轮动意味着钱从 A 流向了 B，而这需要资金流向数据：我们没有，所以我们从不用这个词。 我们能测的是分化：同一天里，微软几乎平开、涨幅 3.27% 全部产生在盘中、收在当日区间的 88% 位置；而超微与英特尔收在当日区间的最低 1%。同一个板块，同一天，两种完全相反的日内形状。月度上更极端：微软 +24.58%、英特尔 −35.40%，同属科技。推理：能确认的是同一个标签下的资产已经不再一起走；不能确认的是钱有没有从一边搬到另一边。推翻条件：如果八月这些票重新开始同涨同跌（日度相关性回升），分化这个描述就该收回。

关于存储：这是死猫跳，还是反转？
这个问题我们不回答方向，但可以把它拆成可查的坐标。依据：闪迪 52 周从 40 到 2354（58 倍），七月 −46.57%，距历史最高回撤 48.40%；美光七月 −28.70%。而 07.31 这天它们的日内形状是高开被卖：闪迪开盘时还涨 8.10%，从开盘算跌 12.20%；美光跳空 +5.18%，从开盘算跌 10.54%，两只都收在当日区间的低位。推理：跳空是隔夜给的，盘中是当天卖的：这两件事的参与者不是同一批人，而只看全天跌幅会把这个区别抹掉。同时，07.31 有人在美光今天这根阴线里，为 1000 线的看涨期权付了 517 万美元（vol/OI 3.66，是当天新建的仓）。推翻条件：不看方向，看两件事：其一，跳空与盘中是否重新同号（连续几天高开高走或低开低走，说明隔夜与日内的人看法一致了）；其二，长端保险价是否回落（闪迪一年期读数从 106.51 往下走，说明市场不再为它的不确定性付这么高的价）。这两件事我们每天都在记，到时候回来对。

九、信用故事讲完了，那存储呢：把七月的钱和八月的位置摆在一起

七月我们写过一句市场追的是英伟达的信用故事，不是存储的供给故事。一个月过去，这句话的两半走成了两个极端：英伟达全月 +0.33%，存储那两只是 −28.70% 和 −46.57%。信用故事的定价没变，供给故事的定价崩了。

但真正值得留到八月的，不是这个结论，是钱在崩的过程里做了三件互相矛盾的事。

第一件：短打的钱在美光身上是偏看涨的。 我们记录的这九个交易日，美光的期权权利金 33 亿美元排全场第三，其中 67% 付在看涨一侧：而它同期跌了 28.70%。这不是有人在抄底（成交数据看不出谁买谁卖），但它至少说明：跌的过程中，付钱的人并没有一边倒地站在下跌那一侧。

第二件：闪迪的留仓换了侧。 看跌与看涨的未平仓比从 1.79 掉到 0.78，留在场上的仓从看跌为主变成了看涨为主，而同期股价近乎腰斩、保险价仍在自己一年的 93.6 分位。跌了一半、赌跌的持仓反而少了、保险还是最贵的一档，这三件事同时成立，是七月存储最难解释的一段。

第三件：长天期的钱压根不在存储上。 到期在三个月以上的新建仓，前四名是纳指 ETF（161 条 / 13 个交易日）、SPCX、特斯拉、标普 ETF：存储一只都没进前四。短打热闹、长钱缺席，这两句要一起说才诚实。

八月的位置已经摆在那里了。 07.31 收盘时，美光的现价站在自家伽马翻转位 上方约 27 点，英特尔在自家翻转位 上方约 7 点，超微 上方约 45 点；而闪迪当天的翻转位算不出来（我们的两把尺子都给不出值，这在本账里是第一次，原样记档）。同一时点，标普与纳指的现价都贴在各自翻转位下方（标普差 0.68 点、纳指差 4.93 点）。

八月最挤的一个日子是 08.21：我们 07.31 一天就记到 562 条押在这个到期日上的异常成交，是次高（08.14 的 256 条）的两倍多。那天是月度期权到期日：七月的 07.17 就是这样一天，当天谷歌的现价、两道墙、翻转位、最大痛点全部挤进 1.4 点之内。

所以八月我们盯三件事，都不需要预测方向，到点自己会有答案：

• 存储的翻转位守不守得住（美光 882、英特尔 89 一线；跌破意味着做市商的对冲方向换边）

• 闪迪的翻转位能不能重新算出来（算不出来本身就是结构极端的一个读数）

• 08.21 那 562 条押注结算成什么（到期日当天，墙会塌成针，然后整个重排）

存储是不是死猫跳，我们不猜。但下个月的每一天，这三个位置都会告诉我们一点新的东西，而它们现在就已经写在账上了。

十、这支温度计到底关你什么事（每期都会讲，这次讲透）

你大概率没买过一年期期权。所以先说结论：这个数字不是给买保险的人看的，是给卖保险的人开的价。

先把两个数分清楚。 大家熟悉的恐慌指数（VIX）量的是未来三十天：今晚会不会出事。它一天能跳 20%，一条新闻就够。我们这支温度计量的是未来一年：这一整年的不确定性，现在值多少钱。它一天动 2 个点就算大动作。

