Eigenvalue Spectrum of Large-Cap Equity Returns: Evidence for Non-Random Clustering Beyond Market-Wide Factors
Question
Does the eigenvalue spectrum of the correlation matrix among large-cap equities deviate from the random-matrix null hypothesis in a manner consistent with structured clustering of returns, and do the dominant eigenvectors reveal interpretable factor structure beyond a single market factor?
Method
We computed the eigenvalue decomposition of the return correlation matrix for 11 large-cap U.S. equities (AAPL, AMZN, BAC, CVX, GOOGL, JNJ, JPM, META, MSFT, NVDA, PFE, XOM) over the period 2010-01-01 through 2024-12-31, using daily adjusted-close returns from yfinance (n = 3772 observations). The null hypothesis is the Marchenko-Pastur (MP) distribution for random correlation matrices: under purely random co-movement, eigenvalues should lie within the MP bounds [0.8949, 1.1109] given the ratio q = n_assets / n_obs = 0.003. Eigenvalues exceeding the upper bound 1.1109 are statistically distinguishable from noise and indicate genuine common factors. We report the top 10 eigenvalues, the count of significant factors, the variance explained by the top factor and by all significant factors, and the loadings of the top two eigenvectors. To assess time variation, we recomputed the eigenvalue spectrum within each calendar year 2010–2024 on the same data and method, counting significant factors per year.
Result
The top eigenvalue is 5.1473, far exceeding the Marchenko-Pastur upper bound of 1.1109. The second eigenvalue is 1.5929, also above the threshold. The third eigenvalue is 0.9728, below the upper bound and within the MP interval, as are all subsequent eigenvalues (0.7274, 0.5580, 0.5024, 0.4561, 0.3919, 0.3428, 0.1652). Thus n_significant_factors = 2: two eigenvalues are statistically distinguishable from random-matrix noise.
The top factor (eigenvalue 5.1473) explains 46.79% of total variance. The two significant factors together explain 61.28% of variance. The remaining nine eigenvalues, consistent with the MP null, account for the residual 38.72%.
Factor 1 loadings (top three by absolute magnitude):
- JPM: -0.343
- MSFT: -0.335
- BAC: -0.331
All 11 assets load negatively on Factor 1 with magnitudes between -0.25 and -0.35 (not shown in full), indicating a broad market factor with near-uniform exposure across the universe.
Factor 2 loadings (top three by absolute magnitude):
- AMZN: -0.399
- XOM: +0.395
- CVX: +0.368
Factor 2 exhibits a clear sector split: technology/consumer (AMZN, MSFT, AAPL, META, NVDA, GOOGL) load negatively; energy (XOM, CVX) and financials (JPM, BAC) load positively; healthcare (JNJ, PFE) load near zero. This is a growth-versus-value or tech-versus-energy contrast.
Time variation (rolling per-year significant factor count):
- 2010–2015: 1 significant factor per year
- 2016–2017: 2 factors
- 2018–2019: 1 factor
- 2020–2024: 2 factors every year
The emergence of a persistent second factor from 2020 onward coincides with the post-pandemic regime of heightened sector dispersion (tech rally, energy volatility, rate-sensitive sector divergence).
Interpretation
The eigenvalue spectrum provides strong evidence that large-cap equity returns are not governed by random co-movement. Two eigenvalues lie well outside the Marchenko-Pastur bounds, rejecting the null hypothesis of a purely noise-driven correlation matrix. The top eigenvalue (5.1473) is 4.6 times the MP upper bound, indicating a dominant common factor—interpretable as the broad market or systematic risk—that drives nearly half of all variance. The second eigenvalue (1.5929), 1.4 times the threshold, captures a persistent sector or style factor orthogonal to the market.
The loadings confirm this interpretation. Factor 1 is a market factor: all assets load with the same sign and similar magnitude, consistent with a common exposure to aggregate risk (economic growth, monetary policy, risk appetite). Factor 2 is a sector/style factor: the sign split between technology/consumer and energy/financials reflects the well-documented growth-value divide. The negative loading of AMZN (-0.399) and positive loading of XOM (+0.395) are nearly symmetric, suggesting a zero-cost long-short portfolio that isolates this dimension of risk.
The time dynamics reveal that the second factor is not a permanent feature but emerges in regimes of elevated cross-sectional dispersion. From 2010 to 2015, only one factor exceeded the MP bound—consistent with a low-volatility, low-dispersion environment where sector differences were compressed. The brief appearance of two factors in 2016–2017 (energy collapse, tech acceleration) and their persistent presence from 2020 onward (pandemic, inflation, rate cycle) align with known macro regimes. The rolling-window evidence shows that factor structure is time-varying and regime-dependent, not a static property of the universe.
