Insider Cluster Trades: Cohen, Malloy, Pomorski Retested

Note on figures cited below. Some statistics in this article come from the V12 universe cohort (227,594 BUY filings, period through 2024). Since 2025-01-01 the live out-of-sample BUY universe (28 markets) shows a 52.96 percent win rate and a +8.59 percent mean 90-day return on 20,688 filings. See methodology for the current numbers and the STRATEGY_PROOF block for the frozen audit reference.

One insider can be early, two can be friends, three in a month starts to look like information.

The signal gets louder when insiders stop acting alone

Insider trading research has a bad habit of sounding mystical. It is not. Most of it reduces to a simple proposition: corporate insiders know more about their firms than the rest of the market, but not every trade they make carries the same amount of information. Tax bills exist. Diversification exists. Option grants exist. So the useful exercise is to identify the subset of trades that look less like administration and more like belief.

That is where cluster trades enter. Cohen, Malloy and Pomorski, in a paper that has aged better than much of the factor zoo, formalised a practical idea: insider signals are stronger when several insiders trade in the same direction within a short window. Three or more insiders in 30 days is the sort of threshold that makes a quant nod and a compliance officer reach for coffee.

The appeal is obvious. A single chief financial officer sale could mean school fees. Three executives buying after a drawdown is harder to dismiss as household budgeting. Aggregation strips out idiosyncratic motives and preserves the common component, namely private information or at least unusually strong internal conviction.

For a 17-market study, that common-sense intuition is only the start. The harder part is making the signal comparable across jurisdictions that disagree on almost everything that matters in practice: who must report, how quickly, what counts as a reportable transaction, and how cleanly the data can be linked to listed issuers.

Why the cluster definition matters

The threshold, three insiders within 30 days, is not sacred. It is useful because it balances rarity and sample size. Set the bar at two insiders and the signal becomes more frequent, but also noisier. Set it at five and you get a purer event, but in smaller markets you may be left with a backtest made of decorative anecdotes.

The 30-day window is similarly pragmatic. It is long enough to catch staggered filings and internal blackout timing, but short enough to preserve the notion of a common informational episode. Stretch the window too far and a cluster becomes merely a quarter.

Why Sharpe is the right wrong metric

The user angle here is the Sharpe of the cluster signal in each market. That is the right instinct. Raw returns flatter illiquid markets, small-cap markets and periods when everything worked. Sharpe at least asks whether the signal paid investors for the path they had to endure.

It is still an imperfect measure. A market with sparse but high-conviction clusters can show a lower Sharpe than a market with frequent, moderate signals simply because diversification works. Transaction costs, borrow constraints for sell signals, and filing lags can also turn a handsome in-sample Sharpe into a rather ordinary live result. Quant finance is full of beautiful ratios that disappear on contact with a broker.

Cohen, Malloy and Pomorski still matter, because they asked the right question

The enduring contribution of Cohen, Malloy and Pomorski was not merely that insiders can predict returns. That result predates them by decades. Their sharper point was that the structure of insider trading activity matters. Insiders trading together contains more information than insiders trading alone.

This may sound almost offensively intuitive. Good papers often do. The achievement was to show it systematically and to distinguish clustered activity from isolated transactions in a way that survived standard controls.

The economic logic is stronger than the slogan

If multiple insiders buy around the same time, several things are more likely to be true.

First, the signal is less likely to be driven by personal liquidity needs. Household shocks are not usually synchronised across the executive suite.

Second, the information set may be broader. Different insiders see different corners of the business. Agreement across functions can indicate that the valuation gap is not a quirk spotted by one person with a hobbyhorse.

Third, the act of trading may itself reflect internal confidence. Insiders know the legal and reputational risks of buying before bad news. When several still choose to buy, that is informative.

The reverse case, clustered selling, is trickier. Sales are generally less informative than buys because insiders sell for many reasons, most of them dull. A cluster of sales can still matter, especially if it follows a long rally and excludes mechanical transactions, but the buy side is where the literature has usually found the cleaner signal.

