Do corporate insiders agree with the AI? Cross-referencing 12 months of insider buying with our 7-framework value-investing consensus.

Original data study: across 48 large-cap US stocks with detailed Form 4 coverage over the trailing 12 months, only 2 finished net-positive on open-market insider buying. Of the 12 stocks that pass all 7 named-investor value frameworks at a B-minus or better, exactly zero have net insider buying. The framework with the highest correlation to insider selling is Peter Lynch's at +0.240. The intersection set is a one-name list. Full numbers, methodology, and reproducibility notes.

This is an original data study. We took every US-listed stock that has at least one SEC Form 4 insider transaction in the invest-like.com production database over the trailing 12 months (49 large-cap tickers as of 26 May 2026), reconstructed net open-market insider buying versus selling for each one, and cross-referenced the result against the same seven value-investing framework scores we used in our 7-framework consensus study and our halal intersection study earlier this week.

The question we wanted to answer is the one almost every retail value investor asks at some point. Corporate insiders are the people closest to the business. The 7-framework AI consensus is supposed to identify the highest-quality compounders in the universe. Do those two signals point at the same stocks?

The data says no. They point in essentially opposite directions on the cohort where we can measure both.

Here is the full study.

TL;DR

• Of the 12 stocks that pass all 7 value-investing frameworks at score >= 60 and have insider-trade coverage in the last 12 months, exactly zero have net-positive open-market insider buying. All 12 are net sellers. The median net dollars across the group is negative $30.7 million and the mean is negative $96.8 million per ticker.
• Of the 49 unique tickers with Form-4 coverage in the trailing 12 months, only 2 are net-positive buyers. Tesla (+$742M, entirely from one Elon Musk open-market purchase block on 12 September 2025) and Berkshire Hathaway Class B (+$251k, two trivial trades). Every other ticker, including every name that scores well on our framework consensus, is net sell.
• Correlation between framework score and the share of insider activity that is selling is positive across all 7 frameworks. The highest is Peter Lynch's at +0.240, followed by Charlie Munger at +0.229 and Philip Fisher at +0.223. Graham's defensive-investor screen has the lowest correlation at +0.072. In plain English: the higher a stock scores on our value-investing AI consensus, the more aggressively its insiders are selling.
• The "named stocks with both signals positive" list has one entry. Tesla. It scores 1 out of 7 on the framework consensus (it passes only Graham, on price-history mean reversion grounds, and fails the other six). It is, by some distance, the most counter-intuitive cross-signal name in the dataset.

The rest of the post shows the methodology, the per-framework correlations, the full intersection set, the sector breakdown, the data caveats, and how anyone can reproduce these numbers through the public API.

Dated snapshot: the underlying Form 4 transactions and framework scores were last refreshed on 26 May 2026. Citations of specific counts below are stable against that snapshot. The insider-trade table covers 49 high-volume tickers in detail; broader Form 4 coverage of the full 12,500-ticker universe is on the roadmap but is not yet ingested. Any conclusions in this post are bounded to the 49-ticker cohort.

Methodology

What SEC Form 4 insider data is

Form 4 is the SEC filing that corporate insiders (directors, named executive officers, and beneficial owners of more than 10 percent of a class of registered equity) must submit within two business days of a transaction in their company's securities. It is the most timely public record of insider buying and selling in US equities. Each filing reports the transaction type, the date, the number of securities transacted, the per-share price, and the identity and role of the reporter.

The transaction-type field is the most informationally dense column in the filing. The codes break down into open-market activity (P-Purchase, S-Sale) and non-discretionary or compensation-related activity (A-Award, M-Exempt for option exercise, F-InKind for tax withholding, G-Gift, C-Conversion, D-Return). For a value-investing signal, only the open-market P-Purchase and S-Sale codes carry direct conviction information. The award and exercise codes do not, because the insider did not choose the timing of the share creation; the compensation committee did.

