Most of the insider trading alpha is gone by the time you see the filing: poof!

An empirical companion to an upcoming literature review on SEC Form 4 trading signals.

The academic literature on legal insider trading is unusually mature. Sixty years of work, replicated across multiple samples and methodologies, has converged on a few consistent claims: insider purchases carry information, insider sales mostly do not, cluster buying by multiple insiders is stronger than individual transactions, and the alpha has been compressing for decades as the market got faster and broader awareness made the signal harder to harvest.

That literature is almost entirely about long historical samples. What does the signal actually look like now, in production, in 2026? A daily sentiment pipeline I run with a collaborator covers 536 US-listed tickers across the nine GICS sectors, and added an SEC Form 4 channel earlier this month. Five days of output (May 19 through May 23, 2026) is far too small to support an effect-size claim, but it’s enough to see whether the signals fire where the literature says they should, and to surface one operational issue not addressed in the academic work.

Three findings, with the appropriate caveat that everything below is n=4 to n=10 and intended as a sanity check, not a result.

The cluster definition needs an extra filter

The standard academic definition of a cluster buy — three or more distinct insiders making open-market purchases in a short window, with role and 10b5-1 filtering applied — flagged four tickers on May 19: TSM (Taiwan Semiconductor), ONON (On Holding), GEHC (GE HealthCare), and SPGI (S&P Global).

TSM was a noise event. The data showed 28 distinct officers buying within the 20-day window — far more than the threshold needs. Manual inspection revealed that 27 of the 28 underlying transactions were executed on the same day (May 8, 2026) at the same price ($71.82) to two decimal places, with individual buy sizes between $1,900 and $11,500. That precision is statistically incompatible with independent opportunistic buying. It’s the unambiguous signature of an Employee Stock Purchase Plan allocation, a directed-share offering, or some similar programmatic compensation event — many officers receiving small allocations on a single corporate trigger date, dressed up by the data structure as a “cluster.”

The Cohen-Malloy-Pomorski “routine versus opportunistic” filter, which is the gold standard refinement in the academic literature, would not have caught this. That filter looks for officers with a multi-year history of trading the same calendar month each year; the TSM officers had no such history. The signature here is different: dozens of officers buying the same security on the same day at the same price, which only happens when a corporate program triggered the buys.

Adding two simple filters — drop individual transactions below $10,000 from the cluster count, and reject the cluster entirely if 80%+ of buys share the same date and price — removed TSM from subsequent runs while preserving the three real clusters. The figure below shows the cluster lifetime across the five-day window:

TSM disappears after May 19 because the new filters were added between runs. GEHC and SPGI persist for one to two days and then age out naturally as their underlying transactions exit the rolling 20-day window. ONON persists across all five days because new officer buys keep entering its window. That persistence pattern — event cluster versus sustained cluster — is itself probably worth treating as a feature in any downstream model.

The alpha is mostly gone by the time you see the signal

The most striking finding in the five days, and the one most consistent with recent academic work, is about timing. For each of the three surviving cluster tickers and for six tickers with large single-director buys, I compared three windows of excess return — the day before the cluster surfaced in the daily pipeline, the day of, and the day after — to the cross-sectional mean across the 536-ticker universe.

The cluster signals (n=4 including TSM) showed +1.61% prior-day excess return on average, and essentially zero same-day or next-day excess. By the time the daily pipeline surfaced the cluster, the market had already moved. Individual cases: ONON had +2.63% prior-day excess and gave back 0.80% the next day. SPGI had +3.91% prior-day excess and gave back 0.56%. GEHC was the exception with positive excess across all three windows.

This matches a January 2026 SSRN paper by Ozlen and Batumoglu that decomposed the insider trading alpha into the portion that lives between the transaction date and public filing, versus the portion that’s accessible after the filing is public. They found that 70 to 80 percent of the apparent return lives in the unobservable window. The implication for production strategies is unambiguous: most of the alpha is captured by people closer to the transaction than a daily-cadence pipeline can be. Run a strategy on filing-date Form 4 data, act at the next open, and you’re working against a strong headwind that the historical academic backtests don’t account for.

There is one important asymmetry in the data, though. Large single-director purchases — six of them in the window, averaging $1M+ in net value, no 10b5-1 plan — showed a completely different timing pattern. Prior-day excess: +0.80%. Same-day excess: +3.55%. Next-day excess: +0.74%. Where coordinated cluster buying is largely priced in by the day before disclosure, large single-director discretionary purchases concentrated their move on the day of disclosure. WGS (a $60 million director purchase) was the most striking case, with +4.00% same-day excess. SHAK was the lone negative case at −3.18%. Five of six were positive.

Why might this be? One reading is that the market reacts faster to coordinated activity than to individual filings — multiple Form 4s arriving in the same window are themselves a more conspicuous public event than a single transaction, even a large one. A six-officer cluster prints in three news cycles; a single $50 million director buy may take a day for the relevant trading desks to notice. That’s a practitioner-relevant subtlety the academic literature does not address.

One transaction was a billion dollars and tells a completely different story

The largest single Form 4 in the window was Volkswagen AG buying $1.0 billion of Rivian common stock — 62,889,522 shares at $15.90 — as a 10% beneficial owner, on May 19, 2026. This was the closing of the strategic investment agreement that VW and Rivian announced in November 2024 and amended in April 2025. The amount, the price, and the timing had been public for 18 months.

The return pattern was different from anything else in the window. Prior-day excess: −2.81%. Same-day excess: −2.39%. Next-day excess: +4.27%. The stock drifted down going into the disclosure, then jumped on the actual filing — the classic pattern of a known event that the market re-prices when the formal closing occurs.

The takeaway for anyone building a Form 4 dashboard or model: 10%-owner transactions need to be visually disambiguated from officer transactions. The same dollar amount carries radically different information depending on whether it came from a CEO buying with their own cash or a strategic investor closing an 18-month-old agreement.

What this doesn’t tell you

Five days is not a sample. Six director buys and four clusters is not a sample. The window happens to include a 4% aggregate rally in the universe, which moves the goalposts on every excess-return calculation. The Cohen-Malloy-Pomorski routine-trade filter isn’t applied here because I don’t yet have three-year insider histories in the local store. None of the numbers above will hold up to inferential statistics applied to a single week.

What the data does suggest, qualitatively, is that the literature’s predictions show up in the right places. Filtered cluster signals carry pre-disclosure drift consistent with Ozlen and Batumoglu. Large director purchases reach the market with a lag. Pure insider sales contain little obvious information. Programmatic compensation events need their own filter that the academic literature hasn’t articulated. And insider activity sometimes diverges from news sentiment in ways that are themselves interesting — eight of the ten high-Form-4 tickers in the window also appeared in the news-sentiment sections of the same dashboard, sometimes agreeing with the insider signal, sometimes contradicting it.

A few directions worth investigating as more data accumulates. With twenty to fifty cluster observations, the per-buy and same-day-same-price thresholds become tunable against an actual false-positive rate. With a few hundred, the +1.61% pre-disclosure drift becomes a distributional estimate with sector and cap-size conditioning rather than a four-observation mean. The timing asymmetry between clusters and large director purchases is a five-versus-six observation sketch right now; a few months of data can tell you cleanly whether it survives. And the cross-signal question — what happens when insider activity and news sentiment disagree — has two anecdotes here pointing in opposite directions (WGS supports the informed-insider read, SHAK supports the news-is-right read) and would, with a year of those divergent cases, become statistical rather than rhetorical.