为什么长的那个更值钱？ 因为短的那个是情绪，长的那个是成本。

做市商必须报价：不管他看多看空，客户要买保护，他就得给一个数。他不猜方向，他只算一件事：我承担这个风险，得收多少钱才不亏。这个数是算出来的，不是感觉出来的。 所以它一天变三次是不可能的；它一旦真的动了，说明有人的成本模型改了。

它跟你的持仓有什么关系，三种情况：

一、你被一条新闻吓到、在想要不要跑的时候。 看短端和长端有没有分家。如果恐慌指数飙了、这个数没动：你慌在了专业资金前面，卖保险的人认为这是一次事故，不是路况变了。如果两个一起涨，那才是重定价。七月这两种都出现过：决议日两个一起冲到 88.4，第二天长端两天回吐 19.9，最后收在全月最低的 68.5。

二、你想给持仓买份保护的时候。 这个数直接告诉你现在买贵不贵。今天 68，意思是这份保险比过去三年 68% 的日子都贵。它不告诉你该不该买，它告诉你你付的是什么价。

三、你买过雪球、结构化票据、或者任何带保本触发字样的产品。 这条最直接：定价你那份合约的，就是这条曲线的长端。 你签的那张纸背后，发行方卖给你的正是一年期以上的波动率。这个数高，意味着同样的结构现在能给你更好的条款；这个数低，意味着条款会变差。你没看过它，但它一直在给你定价。

谁天天在用它： 做市商给长期期权报价用的就是这条曲线的长端；养老金和保险公司做长期对冲买的是一年以上的保护；发行结构化产品的机构，成本直接吃这个数。这些人加起来，是市场上最不需要靠猜方向吃饭的一批人。

所以这支温度计的用处，一句话：它不告诉你明天涨跌，它告诉你，那批必须给未来一年报价、并且要为报错付钱的人，今天报了多少。

十一、Citadel 在那场爆仓里是反派吗

七月最后一周，一支峰值约 450 亿美元的基金把全部公开持仓协议转给了 Citadel，同时缩到约 100 亿。它用了约 4 倍杠杆，收到三家主经纪商的追加保证金通知；转手前六天，创始人还在募资信里喊加码。（据 Bloomberg / CNBC / Semafor 交叉，细节最全的是 Semafor 版本。）

看到大机构围猎某个玩家，第一反应通常是华尔街巨鳄又在收割人。我们不替任何一方说话，但从这本账天天在看的东西出发，有三件事值得摆出来。

一、异常的回报加上公开的持仓，等于暴露的弱点。 一两年用几亿做到几十倍，选股准只是入场券，高杠杆是必然。而极高的收益不只是战绩，它同时是诱饵：当你赚得太快，所有顶级玩家都会开始反推你的仓位在哪。更何况持仓披露写得清清楚楚：哪怕滞后一个季度，方向也够了。在这个市场里，赚得漂亮和被看穿，是同一件事的两面。

二、定向爆破不是恶意，它是一类常规的收益来源。 发现一个杠杆过高、且持仓流动性很差的对手盘，然后把它推到追加保证金那条线：这在机构语言里叫提供流动性并收取风险溢价。听起来冷血，但机制上它和做市商在你恐慌时接你的货是同一件事：当你的流动性覆盖不了你的杠杆，你就成了别人的利润来源。 这不是道德问题，这是结构问题。

三、这类猎手的存在，某种程度上是在保护市场。 如果没有人去定点清除高杠杆玩家，泡沫只会被吹得更大，最后换来的是更大的系统性事故：2008 年就是没人及时清除的结果。这些机构的存在，逼着所有基金在加杠杆的时候真的去做压力测试。健康的市场需要多头，也需要空头。恐惧是维持市场理性的唯一工具：而给恐惧标价，恰恰是这本账每天在做的事。

从我们的数据视角，这件事解释了一个方法上的选择。 我们每天记的是留仓率，不是成交量。成交量只说明那一刻有人付了钱；只有隔夜持仓才说明他愿不愿意、以及扛不扛得住把这笔钱留在场上过夜。一支 4 倍杠杆的基金在崩之前，成交量上看不出任何异样：它出问题的地方从来不在买没买对，而在能不能撑到兑现。这也是我们这个月那句叠句的来处：幅度是市场给的，能不能拿住，是杠杆在事前就决定了的。

