Rotation happens a level down轮动发生在更细的一层

The S&P 500’s 11 GICS sectors split into 25 industry groups, and this year’s rotation happened at that level. Only 8 of S&P’s 36 official GICS index symbols carry any price history on Yahoo Finance, so we rebuilt all 25 groups from the index’s own members, cap-weighted, and checked them against those 8: one-day moves match to the hundredth of a point. The shortcut we refused on the way — pricing the past with today’s members — read the S&P 500’s year at +18.37% against an official +17.24%.

标普 500 的 11 个 GICS 板块之下共有 25 个行业组，今年的轮动正发生在这一层。在 Yahoo Finance 上，标普 36 个官方 GICS 指数代码中只有 8 个带有历史价格，于是我们用指数自己的成分股、按市值加权重建了全部 25 个行业组，再拿这 8 个官方指数核对：单日涨跌幅精确到 0.01 个百分点完全一致。途中我们拒绝了一条捷径——“用今天的成分股给过去定价”——它算出的标普 500 一年回报是 +18.37%，而官方数字是 +17.24%。

Most sector views, including the two already on our Markets page, are ETF prices — the Select Sector SPDRs, one fund per sector. They answer “which sector?” and stop. Yet rotation usually happens a level down: semiconductors against software inside Information Technology, banks against insurers inside Financials. A sector fund averages the two halves, so the move is invisible by construction.

GICS, the classification S&P and MSCI use to file every listed company, splits the 11 sectors into 25 industry groups, and most of them have no ETF on the S&P 500’s own definition. Even the sector funds are not quite their sectors: the Select Sector SPDRs cap single names, and three companies are over half of S&P 500 IT.

S&P’s official indices would be the obvious source (^SP500-4530 is semiconductors). We measured them on Yahoo Finance: 35 of the 36 symbols resolve, but only 8 carry any history. The rest return today’s quote alone — not even a one-week return.

大多数板块视图——包括我们市场页上原有的两个——用的都是 ETF 价格：Select Sector SPDR 系列，每个板块一只基金。它们回答“哪个板块”，然后就止步了。可轮动通常发生在更细的一层：信息技术内部是半导体对软件，金融内部是银行对保险。板块基金把两半平均在一起，这种分化在结构上就看不到。

GICS（全球行业分类标准）是标普与 MSCI 给每家上市公司归类的体系，它把 11 个板块细分为 25 个行业组，而其中大多数都没有按标普 500 自身口径跟踪的 ETF。就连板块基金也不完全等于板块本身：Select Sector SPDR 对单一个股设有权重上限，而标普 500 信息技术板块一半以上的权重集中在三家公司身上。

最直接的数据源本应是标普的官方指数（半导体的代码是 ^SP500-4530）。我们在 Yahoo Finance 上实测过：36 个代码中有 35 个能查到，但只有 8 个带有历史数据，其余只返回当天的报价——连一周的回报都算不出来。

Each group is rebuilt as a cap-weighted index, in which a company counts in proportion to its market value. The inputs are public: weights from the daily holdings of SPY, an ETF that replicates the index; each company’s GICS sub-industry and the date it joined, from Wikipedia’s list of S&P 500 companies; and daily adjusted closes (prices restated for splits and dividends) for all 503 names.

Each member’s weight on the snapshot date becomes a fixed share count — weight divided by that day’s adjusted close — so a group’s value on any day is the sum of shares times price. That is buy-and-hold with constant shares, which is what a cap-weighted index is between rebalances.

Three things that had to be right

Each is pinned by a test that fails without it.

1. Each window counts only companies already in the index when it began. The usual shortcut prices today’s members back through time, which builds in survivorship bias — judging the past by the companies still here. Companies join the S&P 500 after they have risen, so today’s list imports their pre-entry gains into the index’s past.
2. Share counts are struck against the same adjusted series the returns come from. Adjusted prices are rewritten after every split, dividend and spin-off; pairing a raw price with an adjusted one would misweight every company that has split since, by exactly the split ratio. Spin-offs are covered too: DD→Q, FDX→FDXF and HON→HONA show no cliff.
3. Year-to-date starts from last year’s final close, not January’s first, which would throw away 2 January’s move. A figure computed mid-session drops today’s unfinished bar rather than mixing a live quote into a column of closes.

Checked against the official indices

The 8 official indices with history, plus the S&P 500 itself, are an independent answer key. On price-only closes, to match them, one-day moves reproduce to the hundredth — Semiconductors +0.48% against +0.48%, Banks +1.30% against +1.30%, the S&P 500 +0.51% against +0.51% — and one week to three months land within about 0.1 percentage points for every group of ten or more names.

