Insider Cluster Buying — our SEC data

The short version

When several company insiders buy their own stock in the same week, the shares tend to beat the market by about 1.5% over the next month once the buying becomes public — then the edge fades to roughly nothing within a year. (The often-quoted ~2.7% is measured from the insiders' own trade date, a few days before an outsider could act.) It shows up mostly in small, overlooked companies. In plain terms: it's a useful nudge for deciding when to buy a share you already like — not a reason to buy on its own, and not something to hold for the long term.

Abstract

We tested whether clusters of open-market insider purchases (≥2 insiders buying the same company within a week) predict returns — on our own data: SEC Form 4 insider filings + IB daily prices, across four regime-varied years (2017, 2019, 2021, 2022). Vs SPY the signal showed a real ~1-month positive abnormal return that reversed to a large negative by 3–12 months in every year — but the v1.2 refinement (§4b) shows that long-horizon collapse was ~85% small-cap-vs-large-cap beta, not the signal. Against the correct peer group (IWM, small-cap), the −18.6%/12m flattens to roughly +0.6% (not significant). Honest standalone net effect: a real, significant ~+2.7% one-month alpha and a modest ~+1.6% three-month edge, then flat — not value-destructive. The edge concentrates in small, cheap, less-covered names; cluster strength (≥3/≥4 owners) helps modestly; liquidity hurts. For our liquid HL universe it is a modest ~1-month effect — a conviction/entry-timing overlay, not a hold.

1. The claim (and why it matters to us)

Several insiders buying on the open market within days — a cluster — is a costlier, higher-conviction signal than one buyer. If it predicts returns, it's a cheap, public, long-side strand.

2. What the literature claims — and our scepticism

• Lakonishok & Lee (2001): insider purchases (not sales) predict; ~+7.4%/yr in small caps.
• Alldredge & Blank (2019): the cluster anchor — a purchase within two days of a peer earned +2.1%/month vs +1.2% for a solitary buy (a +0.9pp/month cluster premium).
• Our scepticism (now confirmed): the cluster premium is documented only at ≤1–3-month horizons; the popular "4–8% over 6–12 months" line traces to blogs, not a study. The investor who uses it best treats it as low-weight, never a thesis alone ("lots of crappy stocks have insiders buying"). Our data bears this out.

3. Our own analysis ← the heart of the report

Data: SEC Form 4 quarterly datasets (open-market purchases, code P/acquired; cluster = ≥2 distinct reporting owners on an issuer in the same ISO week) → IB daily prices (currently-listed universe; no rate limits). Abnormal = stock return − SPY return from the trade date; bootstrap CIs.

Forward abnormal return vs SPY, mean % (t-stat), by year:

(1-month medians are slightly positive every year (+0.6 to +0.8); by 12 months medians are deeply negative (−9% to −52%) with hit-rates 13–38%. ~51–63% of names per year were delisted/not in IB and dropped — see caveat.)

4. Findings

• The 1-month positive abnormal return is real, broad, and replicates in all four regimes (median positive, mostly significant). This is a genuine standalone effect.
• It reverses to negative by 3–12 months in every year — and 2017/2019 (calm/bull) are the worst. So the long-horizon negative is persistent across regimes, not a 2022 artifact (this overturns the single-year hypothesis).
• Net vs SPY: a ~1-month alpha then sharply negative. But see §4b — the long-horizon negative is largely a small-cap-benchmark artifact; vs IWM it flattens. The figures in this section are the large-cap-benchmarked numbers; the corrected standalone net effect is in §4b.

4b. Refinement — clean benchmark + conditioning (2026-06-02)

Re-run on the cached IB universe (~1,452 listed names; clusters pooled across 2017/2019/2021/2022), benchmarked against IWM (small-cap) instead of large-cap SPY, plus cluster-strength and liquidity splits. Mean % abnormal (t-stat):

Findings: - The brutal 12-month negative was ~85% small-cap-vs-large-cap beta, not the signal. Against the correct peer group (IWM) the −18.6%/12m collapses to roughly flat (+0.6%, not significant). The v1.1 "value-destructive" read was a benchmark artifact (the cluster universe is micro-cap-heavy). - Standalone net effect (vs IWM): a real, significant ~+2.7% one-month alpha and a modest ~+1.6% three-month edge, fading to flat by 6–12 months. A 1–3-month effect, not much more. - Cluster strength helps modestly: ≥3/≥4 owners lift the 1-month edge (~+2.9–3.1%) and turn the longer horizons mildly positive (vs flat) — but the long-horizon gains are small and not significant. - Liquidity HURTS: filtering to higher-priced (more liquid) names shrinks the 1-month edge and turns 6–12 months negative again — the edge lives in the smaller, cheaper, less-covered names. For our liquid HL universe, that is the binding caveat: a modest ~1-month effect at best.

