TSM vs NVDA — same-day co-movement and 1-day lead–lag over the last ~3 years
The supply‑chain story — that TSMC's moves should foreshadow Nvidia’s next session — does not survive a simple test over the past ~36 months. TSM and NVDA show strong same‑day co‑movement (Pearson r ≈ 0.712; OLS slope ≈ 0.87, explaining roughly half of NVDA’s daily variance), but one‑day lead–lag correlations collapse: TSM_t → NVDA_{t+1} is about −0.1336 (and NVDA_t → TSM_{t+1} is similar), while Granger tests return p = 1.00 in both directions.
We arrived at this by converting minute bars to daily closes across 751 overlapping days and measuring same‑day correlations, cross‑correlations k∈[−5,+5], lagged OLS, and Granger causality (lag=1). The full statistics, rolling metrics, charts and regression diagnostics are below; they show a tight same‑day link but no reliable next‑day forecasting edge.
For TSM over the past ~3 years, does the world's biggest chip foundry actually lead its marquee customer NVDA — do TSM's daily moves forecast NVDA's next session (or vice versa) — or do the two just move together with no exploitable sequence? Thesis: their same-day returns are tightly linked, but the cross-correlation at a one-day lag collapses to near zero in both directions, so the supply-chain 'who-moves-first' story offers no forecasting edge.
How this was measured
Resampled TSM and NVDA minute bars to daily closes, then computed close-to-close daily returns over the trailing 36 months in the overlapping window. Measured: (i) same-day Pearson correlation and 60-day rolling correlation; (ii) cross-correlation Corr[TSM_t, NVDA_{t+k}] across k ∈ [-5,+5] with k>0 meaning TSM leads; (iii) 1-day lead–lag correlations in both directions; (iv) OLS regressions for same-day and lagged relationships; (v) Granger causality (lag=1) p-values for TSM→NVDA and NVDA→TSM. Returns, not prices, are used to avoid spurious correlation.
The key numbers
Reading the numbers
TSM and NVDA move together on the same day: Pearson r = 0.7116 (N=751) and NVDA scales with TSM (OLS beta = 0.8699, r² = 0.506). One-day lag stats collapse to small negative values (~ -0.13 to -0.15) and Granger p = 1.0, so no detected lead.
The charts
The 60-day rolling same-day correlation stays high across the window: the 60d r averages 0.6896, runs from a low of 0.4328 to a high of 0.9119, and begins and ends around 0.74 and 0.73. Look at that persistent band near 0.7 — same-day comovement is large and stable. For the question of who leads whom, this chart says the relationship is contemporaneous and consistent, not that one reliably precedes the other.
This bar chart puts the relationship by lag front-and-center: the lag-0 bar is dominant at 0.7116, while the one-day offsets are small and negative (k = -1: -0.1336; k = 1: -0.1463) and the other lags hover near zero. The eye should go to those one-day bars — they do not show a positive, large lead in either direction. That pattern supports the conclusion that there is no exploitable one-day lead from TSM to NVDA or vice versa.
The same-day return scatter is a tight cloud along a positive slope: TSM returns span -0.1289 to 0.1449, NVDA -0.1536 to 0.2054, with N = 751 points. The OLS slope is 0.8699 and r² = 0.506, so roughly half of NVDA's same-day variance is explained by TSM same-day moves. This underlines strong contemporaneous coupling but, combined with the lag results, does not imply a reliable next-day forecasting edge.
Regression summaries (daily returns)
| model | beta | t_stat(beta) | r_squared | N |
|---|---|---|---|---|
| NVDA_t ~ TSM_t | 0.8699 | 27.722 | 0.5064 | 751 |
| NVDA_t ~ TSM_{t-1} | -0.1636 | -3.686 | 0.0178 | 750 |
| TSM_t ~ NVDA_{t-1} | -0.1197 | -4.044 | 0.0214 | 750 |
The takeaway
No — over the last ~36 months there is no evidence that TSM consistently leads NVDA (or vice versa) at a one-day horizon; they simply move together on the same day. The same-day relationship is strong: Pearson r = 0.712 and the OLS slope of NVDA_t on TSM_t is about +0.870 (r² ≈ 0.506, N = 751), so same-day returns explain roughly half of NVDA’s variance. By contrast the one-day cross-correlations are small and negative (TSM_t → NVDA_{t+1} r ≈ −0.1336; NVDA_t → TSM_{t+1} r ≈ −0.1463). The formal Granger tests give p = 1.00 in both directions — effectively no sign of lag‑1 causality. Lagged OLS coefficients are small (≈ −0.164 and −0.120) with tiny r² (≈0.018–0.021) — they can be statistically nonzero at this sample size but carry essentially no predictive power. Practical takeaway: the supply‑chain “who moves first” story does not produce a usable next‑day forecasting edge; focus on same‑day co‑movement or higher‑frequency intraday analysis instead.
The fine print
- High same‑day correlation may reflect shared market/sector factors, not direct causation.
- Daily closes can miss intraday leads; 5–15min bars might reveal microstructure timing.
- Lagged betas are tiny and explain ~1.8–2.1% of variance — large N (751) can make trivial effects statistically significant.
- Granger here used lag=1; multi‑day diffusion or event‑driven asynchronous reactions could be missed.