这一节引用的全部是公开报道与已披露持仓，我们没有任何一方的内部信息；判断的部分是机制解释，不是对任何机构动机的指控。

十二、华尔街不希望你知道的

华尔街真正不希望你知道的，不是哪只票要涨。是这件事：它不靠赢你赚钱，它靠你一直坐在牌桌上赚钱。

老派交易台训练新人，第一堂课往往不在办公室。做法是带他去一趟赌场，什么都不教，只让他玩，然后在旁边看他怎么玩。因为赌场是这个世界上把人性压缩得最狠的地方：它把你未来三十年在市场上会犯的错，用一个晚上演完给你看。

七月的赌桌长什么样。 标普 ETF 全月 +0.03%，而 17 只标的最高与最低差 71.2 个百分点。财报前市场自己用真金白银画的四道波动围栏，破了三道。07.31 那天，闪迪开盘时还涨 8.10%，从开盘算跌 12.20%；美光跳空 +5.18%，从开盘算跌 10.54%。为未来一年买保险的价钱在联储决议那天冲到全月最高 88.4，两个交易日回吐 19.9，最后收在全月最低的 68.5。

这就是高波动的真实形状：不是一路往下，是同一天里，看多和看空的理由同时成立。 涨的时候你觉得自己看懂了，跌的时候你觉得自己被骗了，而这两件事发生在同一根 K 线里。

它来之前，账上有四件可测的事。 其一，长端保险价跟着短端一起涨：短端单独飙是一次事故，两端一起动才是重定价，七月两种都出现过。其二，市场自己画的波动围栏开始被系统性击穿：四道破三道，说明连愿意为幅度下注的人，估的都不够宽。其三，跳空和盘中开始反号：高开被卖、低开被买，说明隔夜那批人和日内那批人看的根本不是同一件事。其四，同一天里有票收在当日区间的 88% 位置，有票收在最低的 1%。

这四件事一件都不预测方向。它们只说明一件事：这张桌子的赔率变了。

赌场为什么能赚钱。 不是因为它让你输。任何一把你都可能赢，赌场的数学优势通常只有几个百分点。它真正卖的不是那几个百分点，是你会回来。它有两样你没有的东西：无限的时间，和无限的次数。所以它从来不需要赢你，它只需要你不离开。

台面上的钱，不是你的钱。 假设某个晚上你顺得离谱，台面上堆到七八百万。只要你还坐在那儿，那就不是你的钱，那是赌场先借给你、好让你继续玩下去的筹码。

而人性是什么？人性是台面上的已经是我的了。

于是最后一把判断错了，全部回吐。你带进去的本金，只有一两万。

第二天别人听说这事，会说他只输了一两万，还好。但你自己知道，你输的是七八百万。

这个落差才是真正的伤害。 账户上少了一两万，心里少了七八百万。从那一刻起，你之后下的每一笔单，都是那个输了七八百万的人在下的。

接下来的动作几乎是固定的：去 ATM 取钱。 取第二次的时候，你已经不在玩原来那个游戏了，你在跟一个数字较劲。你不研究规则，也不研究概率，你只想把那个数字追回来。而高波动的市场刚好会不停给你机会，遍地都是入口，每一个都长得像最后一次翻盘。

这里有一个不需要任何数据的自查，也是这一整节唯一的实操：下这一笔之前问自己，我下它是因为我看到了什么，还是因为我上一笔亏了。 第二种答案出现的那一刻，你已经站在 ATM 前面了。

为什么高波动几乎注定不属于普通人。 机构在这种月份里能活下来，靠的不是看得更准，是三件普通人做不到的事：能对冲，能分批，能不玩。前两件要钱，第三件要心。

这本账上就有现成的例子。07.22 有人在美光 800 线砸下一亿美元的大单，07.31 到期，收盘 823.03，保本线 996。这一笔亏了。 那不是散户，那是能一次下一亿美元的人。高波动不看你有多少钱。

同一本账上还有另一笔：07.20 清晨，有人用 12000 美元买了三千张四美分的深度崩盘保险，一笔打完，全天再无加仓，整个七月一次都没用上。这笔也没赢。但它和上面那笔的区别不在看得准不准，在于下的注是不是自己输得起的那个数。

最后说那件最没人愿意听的。 很多人看不上一年只涨十几个点的东西，嫌慢，觉得那点幅度不值得占仓位。但七月这张表本身就是回答：跨度 71.2 个百分点，指数走了 0.03%。 那些让人夜里睡不着的幅度，最后在指数上互相抵消掉了。

慢的那个，你晚上不会去想它。而正因为你不想它，你就不会在凌晨三点改主意；不改主意，时间才有机会替你干活。资产是这么长起来的，不是靠某一把。

幅度是市场给的，能不能拿满整段，是心境给的。

回到开头那句。华尔街不需要你输，它需要你一直在场：每一次换手、每一次想追回来、每一次凌晨改主意，都是在给这台机器交一次过路费。你看得见筹码，才不至于只看得见运气。

