AI Research AMDAMD_earningsAMD_insider

AMD pre-earnings insider selling (30-day) vs EPS surprise & earnings-day return

11
Earnings events analyzed

For AMD, the idea that insider stock sales right before earnings signal a looming miss is one of those theories that sounds almost too logical to fail. If insiders know bad news is coming, why wouldn’t they get out early? Over the past three years, the data from trades.run offers a blunt answer: it doesn’t work.

Among 11 quarterly events, the 9 quarters that saw insider selling in the prior 30 days actually produced a larger average EPS beat—3.8% versus 1.5% for the two no-sale quarters. Earnings-day returns, however, were worse on average with selling (-4.55% vs -1.55%). But both differences are statistically noise: p-values of 0.37 and 0.52, meaning the pattern could easily be random. Correlations between dollar amount sold and the outcome are weak and insignificant.

The takeaway is that, in this sample, pre-earnings insider selling is a coin flip, not a signal. The full breakdown below covers the methodology, charts, and the numbers behind each of the 11 events.

The research question

For AMD over the past ~3 years, do insider stock sales in the 30 days before earnings predict a below-consensus EPS surprise and a negative earnings-day return? Thesis: Pre-earnings insider selling signals upcoming disappointment, as insiders dump shares ahead of bad news they already know.

How this was measured

AMD quarterly earnings releases over the last ~3 years were paired with insider (Form-4) transactions. For each release, we counted insider disposals (acquisition_or_disposal='D', share_price>0 to exclude grants/vestings) in the 30 calendar days immediately before the reported_date. Earnings-day return was defined as close-to-close from the trading day before the announcement to the first trading day on or after the announcement. We then compared EPS surprise (% above/below consensus) and earnings-day return between quarters with any insider selling and those without (Welch t-test). Additionally, Pearson and Spearman correlations between total dollar notional of insider sales and both surprise and reaction return were computed.

The key numbers

Earnings events analyzed
11
quarters with known reported_date and non-null surprise_percentage
Events with insider selling (30d)
9
Events without insider selling
2
Mean EPS surprise – with sale
3.81
consensus surprise % (positive = beat)
Mean EPS surprise – no sale
1.47
consensus surprise %
Mean earnings-day return – with sale
-4.5457%
Mean earnings-day return – no sale
-1.5477%
Welch t-statistic (surprise)
0.990
Positive = sale group has HIGHER surprise (contrary to thesis)
Welch p-value (surprise)
0.3671
p=0.3671 ≥ 0.05 → not statistically different
Welch t-statistic (return)
-0.761
Positive = sale group has HIGHER announcement-day return
Welch p-value (return)
0.5197
p=0.5197 ≥ 0.05 → not statistically different
Pearson r (notional vs surprise)
0.233
weak (|r|=0.233)
Pearson r (notional vs react. return)
0.428
|r|=0.428 > 0.3 → moderate
Spearman rho (notional vs surprise)
0.235
rank correlation, robust to outliers
Spearman rho (notional vs react. return)
0.200

Reading the numbers

Over 11 earnings events, quarters with insider selling actually saw a higher average EPS surprise (3.81% vs 1.47%) and a more negative average earnings-day return (-4.55% vs -1.55%). Neither difference is statistically significant, so the data do not support the thesis that pre-earnings selling signals disappointment.

The charts

Pre-earnings insider $ notional sold vs EPS surprise
What this chart says

This scatter plot shows the dollar amount insiders sold in the 30 days before earnings (x-axis) against the EPS surprise percentage (y-axis). Most quarters have selling between $2M and $6M, and the points are spread widely with no clear downward trend. In fact, the weak positive correlation (r=0.233) means more selling was actually associated with slightly larger beats, the opposite of what the thesis predicted.

Pre-earnings insider $ notional vs earnings-day return
What this chart says

Here, the same pre-earnings notional sold is plotted against the stock's return on earnings day. The points are scattered across both positive and negative returns, and there is no obvious pattern linking larger insider sales to bigger drops. If insider selling reliably predicted bad earnings-day reactions, you'd expect the dots to slope downward, but they don't.

EPS surprise by presence of insider selling
What this chart says

This box plot compares EPS surprise between quarters with and without pre-earnings insider selling. The group with selling has a much wider range of surprises (0% to nearly 16%) and a higher average (3.81%) than the no-selling group (average 1.47%). However, there are only two data points in the 'no selling' group, making it hard to draw solid conclusions — the two groups' distributions overlap significantly.

Top events by insider selling notional (last 3 years)

Report dateHas saleTotal notional ($)Surprise %React. return
2024-07-3015837061.981.470.0731
2023-08-0115181119.871.750.0509
2025-11-0414416059.3510.29-0.0767
2025-08-0512790187.110-0.0868
2024-04-3012695799.761.64-0.0838

The takeaway

No, insider selling in the 30 days before AMD earnings does not predict a below-consensus surprise or a negative earnings-day reaction. Over 11 quarterly events, quarters with insider selling actually saw a larger average EPS beat (3.8% vs 1.5% for no-sale quarters) and a slightly worse mean earnings-day return (-4.55% vs -1.55%). But with only 9 sale events and 2 no-sale quarters, these differences are nothing more than noise — the p-values are 0.37 and 0.52, meaning you'd see results like this by chance 37% and 52% of the time, respectively. The correlations between the dollar amount sold and either metric are weak-to-modest at best (Pearson r ≈ 0.23 for surprise, 0.43 for return), and both are statistically insignificant. In short, this is a coin-flip result: the sample is far too small and the data too variable to support the thesis that insiders front-run bad news. The practical takeaway is that there is no buyable signal here — and the data actually hint in the opposite direction, though that hint is not reliable either.

The fine print