Insider Buying vs Selling: Which Is the Stronger Signal?

The asymmetry: why a buy says more than a sale

The information content of an insider trade is not symmetric, and the reason is structural rather than statistical. An executive's wealth is already concentrated in the company: salary, bonus, restricted stock, options, reputation, all riding on the same ticker. Selling some of that exposure is the financially rational default, and a sale is compatible with a hundred motivations, most of them boring.

A purchase inverts that logic. When an insider takes after-tax cash and voluntarily increases an already concentrated exposure, there are very few innocent explanations. The insider knows the pipeline, the customer demand, the tone of the last board meeting. Choosing to concentrate further is a costly action, and costly actions are credible in a way that cheap talk never is.

There is a second structural asymmetry: supply. Insiders are granted equity constantly, so the population of sales is enormous and mechanically driven. Open-market purchases are rare by comparison, which means each one carries more selection. In Arkolith's dataset of more than 51,000 insider transactions sourced directly from SEC filings, sales-type events vastly outnumber true open-market buys. Rarity alone is not a signal, but it concentrates whatever signal exists into a small, watchable set. You can browse this per ticker, for example on the insider page for TSM.

What the academic research generally finds

The insider-trading literature is one of the older, more replicated corners of market-anomaly research, and the broad shape of its findings has been stable for decades. Academic work has generally found that stocks bought by insiders earn positive abnormal returns over the following months, that the effect is stronger for purchases than for sales, stronger in smaller and less-covered companies, and stronger when the buyer is a senior officer rather than an outside director. Studies that condition on sales alone have generally found weak or no predictive power once mechanical selling (option-related and scheduled trades) is removed.

A few consistent refinements are worth internalizing. First, the signal decays: the abnormal-return estimates are concentrated in the months after the trade, not years. Second, the effect has generally survived public disclosure: outsiders reacting to the filing (not the trade itself) still captured some of it on average, though less than the insider did. Third, magnitude matters: large purchases relative to the insider's existing holdings have generally looked more informative than token buys. These are average effects across thousands of events, not guarantees about any single trade.

The honest summary: insider buying is among the better-documented signals in public-filings data, insider selling is mostly noise, and that spread is exactly why naive "insider activity" feeds mixing the two are close to useless. For the adjacent question of how trustworthy quarterly institutional filings are, see how accurate is 13F data.

Cluster logic: one buy is an opinion, three buys are a meeting

A single purchase is one person's judgment, possibly idiosyncratic. When several insiders at the same company buy within a short window, the probability that you are looking at private optimism rather than personal portfolio quirks rises sharply. Academic and practitioner work has generally found clustered buying to be more informative than isolated buying, and it is the organizing idea behind screens like our insider buying clusters for June 2026.

A practical scoring rubric:

Two cautions. First, clusters can be choreographed. Companies know investors watch Form 4s, and a round of small, simultaneous director purchases after bad news is sometimes investor relations by other means. Size relative to each insider's net worth and existing stake is the antidote: theater is cheap, conviction is expensive. Second, cluster detection is sensitive to entity resolution: the same human can appear under multiple filer records. That is a data-quality problem before it is a finance problem, which is why every Arkolith transaction resolves to a specific filer and SEC accession number rather than a scraped name string.

Form 4 mechanics: the data behind the signal

The reason insider data is tradable at all is speed. Under Section 16, insiders (officers, directors, and 10%+ owners) must report transactions on Form 4 within two business days of the trade. That is often fast enough to matter, particularly in smaller names. The related forms have different clocks and different jobs:

The SEC's primer on Forms 3, 4, and 5 covers the legal mechanics; investor.gov has a plain-English glossary entry. For how these forms relate to the quarterly 13F (institutional, slow, long-only, due 45 days after quarter end), see 13F vs 13D vs 13G vs Form 4.

