Methodology

The raw source is Form 8-K filings on SEC EDGAR, specifically Item 5.02 (departure and appointment of officers and directors). US SEC registrants must file an 8-K within four business days of a decision to change their chief executive, so every such change at a registrant should appear there. The current study uses 8-K filings from 28 October 2021 onwards.

Foreign companies listed in the US through ADRs report on Form 6-K, which has no Item 5.02, and are not covered. Item 5.02 also covers CFOs, directors and pay, so each filing is screened for a change in the chief executive role. Classification and all analysis are our own.

Text rules locate the Item 5.02 section of each filing and look for wording that announces a change to the CEO role, such as “resigned as Chief Executive Officer” or “appointed … as Chief Executive Officer”. Mentions that describe no change (“will continue to serve as CEO”, a sitting CEO taking on another title), refer to another company or a subsidiary, or appear in biographies and pay tables are excluded, as are changes caused by the company being acquired.

The rules were tuned on random samples of filings and checked on fresh samples not used for tuning. On the latest check, about four in five events were genuine CEO changes. Missed changes are harder to measure; a fix to how the Item 5.02 section is located recovered about 300 events, including Apple's in April 2026, and a fuller test of misses is planned.

As a second check, a language model (Google Gemini) reads each filing and its press release, using only that text, and records whether it describes a CEO change and the key facts: how the CEO left, whether it was planned, who succeeds them and what other news came with it. It judges about 86% of events at listed companies to be genuine CEO changes. In hand checks, all 52 of the events it rejected were indeed not CEO changes, and 39 of 40 it accepted were genuine, so the checked dataset is about 97% accurate. Rejected events are excluded from the Events list and all averages but kept, labelled, for transparency.

Missed changes are measured against the news: CEO-change headlines from 2025 were matched to every listed company and checked against our events. This found several faults in how filings were read (for example, a later mention of “Item 5.02” being taken for the section itself). After fixing them, we estimate the rules find about 88% of CEO changes reported in the news, up from about 74%.

A company's follow-up filings within 90 days of a first announcement are merged into one event, and filings that only refer to an earlier announcement are dropped. Events at companies without shares listed on a US exchange, SPACs and funds are kept in the dataset but excluded from price analysis.

The 8-K is filed when a change is decided, not when it takes effect. OPAL Fuels, for example, disclosed on 21 October 2026 that its co-CEO will retire on 31 March 2027. Markets react to news when it becomes public, so events are dated by the first announcement, and the effective date has no role in the reaction measured here.

Companies often publish a press release on the day of the decision and file the 8-K the same day or up to four business days later, and news sometimes leaks earlier still. Including the day before the filing in one window, and reporting the run-up before it, guards against this; checking the date of the first news coverage is planned.

Day 0 is the first trading day on which the market could react to the filing. EDGAR's acceptance time is converted to New York time; filings accepted at or after the 4pm close count from the next trading day.

Normal returns come from a market model: each stock's daily return is regressed on the S&P 500 (SPY) over trading days −250 to −60. The window ends three months early so that a slide in the share price before a CEO is replaced is not learnt as normal. Beta is the slope of that regression (equivalently, the covariance of the stock with the market divided by the market's variance). Stocks with fewer than 120 days of data in the window are skipped.

Abnormal returns are the actual return minus the market model's prediction, summed over each window. They measure how the market revalued the company on the news, not the quality of the CEO.

Windows that have not finished, or that end because the shares stopped trading, are shown as returns “so far” and never enter averages. Headline averages use events priced at least $1 two days before day 0, and are reported with medians because a few extreme moves can dominate a mean.

Daily prices adjusted for splits and dividends come from Yahoo Finance for companies still listed and from Tiingo for companies that have since delisted, so that firms which later failed or were acquired stay in the sample. Tickers are taken from each filing's cover page as they were at the time, and a delisted ticker is only used if it was trading on the event date, so that a ticker later reused by another company is not picked up.

Headlines come from GDELT, an open index of world news, searched from the day before each filing to three days after for headlines naming the company and the CEO role. Only the headline, site, date and link are kept. Coverage is good for large companies and thin for small ones: about half of events have a matching headline.

Tone is scored with FinBERT, a language model trained on financial news, and the Loughran-McDonald finance word lists are applied to the company's own press release. Headlines seen before the market could react are kept separate from later ones, which often report the price move itself. In an early test, the tone of press releases and of early headlines added little to the facts in the filing (an interim CEO, an insider successor, results released at the same time) when predicting the reaction.

• Small samples: some subgroups (for example, forced departures with outside successors) contain few events, so estimates are noisy.
• Overfitting: testing many features and windows on the same sample raises the chance of finding effects that do not hold out of sample. Comparisons not set out in advance on the Research design page are treated as exploratory.
• A single event's return describes what happened around the announcement; it does not isolate the effect of the CEO change.