Eigenvalue Phase Transition in Large-Cap Equity Returns: Evidence for Structural Coherence Beyond Random-Matrix Noise

Question

Does the eigenvalue spectrum of large-cap equity returns exhibit a statistically significant departure from the Marchenko-Pastur random-matrix null, and if so, does the number of significant factors vary systematically over time in a manner consistent with regime-dependent market coherence?

Method

We computed the eigenvalue spectrum of the return correlation matrix for 11 large-cap U.S. equities (AAPL, AMZN, BAC, CVX, GOOGL, JNJ, JPM, META, MSFT, NVDA, PFE, XOM) using daily adjusted-close returns from yfinance over the window 2010-01-01 through 2024-12-31 (n = 3772 observations). The sample ratio q = n_assets / n_obs = 0.003 is well within the asymptotic regime where the Marchenko-Pastur distribution provides a sharp null hypothesis for the eigenvalue density under the assumption of independent, identically distributed noise.

The Marchenko-Pastur upper bound for this q-ratio is 1.1109; the lower bound is 0.8949. Eigenvalues exceeding the upper bound are statistically distinguishable from random-matrix noise and indicate the presence of genuine common factors. We applied principal component analysis (PCA) to the correlation matrix and counted the number of eigenvalues above the MP upper threshold. To assess time variation, we recomputed the spectrum and factor count within each calendar year (in-sample, same method) from 2010 through 2024.

Result

Full-sample spectrum

The top ten eigenvalues of the correlation matrix are:

  1. λ₁ = 5.1473
  2. λ₂ = 1.5929
  3. λ₃ = 0.9728
  4. λ₄ = 0.7274
  5. λ₅ = 0.5580
  6. λ₆ = 0.5024
  7. λ₇ = 0.4561
  8. λ₈ = 0.3919
  9. λ₉ = 0.3428
  10. λ₁₀ = 0.1652

Two eigenvalues exceed the Marchenko-Pastur upper bound of 1.1109: the first (5.1473) and the second (1.5929). All remaining eigenvalues fall within or below the MP bulk, consistent with noise. The number of significant factors is therefore n = 2.

The first factor explains 46.79% of total variance; the two significant factors together explain 61.28% of variance.

Factor loadings

Factor 1 (the market factor) loads most heavily on:

  • JPM: −0.343
  • MSFT: −0.335
  • BAC: −0.331

All loadings are negative and of similar magnitude, indicating a broad common mode across financials and technology.

Factor 2 (a sector-contrast factor) loads:

  • AMZN: −0.399
  • XOM: +0.395
  • CVX: +0.368

This factor separates technology/consumer (negative loadings) from energy (positive loadings), capturing the well-known energy-versus-growth rotation dynamic.

Time variation

The per-year factor count (number of eigenvalues above the MP upper bound within each calendar year) is:

Year Significant factors
2010 1
2011 1
2012 1
2013 1
2014 1
2015 1
2016 2
2017 2
2018 1
2019 1
2020 2
2021 2
2022 2
2023 2
2024 2

From 2010 through 2015, only one eigenvalue exceeded the MP threshold in each year. Beginning in 2016, a second significant factor emerged, persisted through 2017, disappeared in 2018–2019, then reappeared in 2020 and remained present through 2024. The transition from one to two factors is not monotonic but exhibits clear regime structure.

Interpretation

Phase transition and structural coherence

The eigenvalue spectrum exhibits a sharp phase transition: two eigenvalues lie far above the Marchenko-Pastur upper bound (by factors of 4.6 and 1.4, respectively), while all others fall within or below the noise bulk. This is the signature of a low-dimensional factor structure embedded in high-dimensional noise, consistent with the hypothesis that large-cap equity returns are governed by a small number of common drivers rather than by pairwise idiosyncratic correlations.

The first factor is a broad market mode, loading uniformly (and negatively, by sign convention) across sectors. The second factor is a sector rotation, separating energy from technology and consumer discretionary. The emergence of this second factor is economically interpretable: it captures the oil-price and interest-rate sensitivity that differentiates cyclical energy stocks from growth-oriented technology.

Time variation and regime dependence

The rolling per-year factor count reveals that the second significant factor is not a permanent feature but appears and disappears in a pattern consistent with macroeconomic and volatility regimes:

  • 2010–2015 (one factor): The post-crisis recovery and low-volatility expansion were dominated by a single market factor. Sector dispersion was present but did not rise above the noise threshold.
  • 2016–2017 (two factors): The energy sector experienced a sharp recovery from the 2015–2016 oil crash, and the Trump election introduced policy uncertainty around energy and technology regulation. The sector-rotation factor became statistically significant.
  • 2018–2019 (one factor): The factor count dropped back to one, coinciding with the late-cycle flattening of sector dispersion and the dominance of Fed policy as the single common driver.
  • 2020–2024 (two factors): The COVID-19 shock, the subsequent inflation surge, and the Fed's tightening cycle reintroduced sharp sector divergence (energy versus growth, value versus momentum). The second factor has remained significant since 2020.

This time variation supports the hypothesis that the eigenvalue phase transition is not a static property of the asset universe but a dynamic marker of regime-dependent coherence. When macroeconomic or policy shocks induce sector-level divergence, the second factor crosses the MP threshold; when such shocks subside, it falls back into the noise bulk.

