AMD heavy insider-selling weeks vs forward 20-day returns (last ~3 years)
The headline intuition — that clusters of insider sell filings mark nearby highs — doesn't hold cleanly for AMD over the trailing ~3 years. The six heaviest cash‑sell weeks produced a mean 20‑trading‑day return of about +0.94% versus a baseline mean of +5.82%, and the event distribution is noisy and small, so the gap is not a robust, statistically reliable signal.
I aggregated Form‑4 cash disposals into Friday‑aligned weeks, flagged the top decile by dollar notional (N=6) and measured forward 20‑day returns against the unconditional baseline. The full report below shows the raw weeks, distributions, t‑test, correlation checks and sensitivity notes — the short answer: clustered AMD insider sales look more like routine/diversification activity than a dependable short‑term bearish tell.
For AMD over the past ~3 years, do clusters of Form-4 insider selling actually flag local tops — do forward 20-day returns after the heaviest insider-sell weeks come in below the everyday baseline — or is executive selling just scheduled 10b5-1 noise with no predictive edge? Thesis: forward returns after heavy insider-selling weeks land statistically indistinguishable from baseline, so 'the insiders are dumping' is routine diversification rather than a bearish tell.
How this was measured
Form-4 insider disposals for AMD over the trailing ~3 years were aggregated into Friday-aligned calendar weeks, counting only cash sales (share_price>0). Weekly sell pressure is the sum of dollar-notional (shares×price). Heavy-sell weeks are the top decile by dollar notional (minimum of 3 weeks), restricted to weeks where a forward return can be computed. For each heavy week, the anchor is the first trading day after week end; the forward 20-trading-day return is close[t+20]/close[t]-1. The unconditional baseline is the distribution of all 20-day forward returns across the same price window. Welch’s t-test compares event vs baseline means; a Pearson correlation between log(1+weekly sell $) and weekly next-20d return probes a continuous relation.
The key numbers
Reading the numbers
There were 6 heavy-sell weeks; their mean forward-20d return was 0.0094 vs a 0.0582 baseline (edge -0.04887). The gap is not statistically significant (Welch p=0.3379), so the sample doesn't show a clear predictive sell signal.
The charts
This small histogram plots the six event-week forward 20‑day returns (n=6). Values span from a low of -0.1171 to a high of 0.209 with a mean of 0.0094, so the events include both losses and gains rather than a cluster of large negatives. With only six points the takeaway is that heavy insider-sell weeks produced mixed short-term outcomes, not a consistent crash after selling.
The unconditional baseline histogram (n=720) shows a wider spread of 20‑day returns, from -0.3063 up to 0.8786 and a higher mean of 0.0582. Compared with the event distribution, the baseline has a larger right tail and greater variance (baseline std 0.18897 vs event std 0.11222). Visually this tells you baseline 20‑day moves are more dispersed and on average more positive than the event-week outcomes.
The two-bar panel directly compares mean forward‑20d return for heavy-sell weeks (0.0094) against baseline (0.0582), showing an observed gap of -0.04887. Statistical testing yields Welch t = -1.054 with p = 0.3379, so that observed shortfall is not a reliable, statistically significant difference. In plain terms, the average after heavy selling is lower here, but the evidence is too weak to call it a consistent bearish signal.
The scatter plots weekly sell intensity (log(1+sell$), n=150) vs next-20d return; log sell$ runs from 0.0 to 17.881 and returns vary from -0.242 to 0.8565. The computed Pearson r is 0.0 with p = 1.0, indicating no linear relationship across the weekly grid. Practically, larger logged sell amounts do not map to systematically worse 20‑day returns here, consistent with the idea of scheduled or noisy selling rather than a predictive top signal.
Heaviest insider-sell weeks (top decile) — forward 20d outcomes
| week_end | sell_usd | anchor_date | fwd_20d_return |
|---|---|---|---|
| 2026-05-15 | 58291099.27 | 2026-05-18 | 0.209 |
| 2024-08-09 | 56913024.9 | 2024-08-12 | 0.047 |
| 2025-08-22 | 36893351.43 | 2025-08-25 | -0.011 |
| 2026-02-13 | 31011843.72 | 2026-02-17 | -0.0112 |
| 2024-03-15 | 30230227.16 | 2024-03-18 | -0.1171 |
| 2023-08-18 | 29709967.5 | 2023-08-21 | -0.0604 |
The takeaway
No — clusters of heavy Form‑4 cash sales for AMD over the last ~3 years do not reliably flag local tops. The six heaviest sell weeks produced a mean 20‑day return of +0.94% versus a baseline mean of +5.82%, an edge of -4.89 percentage points that is not statistically significant (Welch t = -1.05, p ≈ 0.338 — roughly a 34‑in‑100 chance this gap is just noise). The event distribution is noisy and mixed: median -1.11%, only 33% of events were positive, event std ≈ 11.22% vs baseline std ≈ 18.90%, and Pearson r = 0.0 (p = 1.0) shows no linear relationship between sell size and next‑20d return. Given the tiny event sample (N=6) and one very large sell week in the top decile, the evidence is thin; overall the data lean toward there being no reliable short‑term predictive edge from clustered insider selling. Practical takeaway: treat clusters of AMD cash sales as likely routine/diversification activity rather than a dependable short‑term bearish signal.
The fine print
- Only six heavy‑sell weeks (N=6) — very thin sample for a mean comparison.
- Many Form‑4 disposals are pre‑scheduled under 10b5‑1 plans; they may not be discretionary signals.
- One very large week (e.g., $58,291,099.27) can dominate the top‑decile totals.
- Forward 20‑day windows from nearby heavy weeks can overlap, inflating the effective sample.