公司內部人士買入自家股票後，股價會上升嗎？對 47,458 宗公開市場購買的測試

An insider buying shares of their own company on the open market, with their own cash, is about the cleanest signal in public filings. I measured what followed 47,458 of those purchases: every Form 4 open-market buy on 3,859 US stocks from January 2020 to August 2025, grouped into one event per stock and disclosure day. The answer depends almost entirely on what you compare against, and on one test most versions of this study skip.

Same purchases, two answers: 13 points behind the index after a year, or 2.6 points ahead of the typical listed stock. Neither is really about insiders. Insiders mostly buy small companies, and from 2020 to 2025 the median listed stock went nowhere while the index rose about 18% a year.

Five findings

1. The timing adds nothing. I shifted every purchase's measurement window 180 trading days into the future and measured again. The one-year excess over the median stock came out at 2.78% instead of 2.61%, and the event-by-event difference (1.40 points, t = 1.58) isn't distinguishable from zero. At one, three and six months it's flat. These companies beat the median stock whether or not an insider had just bought. Insider buying identifies a type of company, not a moment.

2. Big cheques are followed by worse returns. Sorted into ten groups by dollar value, the largest (median $3.5 million) trailed the smallest by 3.3 points over three months, 7.8 over six months and 17.1 over a year. In the shifted windows the same spread was -0.6, +0.4 and -2.0 points, none significant. Unlike the first result, this one belongs to the purchase window.

3. It's the top group, not a gradient. Nine of the ten size groups had a positive one-year excess return, from 0.7% to 4.9%. Only the largest was negative, at -1.8%. 41% of that top group is 10% holders buying with no officer or director involved, against 7% of the smallest. Removing every 10%-owner-only purchase and re-sorting still leaves the largest group 13.4 points behind the smallest after a year.

4. A purchase after a long silence isn't special. 858 purchases ended a company-wide buying drought of more than two years. Their median one-year return was 0.74%: ahead of the weak median stock, behind the median S&P 500 member (7.50%). The same shortfall shows up in the shifted windows, so it's neither a buy signal nor a short signal.

5. The routine vs opportunistic split didn't reproduce. Cohen, Malloy and Pomorski (2012) found the signal sits with opportunistic insiders rather than those who trade on a calendar. On this data the excess return was larger among routine traders (9.55% vs 4.35% after a year under a strict rule), and neither group showed a timing premium. One caveat: three prior years of trading history per person only exists from 2023 in my data, so this cut covers fewer events.

The traps I hit

• Medians, not means. One-year returns have a skew of 15. The mean purchase gained 19.36%, the median 2.36%, and 90 of the 47,458 returned more than 1,000%.
• Don't average microcaps for a benchmark. The mean 21-day return across all listed stocks came out at +137%, because unrecorded reverse splits look like gains of millions of percent. A daily-rebalanced equal-weighted index was worse, since it also harvests the bid-ask bounce. A buy-and-hold median over the same window avoids both.
• Count events, not filing lines. A CEO buying on three days in one Form 4, or four directors filing on the same day, is one event. Counting lines separately splits one big purchase into several medium ones and flattens exactly the size effect.
• Enter at the next close. Most Form 4s are accepted after the market closes, so the disclosure day's close is a price nobody could have paid.
• Count days on the market calendar. One stock printed a bar on July 3, 2026, when the market was closed. Building benchmark windows from the dates any stock printed shifted 131 one-year windows by a session.
• Don't use the window before the purchase as the placebo. Insiders tend to buy after declines, so that window is depressed and would invent a timing effect.
• Survivorship runs one way. Companies delisted before my price store was built are missing, so the level numbers are biased upward. The two main results compare the same stocks against themselves, or purchases against purchases, so most of that bias cancels there.

Check one company

With the Equibles MCP server connected to Claude or ChatGPT, a prompt like "Using Equibles, list the open-market insider purchases at Intel over the last two years, with who bought and the dollar amount" pulls the Form 4 rows, and you can see which side of the size table a purchase falls on.

The full study has the decile tables, the two-year results (where the size spread reaches t = -5.40 against a flat placebo), the role breakdown and a reproducible specification: What Happens After an Insider Buys? Evidence From 47,458 Open-Market Purchases.