Cluster Buying in Form 4 Data: A Useful Feature, Not a Standalone Strategy

Multiple insiders buying the same company is intuitively more interesting than one isolated purchase. The intuition is reasonable; the popular version of the claim is usually too strong.

Using the July 29, 2026 InsiderAlpha research release, purchase days classified as part of a 30-day multi-owner cluster had better median post-transaction outcomes than non-cluster purchase days. The difference remained positive in a within-issuer comparison. But the effect was incremental, the return distributions overlapped, and six or more buyers did not create a special threshold.

Cluster status is a useful research feature. This dataset does not support presenting it as a mechanical trading strategy.

What “cluster” means in this dataset

The cluster model counts distinct reporting owners with transaction code P for the same issuer inside inclusive 7-, 30- and 90-calendar-day windows. is_cluster_buy is true when at least three distinct owners appear inside the 30-day window.

That definition matters. It is not a 72-hour rule, and a cluster row is not the same thing as an independent corporate event. One issuer can contribute several purchase dates inside the same cluster window.

Code P itself means an open-market or private purchase under the official SEC Form 4 instructions. The cluster field describes co-occurrence in reported purchase data; it does not prove coordination among insiders.

Research design

To avoid counting multiple security lines as independent observations, we:

1. selected code P transactions dated January 1, 2007 or later;
2. collapsed the data to one row per ticker and transaction date;
3. classified the company-day as a cluster when its maximum cluster_size_30d was at least three;
4. used the median return when multiple lines mapped to the same company-day;
5. required a calculated 180-day market-price outcome.

The result is 14,485 mature company-day observations: 12,285 non-cluster days and 2,200 cluster-classified days.

Returns are absolute security-price changes from the dataset's transaction-date baseline. They are not benchmark-adjusted and are not executable returns from the filing timestamp.

Cluster versus non-cluster purchase days

Classification	Company-days	Issuers	Median 30D	Win rate 30D	Median 90D	Win rate 90D	Median 180D	Win rate 180D
Non-cluster	12,285	980	+1.92%	59.5%	+4.75%	62.7%	+7.80%	64.9%
30-day cluster	2,200	365	+2.31%	60.2%	+6.84%	65.9%	+9.80%	66.4%

The unconditional 180-day median difference is two percentage points. Cluster days also show a 1.5-point increase in the positive-outcome rate. These differences are directionally consistent, but much smaller than the spectacular returns often attached to cluster-buy headlines.

The middle 50% of outcomes still overlapped heavily:

• non-cluster: -6.43% to +23.95%;
• cluster: -5.72% to +26.83%.

That overlap is why cluster status cannot function as a complete investment thesis.

Does a larger cluster keep improving the outcome?

Not monotonically.

Distinct owners in 30D	Company-days	Median 180D	Positive outcomes
1–2	12,285	+7.80%	64.9%
3	1,062	+9.49%	67.2%
4	450	+10.00%	65.8%
5	246	+11.11%	65.4%
6+	442	+9.96%	65.6%

The six-plus cohort did not dominate the five-owner cohort, and its win rate was close to the other cluster groups. A model that assigns an enormous bonus at six buyers would be fitting a story, not the observed ordering.

Controlling partially for issuer composition

Cluster activity is not distributed evenly across companies. To reduce the chance that the result merely reflects which issuers happen to generate clusters, we examined 188 tickers with at least three mature cluster days and three mature non-cluster days.

For each ticker, we compared its median 180-day cluster outcome with its own non-cluster median.

Matched-issuer diagnostic	Result
Matched tickers	188
Median within-ticker cluster uplift	+4.61 percentage points
Tickers where cluster median was higher	59.6%
Issuer bootstrap 95% interval	+1.87 to +7.82 pp

This is stronger evidence than the raw comparison because the interval remains above zero. It is still observational. Different cluster and non-cluster dates occur in different market regimes, and several dates can belong to one underlying purchase episode.

How to use the feature responsibly

Cluster status is most defensible as one input in a broader model:

purchase code + distinct owners + reported value + role mix + issuer context + filing-time controls

The dataset does not justify replacing that model with cluster size >= 6. It also does not justify assuming insiders coordinated their decisions or possessed the same information.

For an executable study, a researcher would additionally need to:

• identify the first date on which the cluster threshold became publicly observable;
• enter only after the relevant Form 4 acceptance timestamp;
• use trading-day rather than calendar-day horizons;
• control for market, size, sector and liquidity exposure;
• group overlapping purchase days into independent episodes;
• reserve a later period for out-of-sample validation.

Conclusion

The corrected result is useful precisely because it survives without a sensational threshold. Cluster-classified purchase days were associated with moderately stronger outcomes, including within the same issuers, but the signal remained noisy and non-monotonic.

The InsiderAlpha dataset exposes cluster_size_7d, cluster_size_30d, cluster_size_90d and versioned methodology so researchers can test alternative episode definitions instead of inheriting an unexplained flag. Targeted cluster and transaction queries are also available through the API.