Why this is especially interesting across 17 markets

A cross-market revisit is not just replication tourism. It tests whether the signal is a behavioural constant or a legal artefact.

If clustered insider buying works in the U.S. but fades under European disclosure rules, that tells us something about filing delays, market efficiency and perhaps the degree to which public disclosure itself arbitrages away the insiders' edge. If the signal survives in both regimes, that is more impressive.

It also addresses a practical investor question. If you run a global strategy, should you weight cluster trades equally across countries? Almost certainly not. The signal may be conceptually similar, but the investable version depends on local microstructure.

Cross-country Sharpe is a legal dataset wearing a quant costume

The temptation in a 17-market article is to jump straight to a league table. France 0.8, Canada 1.1, Nordics 1.4, everyone applauds, and then someone notices that one market reports amendments as new filings and another reports spouse transactions under the same identifier. The Sharpe was legal plumbing all along.

A serious revisit has to begin with the reporting regimes.

Disclosure rules shape the observable signal

In the U.S., Section 16 insiders report on Form 4, generally within two business days for most transactions. In Europe, MAR Article 19 requires persons discharging managerial responsibilities, and persons closely associated with them, to notify transactions promptly and no later than three business days after the transaction. That sounds similar enough for a slide deck. It is not similar enough for clean cross-country comparability.

Who is captured differs. What transaction types are common differs. How national competent authorities publish the data differs. Whether the market has centralised machine-readable access differs, often painfully.

A cluster strategy is therefore partly a test of information and partly a test of publication mechanics. If one market has faster, cleaner dissemination, its post-filing alpha may be lower because the market incorporates the information more efficiently. That is not a failure of the underlying insider conviction signal. It is a sign that the market can read.

Europe looks harmonised until you inspect the footnotes

MAR was meant to standardise much of this across the EU, and to an extent it did. But implementation and data accessibility still vary by regulator and venue. Some regulators provide structured disclosures. Others provide PDFs with the aesthetic of a tax dispute. For quants, the difference is material.

France, under the AMF, sits within the MAR framework. So do many other European markets in a 17-country sample. The deadline may be the same on paper, but the practical latency from transaction date to machine-ingestible signal can still differ.

Sample construction is where backtests go to die

For a cluster signal, the following choices are not housekeeping. They are the strategy.

1. Trade type filters

Open-market purchases and sales should be separated from option exercises, grants, gifts and other non-discretionary transactions. Many papers do this because a mixed basket dilutes the economic content. A cluster of compensation-related entries is a payroll event, not an alpha event.

2. Directional purity

A proper cluster should usually require a common direction. Three filings in 30 days, one buy and two sells, may indicate confusion rather than consensus. If the objective is to capture shared conviction, mixed-direction clusters should be excluded or classified separately.

3. Insider identity resolution

The same person can appear under multiple name variants. Closely associated persons can be reported differently across markets. Board members who sit on several entities can produce matching errors that create phantom clusters. Anyone who has ever entity-matched European filings knows that modern finance still depends on surnames and regret.

4. Event timing

Do you sort on transaction date or filing date? The answer for an investable strategy is filing date, because investors cannot trade on a transaction they have not seen. But transaction date is still useful for descriptive work. The gap between the two is itself informative about signal decay.

5. Return horizon and portfolio formation

A cluster signal can be tested over 1, 3, 6 or 12 months. The literature often finds the effect strongest over medium horizons rather than overnight. Equal-weighted versus value-weighted portfolios also matter, especially when small caps dominate the event set.

What we can say now, and what needs the full backtest

The brief asks for the Sharpe of the cluster signal in each market. The live data supplied here gives us one anchor point: the database contains 162,574 director dealing filings. That is enough to support a serious cross-country exercise. It is not enough, by itself, to print a market-by-market Sharpe table without pretending that missing numbers are a personality trait.

So the honest article does two things at once. It explains what a robust result should look like, and it marks clearly where the actual computation must still be done.

