Minimum Covariance Determinant Estimation

The ordinary sample covariance matrix uses every observation equally, which means a small number of extreme days — a flash crash, a data error, an earnings surprise — can dominate the entire estimate and distort every correlation in a portfolio, since covariance estimation is notoriously sensitive to outliers. Minimum Covariance Determinant (MCD) estimation fixes this by deliberately searching for the subset of observations, of a chosen size (typically around half the data), whose covariance matrix has the smallest determinant.

The determinant of a covariance matrix is a measure of the overall "spread" or volume of the data cloud it describes — a small determinant means a tight, well-behaved cluster. By searching for the subset of points with the smallest possible determinant, MCD is effectively finding the densest, most "normal" core of the data and building the covariance estimate from that core alone, deliberately setting aside the points that would inflate the spread — which are typically the outliers.

Once that clean core is identified, MCD reweights the full dataset based on distance from that core's covariance estimate, down-weighting or excluding points that look statistically anomalous relative to it, and produces a final robust covariance estimate. This is directly useful for portfolio risk models: an ordinary covariance matrix estimated across a period containing a crisis month will have correlations badly distorted by that one month, while an MCD-based estimate is far less swayed by it.