When the insiders and the news disagree: a first look at the cross-signal

When the insiders and the news disagree: a first look at the cross-signal

Tommi Johnsen and Svetlana Shasharina

What this is about

Two different sources each tell you something about where a stock is going. The first is insider trading filings: SEC Form 4, which executives, directors, and large shareholders are required to submit within two business days of buying or selling shares in their own company. The second is the news cycle: analyst reports, earnings coverage, breaking stories. Each source has its own academic literature, its own track record, and its own well-known failure modes. What is less studied is what happens when they disagree. When the CEO is buying their own stock at the moment the financial press is turning bearish on the company, which side is right?

This post reports a first look at that question using twelve days of data from a production pipeline that emits both signals on a common ticker-day grid. The window is small, and the conclusions are descriptive rather than statistical. But the qualitative pattern across the four cells of the natural contingency table is consistent enough to be worth describing.

Background: what this builds on

There is a sixty-year academic literature on insider trading filings as predictors of forward stock returns. The headline findings are consistent: insider purchases carry information, insider sales mostly do not, and coordinated buying by multiple insiders is stronger than individual transactions. The signal has been compressing over the decades as markets got faster and the data became more accessible, but it has not disappeared.

A 2026 working paper by Ozlen and Batumoglu decomposed the timing of the alpha and found that 70-80% of it lives in the unobservable window between the transaction date and the day the filing becomes public. By the time a daily pipeline can act on the signal, most of the move is already in the price. A previous post in this series documented this pattern at small scale on five days of production data: clusters of officer buying surfaced in the dashboard, but the corresponding stock moves had typically already happened by the time the dashboard caught them.

That earlier work also surfaced an operational issue not addressed in the published literature. The standard academic cluster definition (≥2 distinct insiders making non-plan purchases in a short window) catches programmatic compensation events alongside genuine opportunistic buying. A single example made this concrete: Taiwan Semiconductor showed up on May 19 with 28 distinct officer-CIK buyers, which on inspection turned out to be 27 of 28 officers buying on the same day at the same price ($71.82). This is the unambiguous signature of an Employee Stock Purchase Plan distribution, not a conviction signal. Two simple filters fixed the problem: drop individual transactions below $10,000 from the cluster count, and reject the cluster entirely if ≥80% of qualifying buys share the same date and price. Together they remove the programmatic noise while preserving real clusters. These are the “cluster gates” referenced below.

This post asks the natural next question. Given that the signal compresses fast, can the divergent cases, where the insider knows something the news does not, be more informative than either signal alone?

What “insider activity” means here

For purposes of this analysis, the pipeline classifies a ticker as showing bullish insider activity when at least one of the following is true over the past 20 calendar days:

A cluster of ≥2 distinct officers (CEO, CFO, COO, and other corporate officers) bought shares on the open market with their own cash, not through a pre-scheduled trading plan, with at least one buy above $10,000 and the cluster not consisting of a single same-day same-price compensation event (filtering out ESPP-style mass-officer distributions). Or a single officer bought ≥$1M of their own stock outside a pre-scheduled plan. Or a single director did the same. Or a 10%+ beneficial owner acquired ≥$1M of shares.

Conversely, the pipeline classifies a ticker as showing bearish insider activity when there is a non-plan sell of ≥$10M with no offsetting officer purchases in the same window. Sales are noisier than buys, because insiders sell for taxes, diversification, or scheduled vests as often as for bearish conviction, so the threshold is much higher and the filter is stricter.

All transactions explicitly executed under a Rule 10b5-1 trading plan are excluded throughout. These are pre-announced trading schedules that strip the timing decision from the insider, and the academic literature is unambiguous that they carry minimal signal.

What “news sentiment” means here

The pipeline reads news articles about each of 536 covered companies each day and runs them through two text-sentiment models: Claude (a large language model) and FinBERT (a domain-specific financial sentiment classifier). Each model produces a positive / negative / neutral label per article, which is aggregated to a per-ticker daily sentiment label.

A “news directional change” is detected when a ticker’s sentiment label flips between yesterday and today. These changes are stratified into tiers by conviction. The highest-conviction tier is when both Claude and FinBERT, two independent models with different architectures and training data, agree on the flip. The next tier is when only one of them flipped. There is also a “conviction loss” tier for when a previously directional label drops to neutral, which is interpreted as the original case weakening.

For this analysis we collapse the tiers and treat any directional change as a binary signal: bullish (any positive flip, or loss of a negative conviction) or bearish (any negative flip, or loss of a positive conviction).