The database is large enough for the question

A 162,574-filing base is not toy data. Even after filtering to discretionary trades, deduplicating amendments, mapping issuers and requiring three insiders in 30 days, there should be enough events to estimate market-level performance in the larger jurisdictions and regional aggregates in the smaller ones.

The likely bottleneck is not total observations. It is event sparsity by market after applying quality filters. A 17-market split means some countries will have plenty of clusters and others will not. The smaller the market, the more unstable the Sharpe estimate. This is another reason not to worship a league table.

What a credible result would probably show

Without inventing market Sharpe numbers, the prior from the literature is fairly clear.

First, clustered buys should outperform isolated buys on a risk-adjusted basis in most markets with usable data.

Second, clustered sales should be weaker and less consistent, unless the sample is carefully filtered to remove mechanical and compensation-related selling.

Third, the signal should be stronger in smaller, less-covered firms and in markets where information diffuses more slowly.

Fourth, post-disclosure alpha should be lower where filings are faster, cleaner and more widely followed, because the market processes them more efficiently.

None of these are controversial. They are the sort of results that survive because they are tied to economic mechanisms rather than statistical whimsy.

Why some markets may look better than they really are

A high Sharpe in one country can reflect genuine inefficiency. It can also reflect one of four less romantic explanations.

Sparse events

If a market produces few clusters, a handful of successful episodes can dominate the estimate.

Small-cap concentration

Equal-weighted portfolios in small-cap-heavy markets often look brilliant on paper and expensive in reality.

Publication quirks

If filings are timestamped or aggregated in unusual ways, event windows can accidentally capture pre-existing momentum or stale disclosure.

Survivorship and mapping bias

If issuer identifiers are cleaner for surviving firms than delisted ones, backtests can drift upward. Delistings have an unhelpful habit of being relevant precisely when insiders are informative.

How to run the 17-market test without embarrassing yourself

A revisit worthy of the name should be conservative. The original insight is strong enough that it does not need decorative optimisation.

Event definition

Use issuer-level events where at least three distinct insiders file same-direction discretionary trades within 30 calendar days. Distinct means distinct beneficial owners after identity resolution, not three line items from one executive and his family office.

Signal timestamp

Form the signal on the public filing date, using the close of that day or the next tradable close depending on timestamp granularity. This avoids the common sin of trading on information before it was public.

Portfolio construction

Build market-neutral or benchmark-relative long-short portfolios if shorting is feasible, otherwise long-only on clustered buys. Equal-weight and value-weight both deserve reporting. If only one is shown, equal-weight is usually the more academically sensitive and less institutionally realistic choice.

Risk adjustment

Report raw returns, volatility, Sharpe, hit rate and factor-adjusted alpha where local factor models exist. At minimum, compare against market and size exposures. A cluster strategy that is secretly just a small-cap rebound strategy is still interesting, but it should not wear a false moustache.

Robustness checks

Use subperiod splits, pre and post MAR for European markets where relevant, and alternative cluster thresholds such as 2-in-30 and 3-in-60. If the signal survives mild perturbations, confidence rises. If it disappears the moment you change one parameter, you have found a PowerPoint effect.

Capacity and costs

Estimate turnover and implementation costs. The signal may be statistically elegant and economically tiny. This happens more often than quants admit in public.

The literature’s message is robust, but the modern question is decay

The useful debate is no longer whether clustered insider buying contains information. It does. The more interesting question is how much of that information remains harvestable after disclosure reforms, data commoditisation and the spread of event-driven systematic strategies.

Has the edge been arbitraged away?

Possibly in the largest, cleanest markets, at least partly. U.S. insider data is easy to access and heavily mined. That tends to compress the post-filing window. But “compressed” is not the same as “gone”. Signals tied to genuine private information often degrade slowly because they are capacity-limited and episodic.

In Europe and other non-U.S. markets, the answer may differ by country. Where data remains fragmented or publication formats are poor, the informational edge can persist longer, though implementation costs also rise. Alpha often hides in places where CSV files fear to tread.

