AI Research AMDNVDASPY

AMD–NVDA daily-return correlation on SPY selloff days (last ~3 years)

They do not converge to a perfect hedge. We measured AMD–NVDA daily-return correlation over the last ~3 years (751 trading days), conditioning on SPY tape to see whether correlation climbs toward 1 on the market’s worst days. The thesis — that already-high co-movement would spike and erase diversification exactly when you need it — is only partially borne out.

Unconditional Pearson r is 0.567 and the worst SPY decile shows r=0.582 (N=76), rising to r=0.606 in the bottom 5% and r=0.650 for SPY ≤ −2% days (N=14). A Fisher z test gives p=0.244, so the upward drift into the mid‑0.6s in deep selloffs is plausible but noisy. Below is the full methodology, statistics, charts, and the detailed breakdown of why this is a suggestive lean rather than a definitive collapse of diversification.

The research question

For AMD over the past ~3 years, does its daily return correlation with NVDA climb toward 1 on the market's worst days, so the diversification between the two chip names evaporates exactly when it matters? Thesis: their everyday correlation is already high but jumps even higher on big SPY down-days, confirming that correlations converge in a selloff and offer little cushion when you need it most.

How this was measured

Resampled AMD, NVDA, and SPY minute bars to daily close-to-close returns. Aligned on common trading days and restricted to the most recent ~36 months ending at the latest overlapping date. Computed the unconditional AMD–NVDA Pearson correlation, then conditioned on SPY tape: (i) decile buckets from worst to best SPY days via qcut, reporting correlation within each decile; (ii) explicit tail subsets — bottom 10%, bottom 5%, and SPY ≤ -2% days. A Fisher z-test compares correlation on the worst decile versus the complement (disjoint samples).

The key numbers

Trading days analyzed
751
2023-06-01 to 2026-05-29
Unconditional Pearson r (AMD vs NVDA)
0.567
All days in window
Worst-decile r (bottom ~10% SPY days)
0.582
N=76 days
Rest-of-days r (ex worst decile)
0.479
N=675 days
Fisher z p-value (worst decile vs rest)
0.2440
p=0.2440 ≥ 0.05 → no clear difference
Bottom-5% SPY-day r
0.606
N=38 days
SPY ≤ -2% day r
0.650
N=14 days
Best-decile r (top ~10% SPY days)
0.546
N=75 days

Reading the numbers

Over 751 trading days AMD and NVDA show a moderate positive correlation of about 0.57. On the worst 10% of SPY days the correlation is about 0.582 (a tiny increase) but the Fisher z p-value 0.244 says that increase isn't a clear statistical jump.

The charts

AMD–NVDA correlation by SPY return decile (worst→best)
What this chart says

This bar chart splits days into SPY-return deciles from worst (D1) to best (D10) and plots the AMD–NVDA Pearson r for each decile. The eye is drawn to the leftmost bar (D1 = 0.5816) — the worst-decile correlation is moderately high — and the surprising dip through the middle deciles (D6 = 0.0238, D7 = -0.0344) before rising again into D10 (0.5458). That pattern shows correlation is not a monotonic climb toward 1 on bad days; the worst decile is a bit higher than many deciles but far from perfect correlation and there are mid-market days with near-zero or negative correlation.

AMD vs NVDA daily returns on worst-decile SPY days
What this chart says

This scatter plots the 76 worst-decile SPY days of AMD vs NVDA returns; the cloud of points follows a positive slope consistent with the decile r = 0.5816. Both means are negative (AMD mean = -0.0381, NVDA mean = -0.0363) and the ranges show some dispersion (AMD min -0.1177 to max 0.0498; NVDA min -0.1536 to max 0.0262). In plain terms, on bad market days the pairs tend to move together on average, but they are not perfectly collinear — there is still spread and room for non-identical moves rather than a collapse to r = 1.

Conditional AMD–NVDA correlation by SPY decile

SPY decileN daysPearson r
D1760.5816
D2750.509
D3750.2722
D4750.3036
D5750.2853
D6750.0238
D775-0.0344
D8750.2696
D9750.1153
D10750.5458

The takeaway

No — AMD and NVDA do become slightly more correlated on bad SPY days, but the link does not run to 1 and diversification does not fully evaporate. The unconditional Pearson r is 0.567, while the worst SPY decile shows r=0.582 (N=76) versus r=0.479 on the other 675 days. In deeper tails the number edges up — bottom-5% days show r=0.606 (N=38) and SPY ≤ -2% days hit r=0.650 (N=14) — but those are still far from perfect correlation. A Fisher z test comparing the worst decile to the rest gives p=0.244, i.e., roughly a 24-in-100 chance the observed decile difference is just noise, so this is a suggestive lean not a definitive signal. Given the small N in the tails and the volatility of tail estimates, treat the upward drift as plausible but noisy. Practical takeaway: baseline co-movement is already high (~0.57) and may rise modestly into the mid-0.6s on extreme down-days, so don’t rely on AMD–NVDA divergence as a reliable crash hedge.

The fine print