Peter Lynch's much-cited line on this is that insiders sell for any number of reasons, but they buy for only one. Our Form 4 explainer goes deeper into why open-market purchases are a higher-conviction signal than open-market sales, and how to filter automatic 10b5-1 sales out of the noise. The present study restricts the signal to net P-Purchase dollars minus net S-Sale dollars over the trailing 12 months, deliberately ignoring A-Award, M-Exempt, F-InKind, G-Gift, and other non-discretionary codes.

What the 7-framework consensus is

We score every stock against seven independently implemented value-investing frameworks: Warren Buffett's quality-and-moat lens, Benjamin Graham's defensive-investor screen, Philip Fisher's growth-quality test, Peter Lynch's growth-at-a-reasonable-price thesis, Joel Greenblatt's Magic Formula, Charlie Munger's mental-models filter, and Terry Smith's Fundsmith framework. Each framework returns a 0-to-100 score, an A through F letter grade, and a verdict.

A stock "passes a framework" at score >= 60 on the 0-to-100 scale, which maps to a B-minus or better letter grade. That is the same threshold we used in the 7-framework cross-consensus study and the halal intersection study that this post builds on. The full methodology hub is at /methodology/.

Cohort definition

The 49-ticker cohort in this study is the intersection of two filters.

The first filter is "has at least one Form 4 transaction with non-null price, date in the last 12 months, in the insider_trades table at snapshot 26 May 2026". That is the universe where we can measure open-market insider activity at all. The table is currently focused on the top 50 highest-coverage US tickers by trading volume, which is why this study is bounded to large-caps and not the full 12,543-stock universe.

The second filter, used for the framework-correlation section, is "has at least one score in the strategy_scores table for the relevant framework". 48 of the 49 insider-cohort tickers are scored on Buffett, Graham, Greenblatt, and Lynch. 49 of 49 are scored on Munger, Smith, and Fisher (with the BRK class-share dedupe applied). The full 7-framework score profile is available for all 49.

Throughout the post, BRK.A and BRK.B are reported separately for Form 4 purposes (the Form 4 filings name the share class) but reported under a single Berkshire framework score, because the underlying business is the same.

What "net dollars" means here

Net dollars = sum of P-Purchase securities-transacted multiplied by price, minus sum of S-Sale securities-transacted multiplied by price, summed across all reporters for the ticker over the trailing 12 months. Positive means net buying. Negative means net selling. A-Award, M-Exempt, F-InKind, G-Gift, C-Conversion, D-Return, and other non-discretionary codes are excluded. Reporting bias against compensation-grant noise is the point of this filter.

Headline finding 1: of the 12 stocks that pass all 7 frameworks and have insider data, zero are net buyers

The most surprising number in this study, by some distance, is the count at the all-seven-frameworks-pass threshold. There are 12 tickers in the 49-cohort that pass all 7 value-investing frameworks at score >= 60. Every single one of them is a net seller in the last 12 months.

The median net-dollars across the 12-ticker all-7-pass cohort is negative $30.7 million. The mean is negative $96.8 million. Not one company in this group had an open-market insider purchase outweigh its open-market insider sales over the year.

The intuition for why this happens is fairly clean. The value-investing consensus is built to flag businesses with durable margins, strong ROIC, low debt, and reasonable valuation. When all seven frameworks agree on the same stock, the underlying ticker is almost always a large-cap quality compounder whose price has already moved a long way to reflect those properties. Insiders at those companies tend to receive large stock-based compensation grants, often have founder-era position holdings, and have rational portfolio-diversification reasons to sell into strength. They are not selling because they think the business is bad. They are selling because their personal balance sheet is over-concentrated in a single position that has already won.

In the next section we widen the lens to all 49 tickers to confirm that the pattern is not specific to the highest-consensus group.

Headline finding 2: only 2 of 49 large-cap tickers had net insider buying

Across the full insider-coverage cohort, the picture is even more lopsided than the all-7-pass slice suggested.

The two net buyers are:

The Tesla figure dominates the cohort, but it is structurally a single insider's single concentrated buy event, not a broad insider-buying signal. There is no second director or officer also accumulating. Strip Musk and the data point disappears.