每个行业组都按市值加权指数（公司按市值大小计入）的方式重建。输入都是公开数据：权重来自 SPY（一只完全复制标普 500 的 ETF）每日公布的持仓；每家公司的 GICS 子行业及纳入指数的日期，来自维基百科的标普 500 成分股列表；以及全部 503 只成分股的每日复权收盘价（按拆股和分红调整后的价格）。

每只成分股在快照日的权重被折算成固定的“股数”——权重除以当天的复权收盘价——于是某个行业组在任意一天的市值，就是各成员股数乘以当天价格之和。这相当于股数不变的买入并持有，而这正是市值加权指数在两次调整之间的样子。

必须做对的三件事

每一条都有测试锁定，去掉它，测试就会失败。

1. 每个区间只计入区间开始时已经在指数里的公司。常见的捷径是拿今天的成分股一路往回算，这就自带了幸存者偏差——用今天还留在场上的公司去评判过去。公司往往是大涨之后才被纳入标普 500 的，于是今天的名单把它们入选前的涨幅也带进了指数的过去。
2. 股数用与回报同一条复权价格序列来折算。每次拆股、分红和分拆之后，复权价格都会被整体改写；若把未复权价格与复权价格混用，所有在快照之后拆过股的公司都会被错配权重，误差恰好等于拆股比例。分拆同样覆盖在内：DD→Q、FDX→FDXF、HON→HONA 都没有出现断崖。
3. 年初至今从去年最后一个收盘价算起，而不是 1 月的第一个——那样会把 1 月 2 日当天的涨跌整个丢掉。盘中计算时则舍弃当天尚未收盘的那根 K 线，不把实时报价混进一列收盘价里。

与官方指数核对

有历史数据的 8 个官方指数，加上标普 500 本身，构成一份独立的“标准答案”。为与之匹配，核对时使用不含股息的价格数据：单日涨跌幅精确到 0.01 个百分点完全一致——半导体 +0.48% 对 +0.48%，银行 +1.30% 对 +1.30%，标普 500 +0.51% 对 +0.51%；从 1 周到 3 个月，凡成分股在十家及以上的行业组，误差都在约 0.1 个百分点以内。

Year to the 25 September 2026 close, total return. The sector figure is not wrong; it averages a group up 57% with a group down 2%, and shows neither.

截至 2026 年 9 月 25 日收盘的一年，总回报。板块数字本身没有错，只是它把一个涨了 57% 的行业组和一个跌了 2% 的行业组平均在了一起，结果两者都看不见。

A return alone does not say whether a move mattered: +20% on a group that is 1% of the index is a fifth of a point. So the panel also shows each group’s change in index weight, in percentage points (pp). IT gained 4.2pp over the year, and its three groups sum to it:

单看涨跌幅，说明不了一次上涨对指数有多重要：一个只占指数 1% 的行业组涨 20%，也只相当于指数的五分之一个百分点。所以面板还给出每个行业组的指数权重变化，单位是百分点。这一年里信息技术板块的指数权重上升了 4.2 个百分点，正好是其下三个行业组之和：

The same year, three ways — price-only, to match the official indices:

同一年用三种方法计算——为与官方指数一致，均为不含股息的价格回报：

Dating each company’s entry closes most of the gap, without touching the arithmetic.

按纳入日期计入每家公司之后，缺口基本被补上，而计算公式本身一处都没改。

Below the table, a map plots each group’s return against the S&P 500 over a longer window (across) and a shorter one (up), bubble area its index weight. Ahead of the index on both windows is leading, behind on both lagging; ahead on the longer only is weakening, on the shorter only improving. On the 25 September 2026 close, at one month against three, Technology Hardware, Software & Services and Media & Entertainment led (30% of the index), Semiconductors and Autos were improving (20%), and sixteen of the 25 groups lagged — together, 32% of the index.

表格下方的象限图，横轴是每个行业组在较长区间内相对标普 500 的超额回报，纵轴是较短区间内的，气泡面积代表指数权重。两个区间都跑赢指数为「领先」，都跑输为「落后」；只在长区间跑赢为「转弱」，只在短区间跑赢为「改善」。以 2026 年 9 月 25 日收盘、1 个月对 3 个月来看：技术硬件与设备、软件与服务、媒体与娱乐处于领先（合计占指数 30%），半导体和汽车两个行业组处于改善（20%），25 个行业组中有 16 个处于落后——合计占指数的 32%。

The panel also exposed that the two older sector sections beside it each had their own idea of where a period starts, and both were wrong without looking it. Both turned up while the panel was being built, and were fixed on 26 September 2026:

面板还暴露出，同页两个旧板块模块对“一个区间从哪里开始”各有各的算法，而且都错了，却看不出任何异常。两处都是在搭建面板时发现的，并于 2026 年 9 月 26 日修复：

• The ranking’s “YTD” was a one-year return: it divided by the first weekly close in a year of data, so XLK’s “YTD” was exactly its one-year figure. The Health Care and XLK rows were measured on 25 September 2026.
• The grid dropped every window’s first session, starting at the first close inside the window, so year-to-date lost all of 2 January and disagreed with the panel beside it.

