Random-Matrix Limits on the Information Content of Empirical Correlation Matrices

Comparing the empirical eigenvalue spectrum of an S&P 500 correlation matrix to the Marchenko–Pastur prediction for random matrices, the authors show that the bulk of eigenvalues is statistically indistinguishable from noise; only a handful of large eigenvalues carry genuine structure. We read it as a hard limit on how much cross-sectional co-movement a desk can claim to measure from finite history.

For T observations of N series with ratio q = N/T, the noise band of eigenvalues lies in [λ_−, λ_+] = σ²(1 ± √q)²; eigenvalues falling inside the band are not distinguishable from those of a random matrix.

The result reframes diversification and factor exposure as statistically fragile when N/T is not small. Crucially, the noise band is widest exactly when the desk most wants fine cross-sectional structure — many names, short windows. Measured correlations in that regime are mostly dressing.

Cross-asset and crowding statements in memos should be qualified by the estimation ratio N/T; correlations estimated over short windows on broad universes are treated as low-evidence, not as a basis for confident co-movement claims.

The clean Marchenko–Pastur null assumes i.i.d. Gaussian returns; fat tails and non-stationarity shift the band, so the bound is indicative rather than exact.