Cluster Buy Signals Looked Like the Strongest Insider Signal. Our Own Data Says They Underperform.

Cluster vs. non-cluster returns, by horizon

We recomputed Form4API's own cluster definition — 3 or more distinct insiders (by CIK) buying the same company, same direction, within a trailing 5-calendar-day window — directly from raw transaction records, on the same base population as our insider-buying-returns-study and ceo-buys-weakest-insider-signal studies: open-market buys (code P), excluding derivatives, superseded rows, and 10b5-1-planned trades. n=75,436, 2023-04 through 2026-09. Buys by 1-2 insiders in the same window count as "non-cluster."

Absolute, split-adjusted returns anchored at the first close on or after the filing date. Open-market purchases only (code P), excluding derivatives, 10b5-1 plan trades, and superseded rows. Cluster flag recomputed independently from raw transactions, not from the /v1/signals precomputed output.

Diff = cluster minus non-cluster, in percentage points.

The mean gap is negative — clusters underperform — at every horizon, and it widens as the horizon lengthens. The median gap is more interesting: at 1 week and 1 month, cluster medians were fractionally ahead (+0.18 and +0.28 points), so a short-horizon trader looking only at typical outcomes wouldn't have seen a problem yet. The median gap flips negative at 3 months and stays negative at 6, and every one of these four confidence intervals excludes zero. The clean, defensible version of the headline is the 3- and 6-month result, in both mean and median — that's where the reversal is unambiguous.

By insider count: a threshold, not a slope

Splitting 3-month returns by the exact number of distinct insiders in the window shows where the underperformance actually happens. Per-bucket sample size is not available for this specific cut — only the aggregated 3-month cluster (n=16,751) and non-cluster (n=45,133) totals from the table above are known:

• 1 insider: +8.33% mean / +2.23% median / 56.1% hit rate
• 2 insiders: +8.73% mean / +3.70% median / 60.0% hit rate
• 3 insiders: +5.23% mean / +1.61% median / 55.0% hit rate
• 4 insiders: +5.43% mean / +3.41% median / 56.3% hit rate
• 6+ insiders: +4.43% mean / +1.87% median / 55.1% hit rate

Returns don't decline gradually as more insiders join. They hold steady, and in the 2-insider case even improve slightly, right up through the "non-cluster" side of Form4API's definition — then drop sharply the moment a buy crosses the 3-insider cluster threshold, and stay roughly flat from there through 4 and 6+ insiders. That step-shaped pattern, landing exactly on the definitional cutoff rather than trending downward from 1 insider on, is more consistent with a composition change at the threshold (a different mix of tickers or sectors starts qualifying at 3+ insiders) than with "more insiders buying together dilutes the signal" as a smooth mechanism. We haven't confirmed which explanation is correct — see the caveats section below.

Is this a real effect, or noise?

We bootstrapped both the mean and median difference between cluster and non-cluster returns at each horizon — 10,000 resamples for the means, 2,000 for the medians — to check whether the gap could plausibly be zero. At 3 and 6 months, every confidence interval (mean and median) excludes zero: [−4.23, −1.72] and [−0.83, −0.11] at 3 months; [−4.95, −2.89] and [−2.33, −1.15] at 6 months. At 1 week and 1 month, the mean-difference intervals also exclude zero, but the median-difference intervals sit on the positive side ([+0.06, +0.31] and [+0.05, +0.50]) — a real, if small, edge for clusters in the short-horizon typical case. This isn't sampling noise at 3+ months; it's a measurable, consistent reversal of the direction the academic literature predicts.

Robustness checks

Two checks against the result being an artifact of one ticker or one unusual year:

• Across 1,696 distinct companies in the cluster sample, the single largest ticker accounts for roughly 1.4% of it — this isn't a handful of names driving the average.
• The share of all open-market buys that qualify as a cluster has been stable year over year: 28.5% in 2023, 28.4% in 2024, 25.8% in 2025, 27.4% in 2026 YTD. Whatever is producing this gap has been present at a roughly constant rate across four years, not a one-time event.

What the literature said, and why we might diverge

The academic case for cluster buys rests mainly on Lakonishok and Lee (2001), "Are Insider Trades Informative?" (Review of Financial Studies), which found cluster purchases outperforming single-insider purchases by roughly 2x over 21-trading-day windows — and that result is what Form4API's own cluster-buy guide cited, secondhand, until this study. Three differences are worth naming plainly rather than papering over:

• Lakonishok & Lee used market-adjusted returns (excess over a benchmark) on 1975-1995 data. This study uses raw, unadjusted price returns on 2023-2026 data, for consistency with every other Form4API return study. A market that drifted up broadly over our sample period would inflate the non-cluster side more than the cluster side if cluster buys skew toward lower-beta names (see caveats) — which cuts in the direction of the reversal we found, not against it.
• Markets, data availability, and the insiders being studied are all different 35-50 years later — same-day EDGAR filing didn't exist in 1975-1995.
• Our cluster definition (3+ insiders, trailing 5 calendar days) is Form4API's own operational choice, matching what our production system computes. It's close to but not identical to the parameters used in any single academic paper.

We don't have a confident explanation for why the sign flipped rather than just the magnitude shrinking, as some post-2015 replications reported. The most likely partial explanation — a sector shift in which companies generate cluster buys today — is covered next, and it's unresolved.

What this data can and can't tell you

Methodology

• Population. Same base population as the insider-buying-returns-study and ceo-buys-weakest-insider-signal guides: open-market buys (code P), excluding derivatives, superseded (amended) rows, and 10b5-1-planned trades. n=75,436, 2023-04 through 2026-09, with a matured return required at each horizon reported.
• Cluster definition. 3 or more distinct insiders (by CIK) at the same company, same direction, within a trailing 5-calendar-day window — Form4API's own definition, matching what the production EnrichmentService computes. Recomputed independently from raw transaction records for this study rather than read from the precomputed /v1/signals output.
• Non-cluster. Buys by 1 or 2 distinct insiders in the same window — i.e., buys that don't meet the cluster threshold.
• Return definition. Absolute, split-adjusted price return anchored at the first market close on or after the filing date, measured at 5, 21, 63, and 126 trading days (approximately 1 week, 1 month, 3 months, 6 months).
• Statistical testing. Bootstrap 95% confidence intervals: 10,000 resamples for mean differences, 2,000 resamples for median differences, at each horizon.
• Robustness checks. 1,696 distinct companies in the cluster sample (largest single ticker ~1.4% of it); cluster share of all qualifying buys stable 2023-2026 YTD (28.5% / 28.4% / 25.8% / 27.4%).
• Known limitation. Not sector-adjusted. See caveats above.

Reproduce it

The underlying transaction and return data is on the API. There's no single query parameter that returns "cluster vs. non-cluster" pre-split for this exact comparison — reproduce the cluster flag the same way this study did, from raw transactions, following the manual approach in the cluster-buy-signals guide's detection section: filter to code P, exclude is10b5Plan, group by issuer CIK and date, and count distinct filerCik values in the trailing 5-day window.

Pull the base population with matured returns:

curl "https://api.form4api.com/v1/transactions?code=P&exclude_10b5=true&has_returns=true" \
  -H "X-Api-Key: $FORM4API_KEY"

Each row carries the return fields (return1w/1m/3m/6m, stored as fractions — 0.084 = +8.4%) plus filerCik and issuerCik, which is everything needed to recompute the cluster/non-cluster split and check the numbers above. Full filter grammar is in the screener guide, and the cluster-buy-signals guide covers the exact 5-day/3-insider definition step by step.

Frequently asked questions