The BRK.B figure is small enough that it is statistical noise. Two purchases totalling a quarter of a million dollars across a $535B market cap company is not a directional signal.

The remaining 46 tickers all printed net selling over the trailing 12 months, with WMT setting the upper bound at -$1.30B and CVX, LLY, AVGO, and PLTR all clearing -$200M on the way down.

The flat ticker is AZN (AstraZeneca), where the Form 4 history we have in scope shows zero P-Purchase and zero S-Sale rows in the window. AZN is European-headquartered with a US ADR listing, so its Form 4 footprint is structurally thinner than a US-domiciled company of similar size. We exclude AZN from the correlation arithmetic below to avoid noise.

Headline finding 3: every framework correlates positively with insider selling

If insiders are systematically selling the same stocks the framework consensus rates highest, the per-framework correlations should all be positive against an "intensity of selling" metric. They are.

We define sell-share for each ticker as S-Sale dollars / (P-Purchase dollars + S-Sale dollars) over the last 12 months. A ticker whose entire open-market insider activity is sales has sell-share = 1.0. A ticker whose entire activity is buys has sell-share = 0.0. The metric is bounded and stable for tickers with material activity (we exclude AZN and BRK.B from the correlation calculation because both have near-zero activity dollar totals).

The Pearson correlation between each framework's 0-100 score and sell-share, across the cohort:

None of these correlations are large in absolute terms. Pearson r = 0.240 means roughly 6 percent of the variance in sell-share is linearly explained by Lynch's score. The cohort is also only 47 active tickers (49 minus AZN and BRK.B), which is a small sample by statistical standards. But the sign of the correlation is positive for every single framework. That is the structural finding. There is no framework in our consensus where a higher score is associated with more insider buying. Every one of the seven is at least mildly aligned with the direction of insider selling.

Why is Graham the weakest? Graham's defensive-investor framework includes price-to-book and current-ratio criteria that down-rank capital-light software and platform names (the prototypical large-cap compounder where insider grants are heaviest). Many of Graham's highest-scoring names in the 49-cohort are commodity-cyclical (CAT, TXN, ORCL, ABBV, AMAT, NVDA) where insider activity is less driven by software-style grant programmes. The result is that Graham scores correlate less tightly with sell-share than the more quality-oriented frameworks do.

Headline finding 4: the named intersection set is a one-stock list

The premise of the post is to find the names where both signals are positive: high framework consensus AND positive net insider buying. The intersection set is one name.

Tesla is the only ticker in the 49-cohort with both positive net insider dollars and at least one framework pass. It passes only Graham, and even Graham only because the implementation of the defensive-investor screen rewards price-history mean reversion (Tesla's market cap has compressed from its 2021 peak), not because the underlying balance sheet or earnings profile actually meet Graham's classical defensive-investor criteria. Tesla fails Buffett (50, grade D), fails Lynch hard (22), fails Greenblatt (40), fails Munger (33), fails Smith (39), fails Fisher (34).

The structural meaning of the one-name intersection: a 7-framework AI consensus is essentially never going to agree with insider open-market buying in a large-cap US universe, because the two signals select for different conditions. The frameworks select for businesses that have already won and are still compounding. Insider open-market buying clusters at businesses where the insiders think the market is mispricing them downward (often distressed or out-of-favour names), or at idiosyncratic founder-led companies where the founder is voluntarily concentrating personal capital. Those two populations almost never overlap in any given 12-month window.

If we relax the cross-signal definition from "passes at least one framework" to "passes any of the seven at any threshold", the intersection set is still effectively Tesla. BRK.B has positive net dollars but the dollar amount is trivial and Berkshire fails every framework at score >= 60. There is no second name to add.

Sector breakdown of the insider-coverage cohort

For context, here is the sector tilt of the 49-ticker cohort against the unfiltered universe. The cohort is, by construction, biased toward mega-cap US names that have heavy Form 4 reporting activity.