这一节讲的是行为与机制，不是任何一类资产的推荐，也不构成投资建议。我们不知道你该买什么，我们只知道七月这张表长什么样。

十三、八月坐标（只给日历，不给方向）

08.03：07.31 标普那两组蝶式几何的持仓，结算后涨不涨。
08.07：非农（微软那条看跌带里，08.07 那一档同日到期）。
08.12：通胀数据：也是上面债市那段的第一个检验点。

08.21：月度期权到期日，墙位会整个重排。

09.04：标普那笔两侧单的到期日，也是九月初的非农发布日。

有人已经为这些日子付了钱。我们不猜他们对不对，到点回来对账，落哪边都照记。

十四、收尾

七月的最后一天，指数收在它一个月前出发的地方。同一个月里，有一只票涨了四分之一，有一只票少了将近一半；有人为一年后的保险付出了三年里最贵的价钱，六天后又收在了全月最便宜处；而在另一张桌子上，最长期的那笔钱在月末悄悄要了更高的价。

指数只告诉你平均发生了什么，它从不告诉你发生在谁身上。 我们这个月做的全部事情，就是把发生在谁身上一天一天记下来：包括记错的那些。

八月接着记。

1 · The last day was the month in miniature

SPY closed 07.31 at 747.03, +0.03% for the month, back where it started. Intraday it ran to within 0.31 points of its own gamma flip in the first five minutes, was refused, lost 8.6 points over forty minutes through three levels, and took all of it back by the close. The macro release that morning moved it 0.6 points.

The closing price said nothing; everything in between was the content. An index tells you what happened on average. It never tells you who it happened to.

2 · The scoreboard, mistakes first

Wrong, both the same way: we said Alphabet would move more than Tesla on earnings night, Tesla did; we said neither would leave its volatility fence, both fell through the floor. Both were guesses about size and rank.

Right, all the same way: Nvidia's deep-in-the-money print was a roll (both legs moved overnight); five legs across three names were one program trade; earnings-week implied volatility would crush at the open (it fell 55 to 60 points). All three read structure the money left behind.

Two have no answer. Intel's four-leg stack and Micron's call ladder expired on 07.31, and same-day expiries offer exactly one window to check open interest, our collection missed it. "Unmeasurable", not "denied".

The pattern is uncomfortably clean: 0 for 2 guessing where price goes, 3 for 0 reading what the money is doing. We do not guess the cards; we watch the chips.

A record is worth something only because it is logged before the fact and never edited after. Allow yourself to go back and tidy it, and the whole book is worth nothing. Every miss on the scoreboard is what makes the hits worth reading.

3 · A flat index and the 71 points underneath

Seven of seventeen up, ten down. Above: Microsoft +24.58%, Amazon +13.95%, Apple +6.76%. Below: SanDisk −46.57%, SPCX −36.57%, Intel −35.40%, Micron −28.70%, Tesla −26.01%. In the middle: SPY +0.03%, QQQ −6.57%, Nvidia +0.33%.

This is not "the index was fine and stocks were not", the two extremes cancelled each other out. One caveat on storage: SanDisk went from 40 to 2354 over 52 weeks, 58-fold. July's −46.57% is a descent from altitude, not a mispricing.

4 · Three lenses

Price, a month about levels. On 07.17 expiry, Alphabet's spot, both walls, its flip and max pain sat inside 1.4 points. 07.20 reset everything; 07.29 brought the Fed; earnings week redrew the gamma map overnight (QQQ's flip fell 33.5 points, Micron's 96.5). By 07.31 Apple closed under its own put wall and Amazon sixteen points through its call wall. Not because walls are magic, because real money sits at those strikes, and every expiry forces a full re-read.

Money, what we logged, and what came after. A $12,000 print for 3,000 four-cent crash puts on 07.20; a $100M block at Micron's 800 line on 07.22; Nvidia's roll on 07.23, confirmed the next day; on 07.30, Intel's four legs and SPY's both-sides prints on 09.04, which settled into July's cleanest contrast: same day, same expiry, the puts kept 60% of their volume overnight and the calls kept 8.7%. That expiry sits on August payrolls. Being checkable is the whole moat.

Fear, 73.5, then 88.4, then 68.5. Opened July at 73.5, 70.6 on 07.10, 82.9 at expiry, 88.4 on Fed day, gave back 19.9 over the next two sessions, closed at 68.5, its July low. The spike and the give-back both landed on event day and the day after: fear priced against specific dates, not against the year.

5 · The fences: three out, one in, and a reversal

Anchored to the 07.27 close, printed in advance, unchanged since. At settlement: Microsoft +19.43% vs ±6.17%, out. Amazon +17.37% vs ±5.89%, out. Apple −8.31% vs ±3.35%, out. Meta −6.26% vs ±6.98%, inside.

