NVDA 5-day forward return after >2% down day with volume >90th percentile (past ~3 years)
Over the past three years, NVDA’s worst single-day evaporations—a >2% plunge on volume in the top 10% of all sessions—have been followed by an average five-day return of +3.48%, nearly triple the everyday baseline of +1.23%. The edge of 2.25 percentage points aligns with the idea that sellers panic and buyers capitalize, but this is no clean victory lap. Only 25 events triggered the pattern, and the statistical test returns a p-value near 0.10, meaning the difference could easily be a mirage.
The split in forward returns tells a similar story: 68% of those capitulation days led to a positive five-day stretch, versus 57% for all days in the sample. Directionally compelling, but not conclusive. The full analysis below walks through the methodology, the caveats around overlapping returns and in-sample volume percentiles, and the raw numbers so you can weigh the evidence yourself.
For NVDA over the past ~3 years, do days with both a >2% drop and volume above the 90th percentile produce above-baseline forward 5-day returns, suggesting panic selling gets reversed? Thesis: forward 5-day returns after such capitulation sessions are positive and beat the everyday baseline because panic gets bought.
How this was measured
Resampled NVDA minute bars to daily close and volume. Flagged days where daily return dropped below -2% AND daily volume exceeded the full-sample 90th percentile. Measured forward 5-trading-day return (close[t+5]/close[t] - 1) for those days and compared the distribution to the unconditional 5-day forward return across all trading days in the window. Welch's t-test (two-sided, unequal variance) tests whether the capitulation-day mean differs from the baseline. Overlapping forward returns and the in-sample volume percentile are noted as caveats.
The key numbers
Reading the numbers
Of 746 trading days, only 25 qualified as capitulation. Those 25 averaged a 3.48% 5-day forward return, versus a 1.23% baseline for all days. The 2.25% edge is suggestive but not statistically conclusive (p=0.096).
The charts
This histogram shows the 25 forward 5-day returns following capitulation days. The average is 3.48%, but the individual outcomes range from a 4.9% loss to a 19.6% gain. Notice that the bars are mostly to the right of zero, consistent with the 68% positive fraction, but a few negative returns remain — so the pattern is not universal.
The bar chart directly compares the average 5-day forward return for capitulation days (3.48%) against the full baseline (1.23%). The capitulation bar is taller, showing the 2.25 percentage point edge. However, the p-value of 0.096 tells us this difference could reasonably occur by chance if there were no real effect; it's a promising signal but not a slam dunk.
Recent capitulation days (max 20 shown)
| date | daily_ret | volume | fwd_5d_return |
|---|---|---|---|
| 2023-07-20 | -0.032 | 559,458,610 | 0.027 |
| 2023-08-02 | -0.0411 | 548,010,370 | -0.042 |
| 2023-08-09 | -0.0496 | 600,287,560 | 0.0267 |
| 2023-08-11 | -0.0396 | 547,311,790 | 0.0671 |
| 2023-08-22 | -0.0276 | 767,577,980 | 0.0683 |
| 2023-08-24 | -0.0667 | 1,178,588,200 | 0.0482 |
| 2023-08-25 | -0.0211 | 942,006,710 | 0.0562 |
| 2023-09-15 | -0.0408 | 554,057,920 | -0.0492 |
| 2023-10-17 | -0.0573 | 827,589,810 | 0.0049 |
| 2023-10-18 | -0.0307 | 642,134,360 | -0.0273 |
| 2023-11-21 | -0.0294 | 578,834,460 | -0.019 |
| 2023-11-30 | -0.0321 | 601,406,870 | 0.0008 |
| 2024-02-06 | -0.0406 | 692,249,070 | 0.0762 |
| 2024-02-20 | -0.0634 | 721,476,510 | 0.1534 |
| 2024-03-08 | -0.1009 | 1,161,532,020 | 0.0363 |
| 2024-03-13 | -0.0276 | 652,049,440 | 0.0074 |
| 2024-03-14 | -0.036 | 623,631,320 | 0.0482 |
| 2024-03-27 | -0.0298 | 603,585,010 | -0.0409 |
| 2024-04-19 | -0.0979 | 898,426,970 | 0.1573 |
| 2024-06-06 | -0.026 | 677,590,600 | 0.0785 |
The takeaway
Over the past three years, NVDA’s most punishing days—those with a drop exceeding 2% on unusually high volume—have been followed by an average five-day return of +3.48%, far above the typical +1.23% for any five-day window. That extra 2.25 percentage points is directionally consistent with the idea that panic selling draws in buyers who reverse the move. But the evidence isn't airtight: there were only 25 such events, and the statistical test returns a p-value of 0.096, meaning roughly a 9–10% chance the difference is just random noise. The fraction of positive forward returns was 68% for these capitulation days versus 57% for the baseline, again leaning the right way but not overwhelming. In practical terms, the signal is suggestive enough to monitor but not strong enough to act on systematically without more data or a cleaner way to isolate pure panic events.
The fine print
- The 90th percentile volume threshold is calculated using the entire three-year sample, which introduces a mild look-ahead bias – future data influences today's trigger threshold.
- Because forward returns overlap across consecutive days, the observations aren't independent, making the t-test less reliable than a block-bootstrap or HAC adjustment.
- Only 25 trigger days – a small sample that makes the average return sensitive to a single outlier and limits statistical power.
- The simple price-drop+volume condition doesn't account for market-wide moves, earnings, or macro news, so the signal may not be pure capitulation.
- The last three years (post-2022 AI boom) represent a specific regime; results may not generalize to earlier periods or other stocks.