One rule, in one place

All three sections now call one rule: a period’s base is the last close on or before its anniversary — the close before the window, not the first one inside it — and year-to-date’s anniversary is 31 December. A series too short to reach it reports nothing rather than a shorter period’s return under a longer heading, which is exactly how “YTD” had become one year. Three rules fixed in place would agree today and drift the next time one was edited.

• 排名表里的「年初至今」其实是一年回报：它除以的是一年数据里的第一个周收盘价，所以 XLK 的「年初至今」恰好等于它的一年回报。医疗保健和 XLK 两行数字测于 2026 年 9 月 25 日。
• 轮动表丢掉了每个区间的第一个交易日：它从区间内的第一个收盘价起算，于是年初至今丢掉了 1 月 2 日一整天的涨跌，和旁边的面板说法不一。

一条规则，只写在一处

现在三个模块调用同一条规则：一个区间的基准，是其周年日当天或之前的最后一个收盘价——区间开始之前的那个收盘价，而不是区间内的第一个；年初至今的周年日是上一年的 12 月 31 日。数据太短、够不到周年日的序列直接不报数，而不是把较短区间的回报挂在较长区间的标题下——「年初至今」当初就是这样变成了一年回报。三条规则各自就地修补，今天一致，下次谁一改就会各自漂移。

• Removed members are missing from history. Date-gating handles companies that joined, not those that left, so long horizons drift: Media & Entertainment’s one-year figure is 1.5pp under the official index.
• Small groups can miss by more than a point. Telecommunication Services is four names, and EchoStar alone accounts for its gap — so every row shows its constituent count.
• Share counts are held at the latest snapshot’s. Buybacks and float changes move them a few percent a year at most.
• Ours are total returns, and uncapped. A high-yield sector reads above S&P’s price-only headline index by its dividend, and our IT row can disagree with the capped XLK with neither broken.
• The map’s windows overlap — the longer contains the shorter — so its axes are not independent. We did not chain-link them apart: each window counts its own membership, and the quotient would mix two.
• A quadrant describes where relative returns have been, not where they are going. Nothing here recommends buying or selling any group.

• 已被剔除的成分股不在历史里。按纳入日期过滤解决的是“后来加入”的公司，补不回“中途离开”的公司，所以区间越长越会漂移：媒体与娱乐的一年回报比官方指数低 1.5 个百分点。
• 小行业组的误差可能超过 1 个百分点。电信服务只有四家公司，它与官方指数的差距完全来自 EchoStar 一家——所以页面上每一行都标明了成分股数量。
• 股数沿用最新快照的数值。回购和流通股变化一年最多影响几个百分点。
• 我们用的是总回报，且不设个股上限。高股息板块会比只算价格的标普主流指数高出其间的股息；我们的信息技术板块也可能与设有上限的 XLK 不一致，但两者都没有错。
• 象限图的两个区间是重叠的——较长区间包含较短区间——所以两条坐标轴并不独立。我们没有用链式拆分把两者分开：每个区间只计入自己起点时的成分股，相除会把两套成分混在一起。
• 象限描述的是相对回报过去所处的位置，而不是未来的方向。本文不构成买入或卖出任何行业组的建议。

On the Markets page, under Sector Rotation, the table “GICS Sectors & Industry Groups — S&P 500” opens each sector into its groups and switches between Return, vs S&P 500 and Weight change. The rotation map below it offers four window pairs — 1W/1M, 1M/3M (the default), 1M/6M and 3M/1Y — and trails each bubble back through its last five weekly closes. Every coordinate is a cell from the table’s vs S&P 500 view, so any bubble can be checked against its row. The data is today’s, not 25 September’s.

What pins it: tests/test_gics.py (57 tests, 12 of them on the arithmetic, on prices small enough to check by hand), tests/test_periods.py (12, including a year that ends on a weekend), a 109-check script for the table and map, and a separate comparison against the official indices, run on live market data.

在市场页的「板块轮动」下方，名为「GICS 行业表现 — 标普500一级行业与行业组」的表格可以把每个板块展开为行业组，并在「涨跌幅」「相对标普500」「权重变化」三种视图间切换。表格下方的轮动象限图提供四组区间搭配——1 周/1 月、1 月/3 月（默认）、1 月/6 月、3 月/1 年——每个气泡身后拖着一条轨迹，回溯它过去五个周收盘时的位置。每个坐标都是表格「相对标普500」视图里的一格，任何一个气泡都可以回到它那一行去核对。页面上是最新数据，而不是 9 月 25 日的数据。

测试锁定：tests/test_gics.py（57 个测试，其中 12 个针对计算本身，用的是足以手算核对的价格）、tests/test_periods.py（12 个，包括一年恰好在周末结束的情形）、一个 109 项检查的表格与象限图脚本，以及一个单独的、需用实时行情数据运行的官方指数比对脚本。