Three left the width the market quoted for itself; this round was underpriced for size. But the one to remember is Meta: on 07.30 we printed it as outside at −9.23%, and by settlement it was back inside. That mid-course reading was not wrong, it simply was not a conclusion. Same lesson as our six-point sampling on Fed day, where one session gave three contradictory but equally real answers depending on the minute. We anchor to the close not because it is more accurate, but because it is the only point we do not have to choose.

6 · In the month's last two sessions, the bond market moved

Deleveraging solves a funding problem; how far a rebound travels is decided at a different table. And in July's final two sessions, there was a checkable change at that table: the 30-year minus 10-year Treasury spread widened from 47bp at its monthly low to 55bp, three basis points on Fed day, three more the day after, easing to 53bp on the final session while both legs stepped higher again (10-year 4.74%, 30-year 5.28%, both July highs).

Both legs fell in price and the long leg fell harder: from 07.29 to 07.31 the 10-year went 4.62% → 4.74% and the 30-year 5.14% → 5.28%, the long leg falling harder. That shape is not trading "one more hike"; it reads more like repricing long-run inflation, the policy path, and fiscal credibility.

There is a phrase for this: bond vigilantes. The Treasury market has no retail; it is institutional money, and when that money dislikes a policy it does not issue statements, it votes with price. In a real sense, the bond market is doing part of the Fed's tightening for it, and it never asks equities for permission.

This is July's question mark for August, not an answer, and it is checkable: the 30-year closed above 5.1% for July's final three sessions and at 5.28% on the last; if that level keeps rising the headwind stays on the books, back below 5.0% and the paragraph is closed. August's inflation print on 08.12 and the month's long-bond auctions are the next two dated checks.

July was a hard month, and its last week finally delivered a real rebound. That does not contradict the paragraph above, they are simply two different tables.

7 · What we said, and what happened

Not to take credit, so you can check. A ledger is only worth something if it invites you to look back.

On 07.27 we wrote that the market was pricing Nvidia's credit story, not storage's supply story. That day it was one observation. By month-end: Nvidia +0.33%, storage −28.70% and −46.57%. Two stories, two entirely different prices, and the split ran the whole month.

On 07.25 we wrote that storage had fallen a third while the ticket to ride had gotten more expensive. The point was never direction; it was a measurable fact, the stock falling while its insurance repriced higher. By month-end storage had fallen further, and its long-dated insurance is still the dearest in this universe (SanDisk's one-year reading at 106.51, more than five times SPY's). Cheap it did not get.

On 07.23 we wrote that storage options had acted out Alphabet's script before the CFO read the numbers. Capital expenditure did become one of the month's main lines.

And what did not land: the four-cent crash puts from 07.20 were never needed. We only ever wrote that someone paid that price, never that they would win, which is why nothing needs walking back now. The $100M block at Micron's 800 line settled on 07.31 at 823.03, far below its 996 breakeven. That one lost, and we file it.

8 · Three open questions (evidence → reasoning → what would overturn it)

The index: calm, or a tug of war? Evidence: SPY +0.03% while its components spanned 71.2 points, and the price of a year of insurance ran 73.5 → 88.4 → 68.5. Reasoning: calm would leave both the index and its components still; a tug of war leaves violent component moves that cancel at the index, July was the second. "The market was calm in July" holds at the index level and fails underneath it. Overturned if: August narrows the component spread inside 30 points while the index still goes nowhere. That would be actual calm.

Inside tech: rotation, or dispersion? First, what we cannot say: "rotation" claims money moved from A to B, and that requires flow data we do not have, so we never use the word. What we can measure is dispersion: on the same session, Microsoft opened flat, earned all 3.27% intraday and closed at 88% of its range, while AMD and Intel closed in the bottom 1% of theirs. Monthly: Microsoft +24.58%, Intel −35.40%, both filed under "tech". Reasoning: what is confirmed is that assets under one label have stopped moving together; what is not confirmed is whether money crossed between them. Overturned if: daily correlations recover in August and these names start moving as a bloc again.

Storage: dead-cat bounce, or turn? We do not answer direction, but the question breaks into checkable coordinates. Evidence: SanDisk ran 40 → 2354 over 52 weeks (58-fold), fell 46.57% in July, and sits 48.40% below its all-time high; Micron −28.70%. On 07.31 both showed the same intraday shape, gap up, sold all day: SanDisk was +8.10% at the open and fell 12.20% from there; Micron gapped +5.18% and fell 10.54%. Reasoning: the gap was given overnight and the session sold it, those are not the same participants, and the full-day number erases the distinction. Meanwhile someone paid $5.17M for Micron's 1000-line calls inside that red candle (vol/OI 3.66, newly opened). Overturned by, in either direction: first, whether gap and session turn the same sign again for several sessions; second, whether long-dated insurance comes down from SanDisk's 106.51. We record both daily, and we will come back.