Why the cross-market view still matters

A 17-market comparison can tell us whether the signal is fundamentally human or mostly institutional.

If clustered buys work almost everywhere, then insider coordination is a broad behavioural fact. If they work only where disclosure is messy, then the edge is more about data processing than about insiders per se. Both findings would be useful. One is a finance result, the other a market structure result. Investors can make money from either, provided they know which business they are in.

What this means for practitioners screening director dealings today

For a live process, cluster logic is one of the few sensible ways to move from anecdote to ranking.

A practical hierarchy of signals

If you are triaging filings globally, the order of interest is usually:

1. Clustered open-market buys by multiple senior insiders.
2. Repeated buys by the same senior insider after weakness.
3. Isolated open-market buys by CEOs or CFOs.
4. Clustered sells, but only after excluding mechanical transactions.
5. Everything else, which mostly belongs to compliance archives and trivia nights.

This hierarchy is not elegant, but it is empirically sensible.

The market-by-market Sharpe should drive allocation, not belief

Once the 17-market backtest is computed, the Sharpe table should be used to size exposure and prioritise research effort, not to decide whether the phenomenon is “real”. The phenomenon is real enough. The market-level dispersion will mostly tell you where it is easiest to monetise.

That distinction matters. Investors often confuse signal existence with signal accessibility. A market can contain informative insider clusters and still offer poor realised strategy performance because of costs, illiquidity or disclosure latency. Reality is rude like that.

The next analytical layer

After the basic cluster test, the natural extensions are obvious and useful:

• Weight insiders by role, CEO and CFO clusters should usually matter more.
• Interact with valuation, clusters after large drawdowns may be stronger.
• Control for earnings proximity and blackout windows.
• Separate first buys after long inactivity from routine accumulation.
• Test whether cross-insider functional diversity, finance plus operations plus board, predicts stronger outcomes than three directors from the same social circle.

That last point is especially underexplored. Not all clusters are equal. Three insiders seeing different parts of the elephant should be more informative than three seeing the same tail.

The payoff is straightforward. A 17-market cluster-trade study can do more than reheat a classic paper. It can tell us where insider conviction still survives public disclosure as an investable signal, and where the market has already done the reading. The concrete next step is simple: compute filing-date-based cluster portfolios for each market, publish the full Sharpe table with transaction-type filters and confidence intervals, then test whether the winners are the markets with the worst data plumbing or the deepest behavioural inefficiencies. That is a useful open question, and unlike many in quant finance, it has the decency to be answerable.

Editorial note, this article was assembled without live web research (Grok unavailable in this generation pass). Evergreen sources are cited above; numerical claims are pulled from our own database snapshot as of 2026-05-17.

Insiders Trades backtest cohort

Our analysis of insider trading signals reveals significant insights from a robust dataset. With 220,691 priced trades across 17 markets, the mean 90-day return stands at -1.70%, accompanied by a win rate of 41.82%. The annualized Sharpe ratio for this cohort is -0.08, indicating that the returns do not adequately compensate for the risk taken. This aligns with the broader context of 227,594 BUY filings across the same markets, which also reflect a mean 90-day return of -1.70% and an aggregate Sharpe ratio of -0.08. These figures underscore the challenges inherent in capitalizing on insider trading signals, suggesting that while there may be potential in identifying cluster signals, the overall performance remains lackluster. It is essential to consider that these findings are based on a substantial sample size and a specific 90-day time horizon.

Go deeper

This article is part of our reading an insider transaction series. Adjacent studies:

• How to read an insider transaction: the field-by-field pillar
• CEO vs CFO buys: the role dimension inside a cluster
• Small-cap insider edge: where clusters move prices most
• When do insiders themselves exit?
• The composite score audit: how clustering weighs into the final rank

Apply it live: the cluster buys screener and the recent clusters page surface forming clusters, and the performance page tracks the strategy built on top of them.

Last reviewed · 2026-05-18 · By Simon Azoulay · Sources, SEC EDGAR, AMF BDIF, and 28 other regulators.