# Research note — NVDA bounce after big down days

**To:** Elena Vasquez  
**Subject:** Research note — Does NVDA bounce more than SPY after −3% days?

## 1. Executive summary

After NVIDIA falls **3% or more** in a day, the next **five** trading days show a slightly higher average NVDA−SPY excess return than on other days (**~1.3%** vs **~0.8%**). A two-sample t-test does **not** reject the null of equal means (p ≈ **0.57**). The evidence does **not** support the hypothesis. I would **not** trade a bounce rule on this sample.

## 2. Question

After a NVDA daily return ≤ −3%, is the mean next-5-day NVDA−SPY excess return higher than on ordinary days?

## 3. Hypothesis

After NVDA daily returns of −3% or worse, the mean next-5-day NVDA−SPY excess return is higher than on other days.

## 4. Null hypothesis

After NVDA daily returns of −3% or worse, the mean next-5-day NVDA−SPY excess return equals the mean on other days.

*(Motivation, not part of the formal claim: volatile names often mean-revert after panic prints, so a stronger bounce than the market would be useful if it were reliable.)*

## 5. Method

- Data: NVDA and SPY daily returns, Yahoo Finance, Jan 2024–Jul 2026 (same CSVs as Quant 101).  
- Shock days: NVDA return ≤ −3% (**71** days).  
- Outcome: compound return over the **next** 5 sessions for NVDA and SPY; excess = NVDA − SPY.  
- Compare mean excess after shocks vs other days; Welch two-sample t-test against the null of equal means.

## 6. Results

| Group | n | Mean 5d NVDA | Mean 5d SPY | Mean excess (NVDA−SPY) |
| --- | ---: | ---: | ---: | ---: |
| After ≤ −3% day | 71 | ~1.8% | ~0.6% | **~1.3%** |
| Other days | 570 | — | — | **~0.8%** |

- t ≈ **0.57**, p ≈ **0.57** — we **fail to reject** the null at conventional α = 0.05.  
- Chart: `outputs/charts/quant301_main.png`.

## 7. Conclusion

**Statistically:** the evidence does **not** support the hypothesis. We do not reject the null of equal mean excess returns.

**Practically:** the point estimates differ by only about **0.4** percentage points over five days, with wide overlap. Even if a larger sample tipped the p-value, the effect would need to clear costs before it mattered for trading.

**Verdict:** reject deploying a bounce rule on this evidence.

## 8. Limitations

Shock days cluster in sell-offs (e.g. early 2025), so observations are not independent. Five-day windows overlap. Threshold −3% and horizon 5 were chosen without a hold-out. No costs. One AI-heavy regime.

## 9. Career reflection

I changed the question once — from “does NVDA bounce?” to “does it bounce **more than SPY**?” — because absolute bounce confuses market beta with stock-specific recovery. Choosing the question was harder than coding the forward returns. Failing to reject the null matched a sceptical prior more than a hopeful one. After 101–301 I still find quant work appealing when it looks like this: a narrow hypothesis, a boring test, and an honest “not significant.”