9 · What Wall Street would rather you did not know

What Wall Street would rather you did not know is not which name is going up. It is this: it does not make money by beating you. It makes money by keeping you at the table.

The oldest first lesson on a trading desk was not taught at a desk. You took the new hire to a casino, taught them nothing, and watched them play, because a casino compresses thirty years of market mistakes into a single evening.

What July's table looked like. SPY finished the month at +0.03% while the spread between its best and worst name was 71.2 points. Three of the four earnings fences the market priced in advance were broken. On 07.31 SanDisk was up 8.10% at the open and fell 12.20% from there; Micron gapped +5.18% and fell 10.54% from its open. The price of a year of insurance hit 88.4 on decision day, gave back 19.9 in two sessions, and closed the month at its cheapest. High volatility is not a straight line down. It is a day on which the bull case and the bear case are both true.

Why the house wins. Not because you lose. On any single hand you might win, and the house edge is usually a few percent. What it actually sells is your return trip. It has two things you do not: unlimited time and unlimited hands. It never needs to beat you; it only needs you not to leave.

The chips in front of you are not your money. Say a good night runs your stack up to seven million. While you are still sitting there, that is not your money; it is what the house lent you so you would keep playing. Human nature says: what is on the table is already mine. So the last hand goes wrong and all of it goes back. You walked in with twenty thousand. The next day people say he only lost twenty thousand. You know you lost seven million. That gap is the real damage: the account is down twenty thousand, the judgment is down seven million, and every trade you place afterwards is placed by the person who lost seven million.

Then comes the ATM. By the second withdrawal you are not playing the game any more, you are arguing with a number. The one self-check in this section needs no data at all: am I placing this because of something I saw, or because the last one lost? The moment the answer is the second one, you are already at the ATM.

Why high volatility rarely belongs to retail. Institutions survive months like this not by seeing more clearly but by doing three things most people cannot: hedge, scale in, and sit out. The first two cost money; the third costs composure. Our own ledger has the proof: the $100M block at Micron's 800 line on 07.22 expired 07.31 at 823.03 against a 996 breakeven. That one lost, and it was not retail. Volatility does not care how much money you have. On the same ledger, the $12000 paid on 07.20 for three thousand four-cent crash puts never paid off either, and the difference between the two is not accuracy. It is whether the stake was a number you could afford to lose.

And the part nobody wants to hear. People dismiss an asset that returns low double digits a year as too slow. July is the answer: 71.2 points of spread, and the index moved 0.03%. The moves that keep you awake cancelled each other out. The slow one you do not think about at night, and because you do not think about it you do not change your mind at 3am, and time gets a chance to do the work.

The size of the move is given by the market. Whether you hold the whole of it is given by your composure.

Back to the top. Wall Street does not need you to lose, it needs you present. Every turnover, every attempt to win it back, every 3am change of mind pays a toll to that machine. You are less at the mercy of luck once you can see the chips.

⚠️ This section is about behaviour and mechanism. It is not a recommendation of any asset class and it is not investment advice.

10 · August markers (dates only, no direction)

08.03, whether the open interest behind 07.31's two butterfly structures appears. 08.07, payrolls, and one strike of Microsoft's put band. 08.12, inflation, the first check on the bond paragraph above. 08.21, monthly expiry; walls get rebuilt. 09.04, SPY's both-sides expiry, and September payrolls.

Someone has already paid for these dates. We come back and publish, either way.

11 · Closing

July ended where it began. Inside that same month, one name gained a quarter of its value and another lost nearly half; the price of a year of insurance ran to its dearest in three years and, six days later, closed the month at its cheapest; and at another table, the longest-dated money quietly asked for more.

An index tells you what happened on average. It never tells you who it happened to. Everything we did this month was write down who, including the times we got it wrong.

August, same.

📒 本周回顾：哪些对了，哪些错了

一句话：发过的判断，对了回来对，错了也回来认。

先说错的。

周三那封《财报开牌前，聪明钱在买翅膀》里，我们押了两条：谷歌财报之夜会比特斯拉动得大；两只都不会冲出市场赛前划好的波动带。

第一条，错。实际是特斯拉动得大。

第二条，财报当晚看起来对了：两只都乖乖待在带内。但第二天，特斯拉 −14.5%，谷歌 −7.1%，到周五收盘双双跌穿带子下沿（谷歌收 319.74，带下沿 329.17；特斯拉收 313.03，下沿 359.85）。也错。

