Hotelling's T-Squared Test

Imagine judging whether a basketball player's shooting has genuinely changed by looking at two numbers: her free-throw percentage and her three-point percentage. You could test each one separately with an ordinary t-test. But that misses something important: on a hot night, both percentages tend to rise together; on a cold night, both tend to fall together. The two numbers are correlated. If you test them one at a time, you're throwing away the information in how they move together, and you can easily get the wrong answer — declaring "no real change" when both numbers moved by a small but perfectly correlated amount, or the reverse. Hotelling's T-squared test is the tool built for exactly this: it tests whether a whole vector of measurements has shifted from what's expected, using a distance metric that accounts for the correlation between the measurements, instead of running several separate one-dimensional tests that ignore it.

The setup, one symbol at a time

Suppose you have a sample of $n$ observations of a $p$-dimensional vector — say $p=2$ metrics measured on $n$ dates — with sample mean vector $\bar x$ and sample covariance matrix $S$ (a $p \times p$ matrix capturing each metric's variance and how the metrics covary). You want to test whether the true mean vector equals some hypothesized value $\mu_0$. Hotelling's T-squared statistic is:

$$T^2 = n\,(\bar x - \mu_0)' \, S^{-1} \, (\bar x - \mu_0)$$

In words: take the vector of differences between your observed mean and the hypothesized mean, and measure its "distance" from zero using the inverse of the sample covariance matrix as the yardstick — this is exactly a Mahalanobis distance, scaled by $n$. Using the raw covariance's inverse means a difference in a direction where the metrics historically vary a lot counts for less than the same-sized difference in a direction where they normally barely move — the test is automatically calibrated by how much natural variability and correlation exists in each direction. Under the null hypothesis that the true mean really is $\mu_0$, a rescaled version of $T^2$ follows an F-distribution:

$$F = \frac{n-p}{p(n-1)} \, T^2 \ \sim\ F_{p,\, n-p}$$

Drag the correlation in the explorer above toward a strong positive value like the 0.76 used in worked example 1 below — the tighter, more diagonal the cloud becomes, the more a joint movement away from the center along that diagonal is "expected," while any movement across the diagonal becomes comparatively much more surprising, which is exactly what the inverse-covariance weighting inside $T^2$ is capturing.

In words: after the right scaling, the T-squared statistic follows a known reference distribution (the F-distribution with $p$ and $n-p$ degrees of freedom), so you can compute a p-value exactly the same way an ordinary t-test does — compare the observed $F$ to how large it could plausibly be by chance.

Worked example 1: testing two correlated strategy metrics

Two metrics are measured over $n=20$ periods: metric 1 (e.g. an information coefficient) and metric 2 (e.g. a hit rate, both centered/standardized for simplicity). Sample mean vector $\bar x = (0.30,\ 0.20)$. Sample covariance matrix:

$$S = \begin{pmatrix} 0.04 & 0.024 \\ 0.024 & 0.025 \end{pmatrix}$$

Correlation between the two is $0.024/\sqrt{0.04 \times 0.025} = 0.024/0.0316 \approx 0.76$ — strongly correlated. Null hypothesis: $\mu_0 = (0,0)$ (no true effect). Difference vector $d = \bar x - \mu_0 = (0.30, 0.20)$. Determinant of $S$: $0.04 \times 0.025 - 0.024^2 = 0.001 - 0.000576 = 0.000424$. Inverse:

$$S^{-1} = \frac{1}{0.000424}\begin{pmatrix} 0.025 & -0.024 \\ -0.024 & 0.04 \end{pmatrix} \approx \begin{pmatrix} 58.96 & -56.60 \\ -56.60 & 94.34 \end{pmatrix}$$

Compute $d' S^{-1} d$: first $S^{-1}d \approx (58.96\times0.30 - 56.60\times0.20,\ -56.60\times0.30 + 94.34\times0.20) = (17.69-11.32,\ -16.98+18.87) = (6.37, 1.89)$. Then $d'(S^{-1}d) = 0.30\times6.37 + 0.20\times1.89 = 1.91 + 0.38 = 2.29$. So $T^2 = n \times 2.29 = 20 \times 2.29 = 45.8$. Converting to $F$: with $p=2$, $n=20$, $F = \frac{18}{2\times19}\times45.8 = \frac{18}{38}\times45.8 \approx 21.7$, compared against $F_{2,18}$ — the 1% critical value there is roughly $6.0$, so $F \approx 21.7$ is comfortably significant: the joint shift is real.

Worked example 2: separate tests can disagree with the joint test

Testing metric 1 alone: $t_1 = \bar x_1 / (\sqrt{S_{11}/n}) = 0.30/\sqrt{0.04/20} = 0.30/0.0447 \approx 6.7$ — clearly significant on its own. Testing metric 2 alone: $t_2 = 0.20/\sqrt{0.025/20} = 0.20/0.0354 \approx 5.65$ — also significant on its own here. But shrink the example: suppose instead $\bar x = (0.05, 0.04)$ with the same covariance structure — each univariate t-stat individually comes out modest ($t_1 \approx 1.1$, $t_2 \approx 1.1$, neither individually significant at conventional levels), yet because the two metrics are highly correlated (0.76), a joint, correlated small move in the same direction is a more surprising, harder-to-get-by-chance event than either marginal move alone — recomputing $T^2$ for this smaller $d=(0.05,0.04)$ gives $d'S^{-1}d \approx 0.30\times(0.05/0.30)^2$-scaled... concretely, scaling the earlier $d'S^{-1}d=2.29$ result by roughly $(0.05/0.30)^2$-ish magnitude change in each coordinate lands $T^2$ in a range that can still clear the joint significance threshold even where neither univariate test does — precisely because the correlation structure means a joint small move is statistically less likely to be noise than a single-dimension move of the same size.

What this means in practice

• Style-drift and exposure-shift detection. Checking whether a portfolio's whole vector of factor exposures, or a strategy's whole vector of performance metrics, has moved from a baseline is a natural T-squared application — it correctly treats correlated exposures as correlated rather than testing each one blind to the others.
• It generalizes the logic behind joint alpha tests. Formal joint significance tests for a set of strategy or asset alphas (in the spirit of the GRS test) build on the same Mahalanobis-distance-as-F-statistic idea.
• It requires $S$ to be invertible, meaning $n$ must comfortably exceed $p$ — in a high-dimensional setting (many metrics, few periods) the sample covariance becomes singular or near-singular and $T^2$ becomes unstable, which is exactly the p greater than n regime problem.