The cohort is more concentrated than the universe: nine sectors versus the universe's fifteen, and the top sector (Technology) makes up roughly a third of cohort weight. That is a function of the data ingestion pipeline (high-volume large-caps were prioritised) and not a property of insider behaviour. As Form 4 coverage broadens in subsequent releases, we expect the sector mix to drift toward universe weights.

The interesting comparison is not the cohort-vs-universe sector tilt (which is mostly a coverage artefact), but the cohort-vs-consensus-cohort tilt. The 7-framework consensus list of 9 names that pass all seven at score >= 60 has a different sector mix from the 49-cohort: Technology dominates even more heavily (KLAC, FTNT, LRCX, MA, TSM in the all-7-pass list overlap with the cohort), and Healthcare punches above its universe weight via LLY and IDXX. Financial Services in the all-7-pass list is narrowly populated by payment-network and exchange-and-data names (V, MA), not by traditional banks. Banks, when measured by insider buying, are not heavy on the buy side; they are heavy on the sell side, especially Goldman Sachs (-$64.6M) and Morgan Stanley (-$65.7M) over the year.

What this means for investors

Three observations, framed as observations and not advice.

Observation 1: insider buying and AI consensus point in different directions on this cohort. A value-investing practitioner who believes both signals matter has to reconcile them stock by stock. They cannot both be triggering on the same name without idiosyncratic context (founder concentration, deeply contrarian buyer, or a special situation). The Tesla data point is the canonical exception that proves the rule: a single founder bought $853M of his own company in one day, against a backdrop of a 1-of-7 framework pass and a Buffett-Fit grade of D. Most investors do not have the conviction of a founder, and a framework cohort with a Buffett-Fit of D is not what the consensus screen is designed to surface.

Observation 2: insider selling in mega-cap quality compounders is not the bearish signal retail investors often think it is. When AAPL prints -$143M of insider sales over a year, the dollar amount is meaningful in absolute terms but small relative to the company's $4.5T market cap and to the executive base's stock-based compensation flow. The same applies to NVDA, AVGO, LLY, and every other large-cap name in the 7-framework consensus list. Insiders are diversifying personal balance sheets that have been overweight a single position for years, often as part of pre-arranged 10b5-1 plans. Reading those flows as "the smart money is leaving" usually overweights the signal. The framework consensus is the more reliable input.

Observation 3: when both signals do align, the alignment is significant. A stock that an open-market insider is genuinely accumulating with their own cash, that simultaneously scores well across the 7 frameworks, would be an unusual and high-conviction co-occurrence. The data set since launch has not produced one. That fact is itself the headline. If and when a name emerges that passes the framework consensus AND has multi-officer open-market accumulation that is not explained by 10b5-1 unwind or founder concentration, it will deserve a closer look.

This is not investment advice. The frameworks, the insider data, and the cross-correlations all reflect public reporting and documented methodology. Whether the methodologies still produce excess returns in current conditions is an open empirical question that this post does not attempt to answer.

How to reproduce this study

Everything in this post is reproducible from the public surface of invest-like.com.

• Per-stock 7-framework verdict: /api/public/verdict/[ticker] returns the full per-framework breakdown (score, grade, verdict, per-criterion pass/fail) in JSON. No authentication required. Try /api/public/verdict/NVDA, /api/public/verdict/AAPL, or /api/public/verdict/LLY.
• Cross-framework consensus endpoint: /api/public/consensus returns the top stocks by consensus score across all 7 frameworks.
• Latest verdicts endpoint: /api/public/latest-verdicts returns recent framework verdicts in machine-readable form.
• Methodology hub: /methodology/ documents each value framework's criteria and scoring function.
• Form 4 background: /blog/insider-buy-signal-form-4/ covers the SEC Form 4 filing format itself in plain English and explains why insider buys are a higher-information signal than insider sells.
• Prior data study: /blog/12500-stocks-7-frameworks-cross-framework-consensus/ is the universe-wide 7-framework consensus study this post builds on.
• Bulk machine-readable index: /llms-full.txt contains a flat dump of the most important fields for LLM consumption.

