NVDA earnings-day reaction vs EPS surprise magnitude (last ~3 years)
A clear chunk of NVDA’s earnings beats still closed the day down: 36.4% of beats and 37.5% of “clear beats” (≥5pp) lost ground, and many showed a gap-up that faded intraday. Quantitatively, surprise explains almost nothing of same-day variance (Pearson r = 0.184, R² = 0.034) and the fitted slope implies only about +0.32% per 10 percentage-point surprise — a very small mechanical effect. Beats averaged +1.79% versus the lone miss at −0.40%, but that difference is not statistically robust (Welch p = 1.00).
This study looks at 12 NVDA earnings releases over the last ~3 years, measuring day-0 returns (close-to-close), gap and intraday moves, and correlating those with EPS surprise magnitude. The working thesis — that guidance and crowded pre-print positioning often trump the headline EPS — guides the interpretation; the full statistics, charts, and event-by-event detail follow below.
For NVDA over the past ~3 years, does the size of each quarter's EPS surprise actually drive the earnings-day price reaction, or has 'buy the rumor, sell the news' broken that link — how often does a clear beat still close red? Thesis: the relationship between surprise magnitude and the earnings-day move is weak and several beats sold off, because guidance and crowded pre-print positioning drive the reaction far more than the headline number.
How this was measured
Filtered NVDA quarterly earnings to the last ~3 years with non-null surprise_percentage. For each release, defined day 0 as the first trading day on-or-after the reported_date. Computed earnings-day reaction as close(t0)/close(t-1)−1 (day-0 return), plus gap (open(t0)/close(t-1)−1) and intraday (close(t0)/open(t0)−1). Measured the relationship between surprise magnitude (percentage points) and day-0 return via Pearson/Spearman and an OLS slope (return per 1pp of surprise). Counted how often beats — and 'clear beats' (surprise ≥ +5pp) — still closed red. Also compared beat vs miss day-0 returns using Welch’s t-test (unequal variance).
The key numbers
Reading the numbers
Across 12 NVDA earnings (Aug 2023–May 2026) the link between EPS surprise and same-day return is weak (Pearson r=0.184, R²≈0.034). Notably, about 36.4% of beats still closed the day red, so surprises rarely tell the whole story.
The charts
This scatter places each quarter's EPS surprise on the x-axis (range −100.0 to 30.4348 percentage points, mean 1.2925) against the day-0 return on the y-axis (range −0.0842 to 0.0964, mean 0.0161). The points form a loose cloud rather than a tight upward trend, and the reported Pearson r=0.184 with R²≈0.034 shows surprise size explains only about 3.4% of the return variation. The OLS slope (0.0003201 per 1pp surprise) is essentially negligible, so even big positive surprises do not reliably produce big positive same-day moves.
The box plot compares the single miss (n=1, mean −0.004) to the 11 beats (mean 0.0179) and shows beats have a higher central tendency but a wide spread (beats min −0.0842, max 0.0964). Importantly, some beats closed negative, so the distribution overlaps materially with the miss. The Welch p-value = 1.0 means the apparent mean advantage for beats is not statistically clear in this sample.
These bars show that 36.36% of all beats closed the day red and 37.5% of "clear" beats (≥5pp) also finished down. Put simply, more than one-third of beats — including sizable surprises — still lost ground by the close, which aligns with the observed gap-up-then-fade behavior and supports the idea that guidance and pre-print positioning often dominate the headline surprise.
Event-level summary (last ~3 years; up to 20 rows)
| reported_date | day0_date | surprise_pct | day0_return | gap_return | intraday_return | direction | closed_red |
|---|---|---|---|---|---|---|---|
| 2023-08-23 | 2023-08-23 | 30.43 | 0.0964 | 0.0098 | 0.0857 | beat | false |
| 2023-11-21 | 2023-11-21 | 19.64 | -0.0294 | 0.0006 | -0.0299 | beat | true |
| 2024-02-21 | 2024-02-21 | 11.93 | 0.0806 | 0.0055 | 0.0747 | beat | false |
| 2024-05-22 | 2024-05-22 | 9.48 | 0.0562 | 0.0003 | 0.0559 | beat | false |
| 2024-08-28 | 2024-08-28 | 7.94 | -0.0842 | 0.0038 | -0.0876 | beat | true |
| 2024-11-20 | 2024-11-20 | 8 | -0.0369 | 0 | -0.0369 | beat | true |
| 2025-02-26 | 2025-02-26 | 4.71 | 0.0064 | 0.007 | -0.0006 | beat | false |
| 2025-05-28 | 2025-05-28 | 8 | 0.0429 | 0.0016 | 0.0412 | beat | false |
| 2025-08-27 | 2025-08-27 | 3.96 | -0.035 | -0.0003 | -0.0347 | beat | true |
| 2025-11-19 | 2025-11-19 | 4.84 | 0.0845 | 0.0044 | 0.0798 | beat | false |
| 2026-02-25 | 2026-02-25 | 6.58 | 0.0157 | 0.0034 | 0.0123 | beat | false |
| 2026-05-20 | 2026-05-20 | -100 | -0.004 | 0.0079 | -0.0118 | miss | true |
The takeaway
Short answer: No — over the last ~3 years NVDA's same-day close is only weakly tied to EPS surprise, and a meaningful share of beats still finished the day red. The numbers show a very small linear link (Pearson r = 0.184, Spearman ρ = 0.357) with R² = 0.034, so surprise explains only about 3.4% of day-0 return variance. The OLS slope is ~0.000320 return per 1 percentage-point surprise (about +0.32% per 10pp), meaning even large surprises imply modest expected moves. Beats averaged about +1.79% (N=11) versus the single miss at about −0.40% (N=1), but the Welch test gave p = 1.00 — no statistically significant difference. In practice 36.4% of beats closed red and 37.5% of “clear beats” (≥5pp) did too, and 36.4% of beats showed a gap-up that faded intraday, so selling into beats is common. Bottom line: the evidence leans toward surprise being a weak predictor — useful as one input, but far from a reliable trigger for same-day direction.
The fine print
- Day-0 is the first trading day on/after reported_date; after-hours prints are mapped to the next trading day.
- Sample is small: only 12 earnings events over the period, so correlations are noisy.
- The miss group is a single observation (N=1), so beat-vs-miss comparisons are fragile.
- This analysis doesn't model guidance, revenue surprises or pre-print positioning, which are likely key confounders.