What “excess return” means here

Raw stock returns are not the right way to compare across days when the broader market is moving. If the average stock in the universe gained 2% on a particular day, then a stock that gained 1.5% actually underperformed by half a percentage point, even though its raw return looks positive.

Excess return removes the market move. For each ticker on each day, we compute the cross-sectional mean return across all 536 tickers in the universe (the “universe mean” for that day) and subtract it from each ticker’s raw return. The result tells you how much better or worse this ticker did than the typical ticker that day. A +1% excess return means the ticker beat the universe by one percentage point; a −1% excess return means it lagged.

All the numbers reported below are excess returns. Positive means outperformance relative to the universe; negative means underperformance.

The three hypotheses

When two signals can each be bullish, bearish, or neutral on the same ticker on the same day, you can think of them as filling cells in a contingency table. There are four “joint” cells of interest: the two convergent cases (both bullish, or both bearish) and the two divergent cases (insider buying into bearish news, or insider selling into bullish news). What you expect to see in each cell depends on which of three hypotheses is true.

The independent-signals hypothesis says both channels carry partially independent information. Convergent cells should be stronger than either signal alone; divergent cells should be weaker or noisier.

The redundant-signals hypothesis says both channels measure roughly the same underlying state. Convergent cells should look like the dominant single channel; divergent cells should look like that channel alone with little added.

The informed-insider hypothesis says the divergent cases are themselves informative. When insiders buy against bearish news, they have private information the news has not yet caught up to. The divergent insider-bullish cell should be the strongest of the four — stronger than even the convergent bullish cell.

These hypotheses make different predictions. They are testable; that is the point.

What the data shows

Over twelve days of pipeline runs (May 19 through June 2, 2026), there were 91 ticker-days where both a Form 4 signal and a news directional change fired on the same ticker. Distributed across the four cells:

The same data shown visually:

Figure 1. Three-window excess returns by cell of the contingency table. Prior-day, same-day, and next-day excess returns (gray, navy, green respectively) for each of the four joint signal cells, May 19 – June 2, 2026, 536-ticker universe.

Three observations from this table, in increasing order of confidence.

The divergent insider-bullish cell has the highest next-day excess return of any of the four cells (+2.46%). Seven observations is still a small sample, but the magnitude is large enough that it is not noise. This is the cell where insiders are buying their own stock with their own money at the same time that the news cycle is turning negative; the informed-insider hypothesis predicts exactly this pattern. The original two tickers driving the result, WGS (a small-cap genetic testing company, where a director bought $60M) and UPST (a consumer lending platform, where a director bought a smaller amount), have continued to produce strong forward returns when they reappear: WGS posted +4.50% and +2.89% next-day excess on its two appearances, UPST +2.59%. A new pattern in this expanded window is SHAK, which has now appeared three times in the divergent cell as a director kept buying against repeated bearish news flips; the forward returns there have been mixed (−1.29%, +3.60%, and one event from June 2 still resolving), which is what the persistence of insider conviction looks like at the level of individual stocks rather than cell averages. Verdict: this favors the informed-insider hypothesis.

Several mechanisms could be producing this, and the data does not yet distinguish among them. We consider four below; the list is not exhaustive. The first is the standard academic story: insiders trade on private operational information that the press has not yet caught up to — internal forecasts, pipeline detail, execution quality, key hires landing or not landing. Most of this is not material in the strict legal sense, but it lets the insider hold conviction through a bearish news turn that an outside observer cannot evaluate as confidently. The second is more behavioral on the press side: the financial press is known to overweight recent negative headlines and amplify single bad data points, so when it turns bearish on a company the move sometimes reflects a real change and sometimes reflects the press doing what the press does. Insiders, who watch their own company every day for years, have a better baseline for distinguishing the two and correctly hold through the overreactions. These two theories share an assumption: the news is exogenous to the trade, and the insider’s job is to read it correctly.

