Mahalanobis Distance and Multivariate Outliers

Picture a scatter plot of a stock's daily return against its sector's daily return. Normally these move together — up days tend to come together, down days tend to come together — so the cloud of points stretches out as a diagonal ellipse rather than a circle. Now consider two unusual days. Day A: the stock is up 3% and the sector is down 3%. Day B: the stock is up 6% and the sector is up 5.8%. Which day is the stranger one? Measured by plain distance from the center of the cloud, Day B looks far more extreme — both numbers are large. But Day B sits right on the diagonal, exactly where the normal relationship predicts a big up day for both should land. Day A, despite having smaller individual numbers, breaks the relationship entirely — the stock and its sector moved in opposite directions, which almost never happens. Day A is the real outlier. Ordinary Euclidean distance from the mean gets this backwards, because it ignores correlation. Mahalanobis distance is built to get it right.

large Mahalanobis large Mahalanobis on-diagonal: large Euclidean, small Mahalanobis

The formula, one symbol at a time

For a point $x$ (a vector of $p$ measurements) with population or sample mean $\mu$ and covariance matrix $\Sigma$, the Mahalanobis distance squared is:

$$D^2 = (x - \mu)' \, \Sigma^{-1} \, (x - \mu)$$

Drag the correlation slider in the explorer above toward a strong value like 0.8, matching the worked example below — notice how the cloud tightens into a diagonal band; a point that strays off that band, even slightly, is exactly what Mahalanobis distance is built to flag, regardless of how close it sits to the center in plain Euclidean terms.

In words: take the vector of how far $x$ is from the mean on each variable, and instead of just squaring and summing those raw differences (which is what plain Euclidean distance does), weight them by the inverse of the covariance matrix. The inverse covariance automatically down-weights directions where the variables naturally vary a lot together (so a big joint move there isn't surprising) and up-weights directions that go against the normal correlation pattern (so a small move in an unusual direction counts for a lot). When $\Sigma$ is diagonal (no correlation, unequal variances), this reduces to ordinary scaled Euclidean distance — each variable divided by its own variance. It's the off-diagonal, correlation terms in $\Sigma^{-1}$ that do the real work of catching relationship-breaking outliers.

Worked example 1: computing it explicitly

Two variables, stock return $x_1$ and sector return $x_2$, with mean $\mu = (0, 0)$ (both centered), variance of each $\text{Var}(x_1) = \text{Var}(x_2) = 4$ (in percent-squared units), and covariance $\text{Cov}(x_1,x_2) = 3.2$ (correlation $= 3.2/4 = 0.8$, strongly positive). So:

$$\Sigma = \begin{pmatrix} 4 & 3.2 \\ 3.2 & 4 \end{pmatrix}$$

Determinant: $4\times4 - 3.2\times3.2 = 16 - 10.24 = 5.76$. Inverse:

$$\Sigma^{-1} = \frac{1}{5.76}\begin{pmatrix} 4 & -3.2 \\ -3.2 & 4 \end{pmatrix} \approx \begin{pmatrix} 0.694 & -0.556 \\ -0.556 & 0.694 \end{pmatrix}$$

Take point A $= (3, -3)$ (stock up 3, sector down 3 — off the diagonal). Euclidean distance from origin: $\sqrt{3^2+3^2} = \sqrt{18} \approx 4.24$. Mahalanobis: $D^2 = (3,-3)\,\Sigma^{-1}\,(3,-3)'$. First $\Sigma^{-1}(3,-3)' = (0.694\times3 - 0.556\times(-3),\ -0.556\times3+0.694\times(-3)) = (2.08+1.67,\ -1.67-2.08) = (3.75,\ -3.75)$. Then $D^2 = 3\times3.75 + (-3)\times(-3.75) = 11.25+11.25 = 22.5$, so $D \approx 4.74$ — the Mahalanobis distance is actually larger than the Euclidean one here, because moving off the diagonal is exactly what the correlation structure says shouldn't happen.

Worked example 2: the on-diagonal point

Take point B $= (6, 5.8)$ (both up, close to the diagonal). Euclidean distance: $\sqrt{6^2+5.8^2} = \sqrt{36+33.64}=\sqrt{69.64}\approx 8.35$ — much larger raw distance than point A. Mahalanobis: $\Sigma^{-1}(6,5.8)' = (0.694\times6 - 0.556\times5.8,\ -0.556\times6+0.694\times5.8) = (4.16-3.22,\ -3.34+4.02) = (0.94,\ 0.68)$. Then $D^2 = 6\times0.94 + 5.8\times0.68 = 5.64+3.94 = 9.58$, $D\approx3.10$. Despite having a much larger Euclidean distance (8.35 vs 4.24), point B has a smaller Mahalanobis distance (3.10 vs 4.74) than point A — confirming the earlier intuition precisely: the off-diagonal move is the real outlier, the on-diagonal move is comparatively ordinary, and only Mahalanobis distance gets the ranking right.

What this means in practice

• Multivariate anomaly detection. Flagging an unusual trading day, an anomalous combination of risk factor exposures, or a data point that breaks a normally reliable correlation (a stock decoupling from its sector, a hedge that stops behaving like a hedge) is exactly a Mahalanobis-distance problem, not a per-variable threshold problem.
• It underlies robust covariance estimation. High-breakdown estimators like the Minimum Covariance Determinant algorithm iteratively use Mahalanobis distance to identify and downweight the points least consistent with the bulk of the data.
• It's the building block of Hotelling's T-squared test — that test is precisely a Mahalanobis distance from a hypothesized mean, scaled and referred to an F-distribution.
• It needs a trustworthy $\Sigma$. Since $D^2$ depends on $\Sigma^{-1}$, and $\Sigma$ itself can be distorted by the very outliers you're trying to detect, naive Mahalanobis distance using the ordinary sample covariance can be self-defeating — which is exactly why outlier-resistant covariance estimators exist as an input to a more reliable Mahalanobis calculation.