When correlations concentrate

Exploring the spectral collapse of the equity cross section across the 2008 and 2020 crashes.

Hello and welcome back to another paid post!

Today we will take a look at how when markets crash, the correlation matrix of the equity cross section collapses. The market factor absorbs a rising share of total variance and the effective number of independent factors falls sharply. This spectral concentration replicates cleanly across both the 2007-08 financial crisis and the 2020 Covid crash. But the same signal, run as a causal early-warning indicator, only forecasts the crisis that built up endogenously, and the exogenous shock is invisible to it in advance.

Let’s dive right in.

---

Introduction

Igor Halperin’s July 2026 paper introduces an approach called Observable Matrix Dynamics that tracks the equity cross section through the trajectory of a fixed-size distance matrix and its spectrum. We run two fixed-size matrix observables across the crisis decades using the current S&P 500 universe with daily adjusted closes, and the central empirical finding is that the correlation spectrum collapses onto the market factor at the 2008 and 2020 onsets.

A separate section tests whether these fragility signals actually forecast the crash they concentrate on, and finds that only the endogenously building 2008 crisis is predictable in advance.

“An endogenous fragility measure can forecast a crisis that grows out of the correlation structure itself, and cannot forecast an exogenous shock such as a pandemic, or a dispersed decorrelated unwind.”

---

Data & Methodology

The setup is a rolling Pearson correlation matrix on the S&P 500 cross section over a 504-day window, from which we extract the eigenvalues and read two summary statistics. The first is the market-factor share, the largest eigenvalue divided by the sum of all eigenvalues, which measures how much of total return variance is absorbed by the single largest common factor. The second is the participation ratio, defined as the square of the sum of eigenvalues divided by the sum of squared eigenvalues, which counts the effective number of independent factors. Both are standard diagnostics from the random matrix theory of correlation matrices, and both are known to move sharply at crisis onsets.

We compute these statistics across two crisis periods and average them over the calm year before onset and the first two months of the crash. The universe is current S&P 500 constituents with full history back to each period’s two-year pre-roll, filtered to 349 names for the 2007-08 window and 353 names for the Covid window. This universe is survivorship-biased by construction, which means 2008-era failures like Lehman, Bear Stearns, and Washington Mutual are absent from our sample. If anything this should understate the correlation stress at the 2008 onset, since the survivors are disproportionately the firms that made it through the crisis intact.