The other two mechanisms put the insider’s action in a more active relationship with the news. The third is strategic signaling: an insider may buy specifically because of the bearish news cycle, not despite it — deploying their visible Form 4 filing as a deliberate signal to counteract the negative narrative. Form 4 filings reach financial-news terminals within 48 hours and are read by traders as votes of confidence; an insider who knows their buy will be reported during a bearish news window may be using capital partly as strategic communication and partly as conviction trade. The fourth, and least comfortable, is that the news itself is partly endogenous to the trade. Insiders or their associates may shape or facilitate the news cycle that creates the buying or selling opportunity — bearish coverage before a planned purchase, bullish coverage before a planned sale. The cleanest version is illegal market manipulation. The messier and more common version is everyday investor-relations management of expectations, narrative shaping through analyst contacts, and timing of corporate communications around insider transaction windows. Under this theory, the next-day excess in the divergent insider-bullish cell is not the market discovering what insiders knew; it is the market correcting a temporary distortion that insiders participated in creating. The cell’s stylized facts — directors (typically closer to the company’s external messaging strategy than C-suite officers) dominating the cell, the same insiders appearing repeatedly across multiple events for the same ticker, the asymmetry between the two divergent cells fitting institutional capacity to shape narratives in each direction — are at least consistent with this theory, though they fit the other three as well. The cleanest distinguishing tests require extensions to the panel (multi-day forward returns to detect signaling decay; article source classification to detect engineered narratives) that are not in scope for this post.

The mirror cell — divergent insider-bearish, where insiders are selling into bullish news — shows a smaller version of the same asymmetric pattern. With 53 observations (the largest cell, since large insider sells are the most common Form 4 event), the next-day excess is −0.89%, meaningfully below the news bullish single-channel baseline of +0.40%. Insiders selling into bullish coverage does carry information — the stock moves down despite the press positivity — but the magnitude is much smaller than the bullish-divergence case. The asymmetry between the two divergent cells (+2.46% versus −0.89%) is consistent with the long-established academic finding that insider purchases carry stronger signal than insider sales: insiders sell for many non-information reasons (taxes, diversification, scheduled vests) that do not apply to buys made with their own capital. Verdict: the asymmetry strengthens the informed-insider reading, because the joint signal does its work where insider information is strongest (buys) and contributes little where it is weakest (sales).

The convergent cells show their excess return before the joint signal surfaces, not after. Convergent bullish had a +2.43% prior-day excess and then drift toward zero (−0.63% next-day). Convergent bearish had −1.57% prior-day and −1.79% same-day, then essentially zero next-day. By the time the daily dashboard surfaced the convergent signal, most of the move was already in the price. This is consistent with the established academic finding (Ozlen and Batumoglu, 2026, SSRN) that 70-80% of insider-trading alpha lives between the transaction date and public disclosure. By the time a daily pipeline can act on the signal, the alpha has mostly evaporated. The new contribution here is that the pattern holds for the joint convergent signal too, not just the insider signal alone. Verdict: both the informed-insider and redundant-signals hypotheses are consistent with this pattern. It does not discriminate between them, but it says plainly that by the time a daily pipeline can act on a convergent signal, most of the move is already gone.

Why does the alpha live before disclosure rather than after? Several mechanisms are plausible; we consider two. The first is information leakage: the transaction itself leaks through secondary channels (broker order flow, market-maker inference, unusual volume detected by quantitative funds, clustering around earnings windows) faster than the two-business-day SEC filing window allows the formal disclosure to land. By the time the filing becomes public, the stock has already moved on the inferred trade. The second is correlated information: the insider is not trading on truly private information, but on information also accessible to other informed parties — industry contacts, vendor activity, hiring patterns visible from outside. When the insider buys, several other informed parties also act, often before the insider files. The pre-disclosure move reflects the broader informed flow, not the insider transaction in isolation. The leakage theory predicts the pre-disclosure window will keep compressing as detection methods improve; the correlated-information theory predicts a floor below which it will not fall, set by the existence of informed networks that no disclosure speedup can eliminate.

Single-channel signals are weaker than the joint signals. Looking at the same data but slicing by single channel: Form 4 bullish alone (across all 142 such ticker-days, with or without concurrent news) produced essentially zero next-day excess (+0.10%). News bullish alone (580 ticker-days) produced a small positive (+0.40%). News bearish alone (333 ticker-days) produced −0.83%. The divergent insider-bullish joint cell produced +2.46%. The combination of news bearish and insider bullish swings the post-disclosure return by more than three percentage points relative to news bearish alone: strong qualitative evidence that the joint signal carries information that neither alone does. Verdict: the signals are not independent. If they were, the convergent cells would be stronger than single channels — they are weaker. The independence hypothesis is ruled out.

Why would two signals together produce more than either alone? Several mechanisms are candidates; we examine two. The first is noise cancellation: each signal carries information mixed with noise, and the joint condition filters for cases where both noisy signals point the same way, producing a more reliable composite. The second is informational complementarity: the two signals carry different kinds of information — the insider sees operational reality (internal forecasts, pipeline detail, execution quality), the news cycle captures market psychology and external perception — and the joint condition reveals where these two sources of information interact. The data favors the second. If noise cancellation were the primary mechanism, the convergent cells should outperform their single channels too, since the same filtering logic applies. They do not: convergent bullish next-day excess is −0.63%, weaker than news bullish alone (+0.40%) and weaker than Form 4 bullish alone (+0.10%). The joint signal’s edge lives mostly in the divergent cells, where the two information sources are doing genuinely different work, and that is more consistent with complementarity than with noise reduction.

