← All studies · askamerica.ai
q88 · askamerica

How much more do company insiders sell their own stock before bad results than at other times?

Answer: modestly more, and only in breadth (whether/how many insiders sell), not clearly in intensity (raw transaction volume).

Across 42,002 U.S. earnings announcements (2019-2024, 5,875 companies), the worst decile of two-day stock-price reactions (CAR ≤ -7.66%, used as the "bad results" proxy since this corpus has no consensus-EPS-surprise table) were preceded by at least one insider Form 4 sale in the prior 30 days 20.5% of the time, vs 16.7% before all other earnings announcements. That gap is statistically significant: chi-square = 40.24, p = 2.2e-10, n = 42,002. The average number of distinct insiders selling per pre-earnings window was also significantly higher before bad results (0.458 vs 0.358 insiders; Welch's t = -5.02, p = 5.2e-7). But the average number of sale transactions per window (1.68 vs 1.46) was not statistically distinguishable from chance (Welch's t = -1.21, p = 0.227) — the effect is real but small, and it is driven by more insiders participating, not a surge in trading volume from the same insiders.

Full published report (dashboard + methodology + sources + caveats), delivered via the AskAmerica connector's publish_report: http://127.0.0.1:54431/a/16a3effcf8b67d74f5521d173453c698.html (local loopback link, for the human reader; dashboard image saved alongside this file as dashboard.png).

Data and method

Statistical rigor / the confound named explicitly

Insiders sell for many routine reasons unrelated to inside information — portfolio diversification, tax planning, and above all Rule 10b5-1 pre-scheduled trading plans set up months in advance. Cohen, Malloy & Pomorski ("Decoding Inside Information," Journal of Finance 2012 / NBER WP 16454) find that more than half of all insider trades are "routine" and carry essentially no predictive information, while a minority of "opportunistic" trades carry all the predictive power (82bp/month value-weighted abnormal return). sec.insider_transactions carries no 10b5-1 plan/routine-vs-opportunistic flag, so this analysis cannot separate routine from informed selling — the reported gap is a raw, unconditioned association and should be read as an upper bound on any information-based signal, not proof of one. A related structural confound: many firms' trading windows open right after a quarter's earnings and stay open for weeks, so routine post-window selling naturally clusters in weeks that are also "30 days before next quarter's earnings" regardless of outcome — plausibly why even the "other" group's baseline (16.7%, 1.46 txn) is not near zero. Separately, research on 10b5-1 plans specifically finds plan sales cluster in the 40 trading days before earnings and are followed by roughly -6% average abnormal returns over 6 months (ScienceDirect, "When and how are Rule 10b5-1 plans used for insider stock sales?"), suggesting some of the effect found here could be real information leaking through nominally-"scheduled" trades rather than illegal trading.

Given these confounds, what stands up under a real significance test: (1) the probability that at least one insider sells at all is significantly higher ahead of bad reactions (hard to explain by routine-plan timing alone, since routine plans should hit both groups similarly), and (2) the number of distinct participating insiders is significantly higher. The transaction-count result not reaching significance despite a similar-looking raw gap (1.68 vs 1.46) is itself informative — it shows the breadth signal (more insiders selling, more often) is materially more robust than any intensity signal, and that raw transaction counts are noisy at this sample size due to a handful of firms filing bursts of Form 4s.

What this does and does not establish

Confidence: MEDIUM

The breadth findings (probability of any sale; number of distinct sellers) are statistically significant at real scale (n=42,002, p<1e-6) and computed directly from this session's own SQL against Form 4/8-K/price data. Confidence is capped at medium, not high, because (1) "bad results" is a price-reaction proxy, not an actual EPS-surprise measure, (2) the transaction-intensity result was not significant, and (3) this corpus cannot distinguish routine/10b5-1 sales from informed sales, so the finding is a raw, unconditioned association.

Sources

Data-quality issue found and reported

While building this analysis, found and reported (via report_issue) a SQL validator defect: a query fails with "Unknown identifier 'date'" when one CTE selects the reserved-word column close unquoted (relying on auto-quoting) while another CTE in the same query contains a CAST(x AS DATE) ... AS VARCHAR) chain; explicitly quoting close as "close" fixes it. Also noted: filtering sec.stock_prices by ticker alone with no year/partition predicate hangs indefinitely even for a single-ticker COUNT(*); adding a year predicate makes it instant, suggesting no non-partition index on ticker.