两个错其实是同一个错：财报的账单，是第二天才付的。我们把当晚的带子，当成了整段行情的围栏。

再说对的。三条，全是同一类。

英伟达那笔约 3 亿美元的大单：我们周三当天判断是老仓搬家、不是新赌注（判断台账 07-22 当天立案，先写判断，后看开奖），并且把第二天该看到的两个数字提前写下来：旧仓应减约 8 万张，新仓应增约 8.8 万张。第二天开奖：−78471 和 +88009。两个数都对上了。

三巨头同一天各冒出 1.1 万张的五条腿：我们判断是同一家机构的程序单。第二天五条腿的持仓齐刷刷各增约 1.1 万张。

财报周被炒到天价的短期期权，开盘后必然大幅泄气：最狠的一只，一夜从 100 上下泄到 41。

本周回顾：猜价格往哪走，0 对 2；看懂这笔钱在干什么，3 对 3。我们不猜牌，我们看筹码。

🙋 问得最多的两个问题

一句话：我不预测价格，但钱的每个动作都记了账。

跌到位了吗？

不知道，也不猜。但这一周有三笔钱的动作，值得你知道。

第一笔：保险没涨价。 你的车要是出了事故，第二年续保，保费多半要涨：保险公司认定这条路以后还会出事，才敢多收钱。这一周市场天天挨打，短期的怕上蹿下跳，来回跑了 12.4%；可给未来一年买保护的价钱，全周只动了 1.9%。出了事故，保费却没涨。 卖保险的是全市场最会算账的一批人：他们没提价，意思是在他们的账本上，这一周是一场事故，不是路况变了。你纠结割肉还是加仓的那一刻，这个数不告诉你该怎么做；它只告诉你：天天跟恐惧做买卖的那批人，这周没改价。

第二笔：有人押会动，不押方向。 有人在 8 月底到期的合约上花了五六百万美元，同时买涨和买跌，两边距离现价对称。这笔单 99% 留仓过夜，是真仓；到期日正好罩住下周的美联储会议和整个财报季。

第三笔：极端的险有人上。 周一有人花一万二美元，买了 3000 张 SPY 跌 86% 才赔的保险，4 美分一张。第二天几乎全部留仓。

三笔连起来读：主流的钱没在慌，会动的钱在加价，极端的险一直有人买。 这不是预测，是记账。

存储还跌吗？

先把一个反直觉的事实放正：周五被血洗的存储，其实是本周最强的板块。 美光这一周 +8.5%（周五 −7%），闪迪 +6%（周五 −10.8%）。周五吐回去的，少于这一周涨出来的。

然后看钱怎么动的：

• 周三收盘前最后半小时，有人花约 1 亿美元买了美光的看涨，这笔单赚钱的门槛是月底站上 996。周五美光收 920.95，低于那道门槛 75 美元。下周五开奖，赢输我们都晒。
• 周五大跌当天收盘前，又有人花 960 万美元，买美光回到 1000 的票。跌下来的当天，就有钱进场赌反弹。
• 但想上车的先看票价：买它们期权的成本，闪迪排在过去一年的 96 分、美光 88 分（100 分就是这一年最贵那天）。便宜的是股票，不是上车的成本。
• 长期的票价没慌：闪迪一年多之后到期的期权价格，一周从 118 降到 114，还在轻微降温。

钱的表态翻译成一句：短线在对赌，长线没改主意，入场费仍在最贵档。

🗓 下周看点：钱已经表态的两场

一句话：我不猜结果，只报市场已经付钱买下的预期。

第一场：Meta 财报，周三盘后，同一晚撞上美联储决议。一晚两颗雷。

钱的表态：市场给它划的围栏 ±8.01%（558 到 655 美元），四家巨头里最宽，是苹果（±3.9%）的两倍。翻译：市场为 Meta 准备了全场最大的意外空间。微软同晚开牌，围栏 ±6.62%。

周四苹果、亚马逊接棒。亚马逊的账本歪着：押它涨的持仓是押跌的 3.4 倍。每张合约都有买方和卖方，这个倍数不代表谁会赢，只代表如果结果朝反方向走，被打脸的人多得多。

第二场：存储的两次终审。

周五 07-31：美光那张 1 亿美元大单开奖，996 见分晓。
08-05：闪迪财报。到时候它 96 分的票价，就要接受检验。

我们的挂账（美光 07-31、诺基亚 08-28）到期都会回来对，赢输都晒。

这封信不告诉你该怎么做。它只保证一件事：我们说过的每句话，都会回来对账。

📒 The scorecard: what we got right, what we got wrong

One line: every call we publish comes back for settlement — wins and losses alike.

The misses first.

On Wednesday we made two calls: Google would move more than Tesla on earnings night, and neither would break out of the range the market had priced in.

First call: wrong. Tesla moved more.

