N Versus T Tradeoffs in Panel Signal Tests

A signal test on 500 stocks over 250 trading days looks like it has 125,000 data points — plenty for a confident statistical conclusion, if you treat every stock-day observation as an independent draw. But it isn't: on any given day, the 500 stocks' returns are correlated with each other through common market and factor moves, and any given stock's returns are correlated across consecutive days through persistence and volatility clustering. The "effective" sample size for statistical confidence is much closer to whichever of $N$ (number of stocks) or $T$ (number of time periods) is actually independent — and if most of the correlation lives across stocks rather than across time, the real sample size is closer to $T$ alone, not $N \times T$.

An analogy: polling opinions in one country versus one household

Imagine trying to estimate the average height of adults by measuring 500 people from a single household on 250 different days. You have 125,000 measurements, but the true information content is roughly "one household's heights, measured many times" — the 500 "people" don't actually exist as 500 independent draws; if it's really one household of, say, 4 people measured repeatedly, your true independent sample size is 4, not 125,000. A panel of stock returns has an analogous problem in the other direction: on any single day, hundreds of stocks share the same market-wide shock, so cross-sectionally they're not independent either — the panel's real information content depends on how much genuinely independent variation exists along each dimension, and that has to be diagnosed, not assumed away by multiplying $N \times T$.

The tradeoff, one symbol at a time

In a stock-date panel with $N$ stocks and $T$ time periods, standard errors computed as if all $N \times T$ observations were independent (the naive OLS or naive correlation standard error) are typically too small — sometimes dramatically so — because of two separate sources of correlation: cross-sectional correlation (stocks moving together within a given date, from common factors) and time-series correlation (a given stock's residuals or signal values persisting across consecutive dates). The correct standard error needs to account for whichever of these dominates:

• If observations are correlated mainly within a date (across stocks) but roughly independent across dates, then the effective sample size for inference is closer to $T$ — you have $T$ genuinely independent "cross-sections," even though each cross-section contains $N$ stocks.
• If observations are correlated mainly within a stock (across time) but roughly independent across stocks, the effective sample size is closer to $N$.
• If both types of correlation are present (the realistic case), standard errors need to be clustered on both dimensions — double clustering, sometimes called Thompson or Cameron-Gelbach-Miller standard errors — to avoid understating uncertainty from either source.

Worked example: overstated confidence from a naive standard error

Suppose you test a signal on $N = 500$ stocks over $T = 12$ monthly periods (6,000 stock-month observations) and get a cross-sectional regression coefficient with a naive standard error assuming all 6,000 observations are independent, giving a t-statistic of 4.5 — seemingly quite significant. But suppose the true driver of any given month's cross-sectional result is largely a single common factor shock that month, meaning the 500 stocks within a month are highly correlated with each other and contribute much less independent information than 500 separate draws would. Clustering standard errors by month (acknowledging that within a month, the 500 observations share common shocks) might reveal the effective sample size for inference is closer to $T=12$ than to 6,000, inflating the standard error roughly proportional to $\sqrt{500}$ and dropping the t-statistic to something like 4.5 / 6 ≈ 0.75 — no longer significant at all, once the panel's true cross-sectional correlation structure is accounted for.

Worked example: recognizing which dimension actually helps

Suppose instead you extend the sample to $T = 120$ months (10 years) while keeping $N = 500$ stocks fixed. If the signal's true predictive power is stable and each month's cross-section is roughly independent of other months (a defensible assumption for many signals, since monthly market regimes aren't strongly autocorrelated at that frequency), then going from $T=12$ to $T=120$ genuinely improves statistical confidence by roughly $\sqrt{10}$, since you've added ten times as many independent cross-sections. By contrast, adding another 500 stocks to each existing month's cross-section ($N=500 \to N=1000$) adds much less independent information than it appears to, since those stocks are correlated with the existing 500 within the same month through the same common factor shocks — the marginal statistical benefit of a larger $N$ within an already-large cross-section is much smaller than the benefit of another independent time period.

What this means in practice

This distinction is central to honestly reporting statistical significance for any cross-sectional signal test, and it's the reason standard errors clustered by date (or by both stock and date) are standard practice in academic and practitioner panel-data work rather than plain OLS standard errors, which silently assume away exactly the correlation structure that's almost always present in financial panels (see panel data and fixed effects). Before trusting a large stock-date sample size as evidence of high confidence, check which dimension — $N$ or $T$ — the correlation structure actually allows to contribute independent information, since a test with an enormous $N$ but only a handful of genuinely independent time periods can be far less reliable than its raw observation count suggests.