What the result does not support: the original research angle posited a "physics gravity model" in which market cap acts as mass and correlation distance as gravitational distance, with the hypothesis that dominant eigenvectors would align with market-cap-weighted distance metrics. The computed result does not test or validate this analogy. The eigenvalue spectrum and loadings are standard outputs of principal component analysis on a correlation matrix; they do not incorporate market-cap weighting in the construction of the correlation matrix, nor do they compute a distance metric or test for alignment with a gravity-inspired clustering. The result is a rejection of the random-matrix null and an identification of interpretable factors (market, sector/style), but it is silent on whether these factors correspond to a cap-weighted "gravitational" structure. The gravity metaphor remains untested.
The result does support the broader claim that large-cap equity returns exhibit low-dimensional structure: 61.28% of variance is captured by two factors, with the remaining variance consistent with idiosyncratic noise. This is a well-known empirical regularity in equity markets, formalized in factor models (CAPM, Fama-French, APT). The novelty here is the quantitative bound from random-matrix theory: the MP threshold provides a rigorous statistical criterion for distinguishing signal from noise, rather than an ad hoc choice of factor count. The fact that exactly two eigenvalues exceed the bound, and that the third eigenvalue (0.9728) lies comfortably within the noise interval, gives confidence that the two-factor model is not over-fitted.
Relation to the Literature
No closely related papers were retrieved for this computation. The result stands on its own as an application of random-matrix theory (Marchenko-Pastur distribution) to equity return correlations, a technique introduced in econophysics and quantitative finance over the past two decades. The finding of a dominant market factor and a secondary sector/style factor is consistent with decades of empirical asset pricing research (Sharpe's single-index model, Fama-French factors, principal component analysis of returns), but the use of the MP bound as a formal significance threshold is a more recent methodological contribution. The time variation in factor count (one factor in low-dispersion regimes, two in high-dispersion regimes) aligns with conditional factor model literature showing that cross-sectional return dispersion and the number of priced factors vary with the business cycle, volatility regime, and monetary policy stance.
The absence of a gravity-model literature citation reflects the fact that the "market cap as mass, correlation as distance" analogy is a heuristic metaphor, not a formalized theory with testable predictions in the asset pricing literature. The computed result does not engage with or validate that metaphor; it is a standard eigenvalue decomposition with a random-matrix significance test.
Limitations
Sample size and universe: The analysis covers 11 large-cap U.S. equities, a small and highly liquid subset of the market. The q-ratio (n_assets / n_obs = 0.003) is extremely low, which tightens the MP bounds and increases statistical power, but the small cross-section limits the generality of the factor structure. A broader universe (e.g., S&P 500 constituents) would test whether the two-factor result scales or whether additional sector, size, or liquidity factors emerge. The 15-year window (2010–2024) spans multiple regimes, but the rolling per-year analysis is in-sample within each year—there is no out-of-sample validation of the factor count or loadings.
Methodology: The eigenvalue spectrum is computed on the correlation matrix, not a covariance matrix, so the factors are scale-invariant and do not account for differences in volatility across assets. The result does not incorporate market-cap weighting in the correlation matrix construction, despite the research angle's emphasis on cap-weighted "mass." A cap-weighted correlation or distance metric would be a different computation and might yield different factor structure. The MP bounds assume i.i.d. returns, which is violated by autocorrelation, heteroskedasticity, and time-varying volatility in equity returns; the bounds are asymptotic (large n, large T) and may be conservative for finite samples.
Interpretation of Factor 2: The sector/style interpretation of Factor 2 is based on the sign pattern of loadings (tech negative, energy/financials positive) but is not formally tested against sector classifications or style indices. A more rigorous approach would regress the factor returns (eigenvector-weighted portfolio returns) on sector or style benchmarks to quantify the alignment. The loading magnitudes are modest (max 0.40), indicating that Factor 2 captures only a fraction of sector-specific variance; much of the sector variation remains in the residual eigenvalues.
Gravity model: The result does not test the gravity-model hypothesis. To do so would require: (1) constructing a market-cap-weighted distance metric (e.g., inverse of cap-weighted correlation, or a metric derived from cap ratios), (2) computing a "gravitational potential" or clustering measure based on that metric, and (3) testing whether the dominant eigenvectors align with (are spanned by) the cap-weighted distance structure. The current result is a standard PCA eigenvalue decomposition; it neither confirms nor refutes the gravity analogy.
Strengthening the result: Out-of-sample validation (e.g., compute eigenvalues on 2010–2019, test factor stability on 2020–2024) would assess whether the two-factor structure is robust or sample-specific. Expanding the universe to 50–100 assets would test scalability. Incorporating market-cap weights into the correlation matrix (e.g., a cap-weighted Mahalanobis distance or a gravity-inspired kernel) would directly test the physics analogy. Comparing the MP-based factor count to alternative methods (e.g., scree plot elbow, parallel analysis, cross-validation of factor models) would validate the threshold choice. Finally, computing the economic significance of the factors (e.g., Sharpe ratio of factor-mimicking portfolios, out-of-sample R² in return prediction) would assess whether the statistical factors are tradable or merely descriptive.
Research evidence, not investment advice.