Second call looked right that night — both stayed inside their bands. Then came the next day: Tesla −14.5%, Google −7.1%, and by Friday's close both had fallen through the bottom of their bands (Google closed 319.74 vs. a floor of 329.17; Tesla 313.03 vs. 359.85). Wrong again.

Both misses are the same miss: earnings bills come due the next day. We mistook one night's range for the whole ride's fence.

And one piece of homework we failed: a call required checking open interest before Friday's open. We missed the window; the contracts expired and settled to zero. Permanently unverifiable. Logged as cannot-judge. No excuses.

Now the hits. Three, all the same kind.

That roughly $300M Nvidia print: we called it a position moving house, not a new bet — logged the same day, judgment first, verdict later — and wrote down the two numbers the next morning should show: old position down ~80k contracts, new one up ~88k. The verdict: −78,471 and +88,009. Both matched.

Five legs of ~11.2k contracts each, across three mega caps, in one day: we called it one institution's programmed trade. Next morning, all five legs' open interest rose in lockstep.

Earnings-week options priced at absurd levels must deflate at the open: the worst one went from around 100 to 41 overnight.

The scorecard in one line: guessing where prices go, 0 for 2. Reading what the money is doing, 3 for 3. We don't guess cards. We count chips.

One more confession: we also cited a wrong closing price this week (an after-hours print instead of the official close). The ledger now carries a dated correction; the verdict stands. We publish our bookkeeping errors too — this ledger is worth something precisely because it can't be quietly rewritten.

🙋 The two questions we get asked most

One line: we don't predict prices. We keep books on what the money does.

Done falling?

We don't know, and we won't guess. But three money moves this week are worth knowing.

First: insurance didn't get more expensive. Crash your car and your premium jumps at renewal — the insurer has decided your road is dangerous. The market took hits all week; short-term fear swung 12.4% top to bottom. Yet the price of one-year protection moved just 1.9%. An accident, and no premium hike. The people selling insurance are the best accountants in the market: no reprice means that in their books, this week was a crash, not a change in the road. The moment you're debating whether to cut or add, this number won't tell you what to do — it tells you the people who trade fear for a living didn't change their price this week.

Second: someone is betting on movement, not direction. Someone spent $5–6M on end-of-August contracts, buying calls and puts in equal size, symmetric around the current price. 99% held overnight — real positions. The expiry covers next week's Fed meeting and the rest of earnings season.

Third: tail risk still has buyers. On Monday someone paid $12,000 for 3,000 contracts of SPY protection that only pays if the market falls 86%. Four cents apiece. Nearly all of it held the next day.

Read together: mainstream money isn't panicking, movement money is paying up, and crash insurance keeps finding buyers. That's not a forecast. That's bookkeeping.

Is memory done falling?

First, a fact that cuts against Friday's tape: the stocks bleeding on Friday were the strongest group of the week. Micron finished the week +8.5% (Friday −7%); SanDisk +6% (Friday −10.8%). Friday gave back less than the week added.

Now the money:

• In the last half hour of Wednesday, someone spent about $100M on Micron calls that only pay if it stands above 996 by month-end. Friday's close: 920.95, 75 dollars below the line. Verdict next Friday; we post it win or lose.
• During Friday's selloff, before the close, another $9.6M bought tickets on Micron returning to 1,000. The day it fell, money showed up to bet on the bounce.
• But check the ticket price before boarding: options on SanDisk cost more than on 96% of days in the past year; Micron, 88%. The stock got cheaper. Boarding didn't.
• Long-dated prices aren't panicking: SanDisk's year-plus options drifted from 118 to 114 over the week. Cooling, slightly.

The money's statement in one line: the short term is a knife fight, the long term hasn't changed its mind, and the entry fee is still near the top of the range.

🗓 Next week: where the money has already voted

One line: we don't guess outcomes. We report the expectations the market has already paid for.

Act one: Meta reports Wednesday after the close — the same night as the Fed decision. Two grenades, one evening.

The market's fence for Meta: ±8.01% ($558–$655), the widest of the four mega caps reporting next week — twice Apple's (±3.9%). Translation: the market has budgeted Meta the most room for surprise. Microsoft reports the same night, fenced at ±6.62%.

Thursday, Apple and Amazon take the stage. Amazon's book is lopsided: bets on up outnumber bets on down 3.4 to 1. Every contract has a buyer and a seller — the ratio doesn't say who wins, only that if it goes the other way, far more faces get slapped.

Act two: two verdicts on memory.

Friday 07-31: that $100M Micron bet settles. 996 is the line.
08-05: SanDisk reports. Its 96th-percentile ticket price goes on trial.

Open cases (Micron 07-31, Nokia 08-28) all come back for settlement — posted win or lose.

This letter won't tell you what to do. It guarantees one thing: everything we say comes back for settlement.