4c. The tradable result — entering when the cluster is public (2026-06-07)

The +2.7% one-month figure above is measured from the insiders' own trade date. An outside investor can't act until the cluster is public — the second owner's Form 4 filing, a median 3 trading days later. Re-running the entry at that filing date (identical prices) roughly halves the edge but keeps it real, and standard errors clustered by week (events in the same week aren't independent) leave it clearly significant:

So the honest, tradable one-month edge is about +1.5% versus the small-cap benchmark, it survives a conservative (clustered) error treatment, and it is gone beyond a month — exactly an entry-timing overlay, now measured the way you could actually trade it.

4d. Does conditioning on quality extend it? — no (2026-06-07)

The hopeful conjecture was that an insider cluster in a high-quality company might carry beyond a month. We tested it: splitting the clusters by company quality (gross-profits/assets + leverage, from free SEC filings — 1,459 clusters with usable fundamentals), entering at the filing date vs the small-cap benchmark:

Quality conditioning did not extend the edge — quality-name clusters fade just like the rest; both are ~1–3-month effects that give it all back by 6–12 months. So the one-month, entry-timing reading stands, and we won't claim a quality-conditioned hold the data doesn't support. (SEC fundamentals resolve mainly for larger, established filers, so this answers the conditioning question for that slice; a cheapness/value split is the one test still open.)

5. Conclusion — a real ~1-month alpha vs the correct benchmark; an entry-timing overlay

Benchmarked correctly (vs small-caps), insider-cluster buying earns a real ~+2.7% one-month alpha and a modest ~+1.6% three-month edge, then fades to flat — not the value-destruction the large-cap benchmark implied (§4b). Reported cleanly, that is the honest net effect. It is not a 1–18-month hold-thesis. Its legitimate role (per the multi-strand model in ../families_and_independence.md and the way the best practitioners use it) is a conviction + entry-timing overlay on a name already justified by durable strands (quality, value, capital allocation, management) — harvested with entry discipline (small equal-weight, enter on a %-pullback not the pop, stops). This mirrors our 13F verdict: naive following fails; only conditioned following works.

6. Scope and limitations

Honest boundaries — they sharpen the verdict rather than undermine it: - It is a short-horizon, entry-timing effect — and measured the tradable way (entry at the public filing date, §4c) the one-month edge is ~+1.5%, gone beyond a month. That is precisely why we treat it as a light entry-timing overlay, not a large or standalone edge. - Returns are gross, and the edge concentrates in small, less-liquid names, where dealing spread and market impact matter most. Another reason it belongs as a small overlay on names already justified on fundamentals, not a strategy traded on its own. - Standard errors are clustered by week (events in the same week aren't independent): the one-month t is 7.6 at the trade date and 4.5 entering at the filing date (§4c) — comfortably significant either way. - Survivorship is a known boundary. Roughly 51–63% of names per year had delisted or left our price set and are excluded; delistings mix company failures (zeros) with takeovers (premiums), so they do not cleanly bias the result either way. The long-horizon figures are best read as indicative; the one-month result — where our verdict sits — is the most robust. - Benchmark. Measured against the correct small-cap peer (IWM); the earlier vs-large-cap (SPY) read conflated small-cap beta and is shown only for contrast.

7. How we would extend this

Post-filing entry, clustered errors and quality-conditioning are now done (§4c–§4d). Still open — all on free data:

1. Cluster strength — split by ≥2 / ≥3 / ≥4 owners and by key-executive (CEO/CFO) buys; already in the filing data.
2. Cheapness/value conditioning — quality alone didn't extend it (§4d); does conditioning on cheapness (a value tilt) do any better? — using free SEC financials + point-in-time valuation.

Quality-conditioning has had its answer (no). The remaining question is whether value does better — but the working assumption now is that this is a one-month, entry-timing signal, full stop.

Sources

• Pulled ourselves: SEC Form 4 insider datasets (clusters); IB daily prices (read-only).
• Cited: Lakonishok & Lee (2001, RFS 14(1)); Alldredge & Blank (2019, J. Financial Research 42(2)).

Changelog

• v1.2 (2026-06-02) — refinement (§4b): re-benchmarked vs IWM (small-cap). The v1.1 long-horizon "value-destruction" was ~85% small-cap-vs-large-cap beta — vs IWM the 12m negative flattens to +0.6% (ns). Honest standalone net effect = ~+2.7% 1mo / +1.6% 3mo, then flat. Cluster strength (≥3/≥4) helps modestly; liquidity (price ≥$5/$10) hurts — edge lives in small/cheap names.
• v1.1 (2026-06-02) — multi-year IB result (2017/2019/2021/2022): 1-month edge real & cross-regime, reverses to negative by 3–12mo every year → standalone ≈ 1-month effect (the 13F parallel). Refinement roadmap added. Survivorship reframed as ambiguous (zeros + buyouts).
• v1.0 (2026-06-01) — 2022 date-controlled result; multi-year pending price data.