Is a buyback announcement a buy signal? I tracked 430 recent ones against SPY

A buyback announcement reads like a company putting its own money behind its shares. I checked what actually followed recent ones: the first verified buyback announcement from each of 430 US-listed companies between July 2025 and June 2026, taken from the company’s own SEC filings, with each stock measured against SPY over the same dates.

股票回购公告读起来就像一家公司在用自己的钱支持其股价。我核查了近期公告后的实际表现：选取了 2025 年 7 月至 2026 年 6 月间 430 家美国上市公司的首次经核实回购公告，数据取自公司自身的 SEC 文件，并将每只股票在同期与标普 500 指数（SPY）进行了对比。

After Typical stock SPY Median excess Mean excess Beat SPY 5 sessions 1.24% 0.40% +0.76 pts +1.51 pts 54.5% 20 sessions 1.46% 0.74% +1.02 pts +1.85 pts 53.1% 63 sessions 5.26% 7.10% -0.92 pts +2.00 pts 47.9% The 5- and 20-session rows cover 431 companies; one hadn’t yet traded 63 sessions after its announcement.

（表格数据：5 个交易日后，典型股票 1.24%，SPY 0.40%，中位数超额收益 +0.76 个百分点，平均超额收益 +1.51 个百分点，跑赢 SPY 比例 54.5%；20 个交易日后，典型股票 1.46%，SPY 0.74%，中位数超额收益 +1.02 个百分点，平均超额收益 +1.85 个百分点，跑赢 SPY 比例 53.1%；63 个交易日后，典型股票 5.26%，SPY 7.10%，中位数超额收益 -0.92 个百分点，平均超额收益 +2.00 个百分点，跑赢 SPY 比例 47.9%。5 个和 20 个交易日的行涵盖了 431 家公司；其中一家公司在公告后尚未交易满 63 个交易日。）

Five findings 1. There’s a small, uneven pop. The average stock beat SPY by 1.5 points over five sessions and 1.8 over 20. The typical stock did about half that, and barely more than half of them beat the index at all. 2. By three months the typical stock lags. After 63 sessions the median stock trailed SPY by 0.92 points and only 206 of 430 beat it.

五项发现：1. 存在小幅且不均衡的上涨。平均而言，股票在 5 个交易日内跑赢 SPY 1.5 个百分点，20 个交易日内跑赢 1.8 个百分点。典型股票的表现约为上述数据的一半，且仅有略多于一半的股票跑赢了指数。2. 三个月后，典型股票开始落后。在 63 个交易日后，中位数股票落后 SPY 0.92 个百分点，430 家公司中仅有 206 家跑赢。

3. One stock carries the average. A single company gained more than 300% against SPY in 63 sessions, which alone adds about 0.8 points to the mean. Trimming the top and bottom 5% cuts the 63-session mean from 2.00% to 0.66%; the median stays negative. 4. The median’s confidence interval crosses zero every time. Bootstrapped 10,000 times, the median excess return’s 95% interval was -0.01% to 1.68% at five sessions, -0.45% to 2.57% at 20 and -2.70% to 1.06% at 63. The mean’s interval excludes zero at five and 20 sessions, which is the honest summary: a basket reacts, a single pick is a coin flip.

4. 单只股票拉高了平均值。有一家公司在 63 个交易日内相对于 SPY 涨幅超过 300%，仅此一项就为平均值贡献了约 0.8 个百分点。剔除表现最好和最差的 5% 后，63 个交易日的平均值从 2.00% 降至 0.66%；中位数则保持负值。4. 中位数的置信区间始终跨越零点。经过 10,000 次自助抽样（Bootstrap），中位数超额收益的 95% 置信区间在 5 个交易日时为 -0.01% 至 1.68%，20 个交易日时为 -0.45% 至 2.57%，63 个交易日时为 -2.70% 至 1.06%。平均值的区间在 5 个和 20 个交易日时排除了零，这得出了一个诚实的总结：一篮子股票会有反应，但挑选单只股票就像掷硬币。

5. Authorization isn’t purchase. A board authorization is a ceiling, not an obligation. Older research found the operating improvement after an announcement concentrated in companies that actually bought shares in the same quarter, and a study of 1994 to 2014 announcements found the long-run drift much weaker after the early years.

6. 授权不等于购买。董事会授权只是一个上限，而非义务。早期的研究发现，公告后的经营改善主要集中在当季实际回购股份的公司中；而一项针对 1994 年至 2014 年公告的研究发现，长期趋势在早期之后变得非常微弱。

The traps I hit: No hindsight from late filings. An event counted only if the filing stated the announcement date and arrived within 14 days of it. Otherwise a 10-Q mentioning an old program months later sneaks information into the past. One event per company. Serial announcers would otherwise dominate the sample. Exact listings and split-safe windows. Each ticker resolved to its exact listed security, only traded closes with volume were used, and no window crosses a stock split. Start before the news. Returns run from the last close before the announcement date, because day-level dates can’t tell a before-open release from an after-close one. Report the median. With one +300% outlier in 430 stocks, the mean describes a basket you’d never hold.

我遇到的陷阱：避免事后诸葛亮。只有在公告日期后 14 天内提交的文件才会被计入。否则，一份在数月后提及旧计划的 10-Q 季报会将信息“潜入”过去。每家公司仅计一次事件，否则频繁公告的公司会主导样本。精确的上市信息和避开拆股窗口。每个代码都对应其确切的上市证券，仅使用有成交量的收盘价，且时间窗口不跨越股票拆分。从新闻发布前开始计算。收益从公告日前最后一个收盘价开始计算，因为按天计算无法区分开盘前还是收盘后发布。报告中位数。在 430 只股票中出现一个 +300% 的异常值时，平均值所描述的篮子是你永远不会持有的。

What to check instead of the headline: How big is the authorization against the market value? Did the company actually buy shares in the following quarters? Is the diluted share count falling, or are buybacks just offsetting stock compensation? Is it paid from free cash flow or new debt, and at what valuation?

除了标题，你应该检查什么：授权金额相对于市值有多大？公司在随后的季度中是否真的购买了股票？稀释后的股份总数是在下降，还是回购仅仅抵消了股票薪酬？回购资金是来自自由现金流还是新增债务，以及估值如何？

With the Equibles MCP server connected to Claude or ChatGPT, a prompt like “Using Equibles, show Apple’s buyback authorizations since 2022, how much it actually spent each year, and its diluted share count over the same period” answers the second and third questions for one company. The full study has the charts, the bootstrap table and the literature it builds on: What Happens After Stock Buyback Announcements? Evidence From 430 Companies.

通过将 Equibles MCP 服务器连接到 Claude 或 ChatGPT，使用类似“使用 Equibles，显示苹果公司自 2022 年以来的回购授权、每年实际支出金额及其同期的稀释后股份总数”的提示词，即可回答单家公司的后两个问题。完整的研究报告包含图表、自助抽样表以及所依据的文献：《股票回购公告后会发生什么？来自 430 家公司的证据》。