Inside a Form 4, the field that matters most is the transaction code. Code P is an open-market purchase, the event the whole literature is built on. Code S is an open-market sale. Codes like M (option exercise), F (shares withheld for taxes), A (grant or award), and G (gift) are mechanical events that a signal pipeline should label and usually exclude. Many "insider selling" headlines are actually M plus F plus S sequences: an exercise-and-cover that says nothing about the executive's view.

Filtering the noise before you score anything

A useful insider-buying signal is mostly a sequence of exclusions. Scheduled trades: Rule 10b5-1 plan transactions are committed in advance precisely so they do not reflect current private information, and filings indicate plan-based trades. Mechanical events: exercises, withholdings, grants, conversions, and gifts, identifiable by transaction code. Token purchases: small lots bought the day after bad news are often theater, so size every buy against the insider's existing stake. Habitual buyers: some insiders buy small amounts every month regardless of conditions; a regular pattern carries less event information than a first purchase in years.

On the sell side, near-total exclusion is reasonable for signal purposes, with one family of exceptions: abnormal selling. A founder's first-ever sale, several senior officers selling unusually large stake fractions in the same window, or discretionary sales breaking a long pattern are the rare sale events generally treated as worth examining. Even there, the hedged framing matters: selling clusters are suggestive, not damning, because liquidity events (lockup expiries, secondary offerings, tax deadlines) cluster too.

The last filter is corroboration: a purchase cluster is more interesting when institutional money is moving the same direction. Cross-referencing Form 4 buys against the 13F record (Arkolith's Q1 2026 dataset covers 1,824 institutional filers and 1.87 million long positions worth $53.7 trillion in reported value) lets an agent ask whether smart money agrees, using the investor leaderboard or a specific fund's filing history.

Querying insider data programmatically

This entire workflow (pull transactions, classify codes, detect clusters, cross-reference institutions) is the kind of multi-step, source-grounded task that agents do well and humans do tediously. Arkolith exposes the data over REST and MCP with every datapoint carrying provenance back to its SEC EDGAR accession number, which matters when an LLM is in the loop: a model that can cite the filing cannot quietly invent the trade. (More on that failure mode in stopping AI from hallucinating market data.)

Find an entity to anchor on:

curl -H "Authorization: Bearer YOUR_KEY" \
  "https://arkolith.com/api/v1/search?q=taiwan+semiconductor"

Then pull institutional positioning to cross-reference against insider activity, using the fund's CIK:

curl -H "Authorization: Bearer YOUR_KEY" \
  "https://arkolith.com/api/v1/funds/0001067983/holdings"

The same surface is exposed as MCP tools, so a Claude agent can run the whole loop conversationally: search the entity, pull the transactions, exclude the mechanical codes, flag the clusters, and cite the accession number for every claim it makes. Keys are prepaid and metered per call, so a cluster-detection job costs what it uses. Setup takes a few minutes via the quickstart, and the full tool and endpoint reference lives in the docs.

Frequently asked questions about insider buying signals

Is it legal for insiders to buy their own company's stock?

Yes, provided they are not trading on material non-public information and they disclose the trade. Section 16 requires officers, directors, and 10%+ owners to report transactions on Form 4 within two business days. The disclosure requirement is exactly what makes the signal publicly available.

How quickly does an insider purchase become public?

The Form 4 must be filed within two business days of the transaction, and it appears on SEC EDGAR essentially immediately upon filing. That short lag is why insider data is considered actionable, unlike 13F holdings, which arrive up to 45 days after quarter end.

Does insider selling ever carry a signal?

Occasionally, but only when it is abnormal: a founder's first-ever sale, multiple senior officers liquidating unusual fractions of their stakes in the same window, or discretionary sales breaking a long pattern. Routine, scheduled, or option-related selling has generally shown little predictive power in academic work and should be filtered out, not scored.

How many insiders does it take to make a cluster?

There is no official threshold, but practitioners commonly treat three or more distinct insiders buying within roughly a month as a cluster. The logic is independence: each additional discretionary buyer is a separate judgment drawing on the same private information, which is harder to explain away than a single purchase.

This article explains public filings and data concepts. It is not investment advice.