Verdicts on the three hypotheses

With the data reviewed, the three hypotheses introduced earlier sort cleanly.

The independent-signals hypothesis is rejected. It predicted convergent cells stronger than the single channels and divergent cells weaker. The data shows the opposite: convergent cells weaker than single channels, divergent insider-bullish much stronger. The signals are not adding up; they are interacting.

The redundant-signals hypothesis is also rejected. It predicted joint cells should look like the dominant single channel. They do not. The convergent cells diverge from the single channels in characteristic ways (the pre-priced pattern in Observation 2), and the divergent insider-bullish cell carries information neither channel alone produces. The signals are not measuring the same underlying state.

The informed-insider hypothesis is provisionally supported. It predicted the divergent insider-bullish cell would be the strongest of the four. At seven observations, it is. But within this hypothesis there is real ambiguity about which mechanism is doing the work — private information, press overreaction, strategic signaling, and engineered narrative are all consistent with the cell being strongest, and the current data cannot yet distinguish them. The data establishes that something more than either signal alone is happening in the divergent insider-bullish cell; it does not yet establish what.

Caveats, in order of importance

Twelve days is not a sample. Seven observations in the most interesting cell is not a sample. The window happens to include a strong rising-market regime, which affects how excess returns sort within each cell. The cluster gate that filters programmatic compensation events ($10K per-buy floor, 80% same-day-same-price exclusion) is calibrated against a single example from earlier work and is not yet tested at scale. Effect sizes, meaning the actual magnitude of the returns reported above, should not be cited; the right reading of this post is “the qualitative pattern is consistent with the informed-insider hypothesis,” not “insider divergence produces +2.46% next-day excess returns.”

The news-sentiment signal is itself a layered construct involving two text models, multiple aggregation steps, and tier classification that this post collapsed for simplicity. Different model choices, different tier weightings, or different thresholds for what counts as a “directional change” might produce somewhat different categorizations. The pattern’s robustness to these implementation choices is an open question.

The 10%-owner subcategory of Form 4 bullish signals is informationally distinct from officer-led buying (it often represents pre-announced strategic transactions like equity placements closing) and was included in the bullish cell without distinguishing. Within these twelve days the 10%-owner cases were rare enough not to dominate, but at larger sample sizes this distinction will matter.

What’s next

A reasonable amount of additional time, weeks rather than days, will continue to multiply the cell counts. The divergent insider-bullish cell, currently at n=7, should reach n=15-25 within several weeks, and n=50 or so within a quarter. At that point the question of whether the qualitative pattern holds at scale becomes statistical rather than illustrative. The data is accumulating continuously; this post is the first checkpoint.

Sample growth is only one of the directions where the panel can become more informative. Several extensions, naturally suggested by the theories above, would let future analysis discriminate among them. Extending forward returns from one day to five or ten would reveal whether the divergent insider-bullish cell shows sustained drift (consistent with private information being durably re-priced) or characteristic decay with partial reversion (consistent with strategic signaling, where the next-day move fades once the signal’s coordination effect has run its course). Classifying news articles by source quality and directional intent — distinguishing investigative reporting from analyst notes, paid commentary from independent journalism, short-seller research from neutral coverage — would test whether the bearish news cycles preceding divergent insider-bullish events look qualitatively different from ordinary bearish news, speaking to the engineered-narrative theory. A third extension would condition the contingency table on news-tier (§2 high-conviction agreement of both sentiment models versus §3 single-model flips) to test whether the cell effect strengthens at higher conviction, as the informational-complementarity reading would predict. None of these extensions requires fundamental changes to the pipeline architecture; they are the natural second-checkpoint work.

If the pattern holds, the practical implication is that divergence between insider activity and news sentiment is the signal worth attending to, not agreement. The convergent cases mostly represent moves that have already happened; the divergent cases are where private information is leaking out ahead of the public narrative. That is an actionable distinction. If it survives more data.

If the pattern doesn’t hold, the most likely failure mode is that the divergent insider-bullish cell collapses to small or zero excess return as the sample grows, suggesting the current +2.46% is sampling noise. That outcome is fine; the test is what matters.