Random Matrix Theory of Equity Cross-Correlations
Which apparent equity correlations are indistinguishable from random-matrix noise?
Compare the empirical eigenvalue spectrum with a random-matrix null matched to panel dimensions and dependence assumptions. In the cited 1994?1995 equity sample, many modes resembled noise; that result cannot label every low eigenvector in another market.
The reported noise-compatible modes come from the 1994?1995 equity panel and its specified null.
Evidence map
| Aspect | Finding |
|---|---|
| What it is | The application of random matrix theory to the empirical correlation matrix of a large stock universe, asking which eigenvalues and eigenvectors carry information and which are indistinguishable from those of a random matrix. |
| Key result / formula | Plerou, Gopikrishnan, Rosenow, Amaral & Stanley (1999) built the correlation matrix of the 1000 largest US stocks over 1994-1995 from 30-minute returns and compared its spectrum with the prediction for a random matrix of the same N/T. |
| Why it matters for backtesting | A factor, cluster or apparent structure extracted from a correlation matrix should be compared with a null matched to the sample size and data properties. |
What it is
The companion note on wide panels states the Marchenko-Pastur band; this note records what was actually found in equity data, and the review that consolidated the toolkit.
Key result / formula
The bulk of eigenvalues fell inside the noise band, and its finer statistics — the spacing of consecutive eigenvalues, the distribution of eigenvector components — matched the Gaussian orthogonal ensemble: the "universal" part of the title. The "non-universal" part is the handful of eigenvalues above the band: the largest, far above the edge, with an eigenvector loading on nearly all stocks (the market mode), and a few below it whose eigenvectors are not spread across all stocks. Bouchaud & Potters (2009) review what followed: generalisations of the Marchenko-Pastur law to correlated and heavy-tailed data, the statistics of the largest eigenvalue, random singular value decomposition for rectangular matrices, free-probability tools, and the direct use in portfolio construction — where the in-sample risk of a minimum-variance portfolio built on the raw matrix underestimates the realised out-of-sample risk by factors that depend only on N/T in the pure-noise case, and cleaning the spectrum removes most of the gap.
Why it matters for backtesting
In the 1994-1995 equity panel studied by Plerou and colleagues, much of the spectrum was compatible with a random-matrix benchmark and several extreme modes differed from it. This does not classify every low eigenvector in a new panel as noise, nor does an in-sample Sharpe ratio decide that question. Set a suitable null, estimate uncertainty, and test whether a proposed component is stable across disjoint periods and useful out of sample. The textbook band assumes independent entries and can be distorted by heavy tails, serial dependence and volatility clustering. Document those assumptions before assigning an economic meaning to an eigenvector.
Source
Plerou, Gopikrishnan, Rosenow, Amaral & Stanley, "Universal and Nonuniversal Properties of Cross Correlations in Financial Time Series", Physical Review Letters 83(7), 1999, 1471-1474; Bouchaud & Potters, "Financial Applications of Random Matrix Theory: a short review", 2009, arXiv:0910.1205 (in The Oxford Handbook of Random Matrix Theory, Oxford University Press, 2011). Primary source