If you replicate the methodology and find different numbers, please email zaid@invest-like.com with your methodology and we will publish a correction or a comparison post.

FAQ

Why is the cohort only 49 tickers when the full universe is 12,543? Because the insider_trades table in production today only ingests detailed Form 4 transactions for the top US large-caps by trading volume (essentially the S&P 100 plus a few high-interest names). Full Form 4 ingestion for the 12,500-stock universe is on the roadmap but is not yet live. Bounding the study to the 49 high-coverage tickers is honest: if we extrapolated to micro-caps and ADRs we would be inventing data. When coverage broadens, we will re-run this study and update the post.

Did you exclude 10b5-1 sales? Doesn't that make the "insider selling" number noisy? Partially. The Form 4 filings do flag 10b5-1-related transactions in narrative footnote text, but the production database does not parse the footnote field into a structured flag in the current schema. So this study treats every S-Sale equally. A future iteration of the study would split S-Sales into "discretionary" and "pre-arranged 10b5-1" and re-run the correlations. We expect the discretionary subset to be smaller in dollar terms but to correlate more cleanly with negative sentiment. As-is, the headline finding (the framework consensus and insider buying point in different directions) is still robust because the 7-framework cohort prints zero net buyers across all 12 tickers, regardless of whether the sales were 10b5-1 or not.

Why does TSLA show $742M of net insider buying when Tesla is famously a heavy insider-selling story? Because of a single event: Elon Musk personally bought roughly 2.2 million shares of TSLA on 12 September 2025 at an average price of $390.66, in 13 separate filings across the trading session, totalling approximately $853M of P-Purchase volume. Net of the $111M of other Tesla S-Sale activity that occurred during the rest of the trailing 12 months, the net figure is +$742M. That is one founder's idiosyncratic capital allocation decision, not a broader insider buying signal at the company. Strip Musk and Tesla's net figure flips negative.

Why does Graham have the lowest correlation with insider selling? Because Graham's defensive-investor framework is the most explicitly statistical of the seven and least quality-tilted. It rewards low price-to-book, high current ratio, long dividend history, and earnings stability. Many large-cap software and platform names (where insider selling and stock-based comp are heaviest) score poorly on Graham because of the price-to-book criterion, while many cyclical or commodity names with cleaner balance sheets score well on Graham despite less insider selling. The result is a weaker correlation. The quality-tilted frameworks (Lynch, Munger, Fisher, Smith) correlate more cleanly because they select for the same large-cap compounder population where insider grants drive the most sell flow.

Are these numbers going to change? Yes. Form 4 filings update daily, insider transactions roll into and out of the trailing 12 months, and framework scores refresh as fundamentals get re-ingested quarterly. The exact dollar figures will drift. The structural finding (essentially no overlap between the 7-framework consensus and net insider buying in the large-cap US cohort) has been stable across every snapshot we have observed, and the direction of the per-framework correlations has been stable as well.

Can I cite these numbers? Yes. Cite as invest-like.com 2026, "Do corporate insiders agree with the AI? Cross-referencing 12 months of insider buying with our 7-framework value-investing consensus", snapshot 26 May 2026. Link target is this page. The underlying data is publicly accessible through the API endpoints listed above, and the licence terms are at /press/. If you want a copy of the underlying snapshot CSV for academic use, email zaid@invest-like.com.

Educational disclaimer

This post is an empirical study of how SEC Form 4 insider transaction data and seven documented value-investing frameworks overlap on a 49-ticker cohort of large-cap US equities. Nothing in this post is a recommendation to buy or sell any security. The framework scores are our implementations of published methodologies and may differ from any individual investor's interpretation of the underlying texts. The Form 4 data is reported by issuers and may contain reporting errors that the SEC EDGAR system corrects after the fact. Readers should consult a licensed financial advisor for personalised investment advice.

Data snapshot: 26 May 2026. Methodology hub: /methodology/. Source data: /llms-full.txt. Public API: /api/public/openapi.json.