methodology
Marchenko-Pastur Filter
A random-matrix-theory tool that filters out covariance matrix eigenvalues which look like statistical noise.
The Marchenko-Pastur filter is a technique from random matrix theory that separates real structure in a correlation matrix from patterns that pure noise would produce. It keeps the strong, meaningful components of the matrix and flattens the weak ones that cannot be told apart from randomness. The optimizer then works with a cleaner picture of how assets relate.
Any correlation matrix can be broken into eigenvalues, each measuring how much of the total variation one independent pattern explains. In 1967 Vladimir Marchenko and Leonid Pastur showed that if returns were pure noise, these eigenvalues would fall inside a known band whose upper edge is (1 plus the square root of N/T) squared, where N is the number of assets and T the number of observations. The filter treats eigenvalues at or below that edge as noise, replaces them with their average, keeps the larger ones, rebuilds the matrix and resets the diagonal to 1.
The more assets you have relative to your history, the wider the noise band and the more eigenvalues get flattened. When stocks share a common market movement, the top eigenvalue sits far above the band, as the example below shows; weaker patterns, such as sector effects, may or may not clear it. Filtering helps most for large universes with short histories and matters little for a small portfolio with a long lookback. It can also remove weak but real structure, so it is a trade-off.
Example with illustrative numbers: 100 stocks with one year of daily data (T = 252) give N/T of about 0.40, so the noise edge is (1 + 0.63) squared, about 2.66. With 10 years of data (T = 2,520), the edge falls to about 1.44, so more structure survives. If every pair of stocks has a correlation of exactly 0.3, the largest eigenvalue is about 30.7 (1 plus 99 multiplied by 0.3), far above either edge.
On GNG, the Portfolio Optimizer offers Marchenko-Pastur covariance denoising as a switch in its Solver knobs, off by default, described as clipping noise eigenvalues and recommended for universes of 100 names or more (harmless for smaller ones). When it is on, it cleans the correlation part of the Ledoit-Wolf estimate for universes of 30 names or more, and it appears only for engines that use a covariance matrix, not for HRP or HERC.
Ledoit-Wolf Shrinkage is the first cleaning step, and Correlation explains the matrix being filtered.
Related terms
More glossary terms
Get full access to GNG Research
Create a free account to access portfolio tracking, advanced tools, and more.