Relation to the gravity-model hypothesis

The computation question asked whether the eigenvalue spectrum marks a boundary where "gravity-model coherence breaks down." The result suggests the opposite interpretation: the presence of significant factors above the MP threshold is evidence for coherence, not against it. In a gravity model, market capitalization acts as mass and correlation distance as gravitational distance; the eigenvalue spectrum then measures the strength of the gravitational field. A single dominant eigenvalue corresponds to a uniform gravitational pull (the market factor); a second significant eigenvalue corresponds to a dipole or sector-level tilt in the field.

The breakdown of gravity-model coherence would manifest as a flat eigenvalue spectrum, with all eigenvalues near the MP bulk and no significant factors. We do not observe this. Instead, we observe a sharp hierarchy: two factors far above the threshold, then a clean drop into noise. This is the signature of a low-rank gravitational field, not of incoherence.

The time variation in factor count does, however, support a weaker version of the hypothesis: the dimensionality of the gravitational field is regime-dependent. In low-volatility, low-dispersion regimes, the field is one-dimensional (a single market factor). In high-volatility, high-dispersion regimes, the field is two-dimensional (market plus sector rotation). The phase transition is not a breakdown but a change in the rank of the coherence structure.

Limitations of the gravity analogy

The gravity-model framing is heuristic, not mechanistic. Physical gravity is a conservative force with a well-defined potential; financial correlation is not. The eigenvalue spectrum measures linear covariance structure, which is a second-moment property; it does not capture tail dependence, asymmetry, or higher-order cumulants. A more complete test of the gravity hypothesis would require:

  1. Explicit distance metrics: Define correlation distance (e.g., d = √(2(1 − ρ))) and test whether the correlation matrix is well-approximated by a distance-based kernel (e.g., exp(−d/ℓ) for some length scale ℓ).
  2. Mass-weighting: Weight the correlation matrix by market capitalization and test whether the weighted spectrum exhibits stronger coherence than the unweighted spectrum.
  3. Out-of-sample validation: The per-year factor counts are in-sample within each year. An out-of-sample test would estimate the factor structure in year t and test its predictive power for correlations in year t+1.

Limitations

Sample size and universe

The sample contains only 11 assets, all large-cap U.S. equities. The q-ratio of 0.003 is extremely small, which sharpens the MP threshold but also means the result is specific to this universe. A larger universe (e.g., the S&P 500) would have a higher q-ratio, a wider MP bulk, and potentially more significant factors. The two-factor structure observed here may be an artifact of the small, sector-diverse sample rather than a universal property of equity markets.

In-sample factor counts

The per-year factor counts are computed in-sample within each calendar year. This is appropriate for measuring the presence of structure but does not test whether the structure is predictive. A factor that is significant in-sample may be spurious or overfitted. An out-of-sample test—estimating the factor structure in a training window and testing its explanatory power in a holdout window—would strengthen the claim that the time variation is economically meaningful rather than a statistical artifact.

Eigenvalue spacing and universality

The Marchenko-Pastur distribution describes the bulk eigenvalue density but does not predict the spacing or repulsion of eigenvalues above the bulk. Random matrix theory predicts that eigenvalues of Gaussian random matrices exhibit level repulsion (the Wigner surmise), but this is a fine-scale property that requires much larger samples to test. The result here establishes that two eigenvalues are above the MP threshold but does not test whether their spacing is consistent with a specific universality class (e.g., Gaussian orthogonal ensemble versus Wishart ensemble).

Correlation versus causation

The time variation in factor count correlates with known macroeconomic and volatility regimes, but the computation does not establish causation. The emergence of the second factor in 2016, 2020, and beyond is consistent with oil-price shocks, policy uncertainty, and inflation, but it could also reflect changes in market microstructure (e.g., the rise of passive investing, the growth of sector ETFs) or data-generating properties (e.g., changes in the volatility of individual stocks). Distinguishing these mechanisms would require external regressors (e.g., VIX, oil prices, sector dispersion indices) and a formal attribution framework.

Absence of liquidity and volatility data

The computation question asked whether the eigenvalue threshold "correlates with realized liquidity and volatility regimes," but the result does not include liquidity or volatility data. The per-year factor counts are suggestive—2020 and 2022 were high-volatility years, and both exhibit two significant factors—but this is qualitative pattern-matching, not a quantitative test. A rigorous test would regress the per-year factor count on per-year realized volatility (e.g., average VIX or cross-sectional return dispersion) and liquidity proxies (e.g., bid-ask spreads, Amihud illiquidity).

Conclusion

The eigenvalue spectrum of large-cap equity returns exhibits a clear phase transition: two eigenvalues (5.1473 and 1.5929) lie far above the Marchenko-Pastur upper bound of 1.1109, while all others fall within the noise bulk. This is evidence for a low-dimensional factor structure—a market factor and a sector-rotation factor—embedded in high-dimensional noise. The number of significant factors varies over time, rising from one to two in periods of macroeconomic stress and sector divergence (2016–2017, 2020–2024) and falling back to one in low-volatility, low-dispersion periods (2010–2015, 2018–2019). This time variation is consistent with regime-dependent coherence: the dimensionality of the common factor structure is not fixed but responds to the economic environment.

The result does not support the hypothesis that the eigenvalue spectrum marks a breakdown of gravity-model coherence. Instead, it supports the interpretation that the spectrum measures the rank of the gravitational field: one-dimensional in calm regimes, two-dimensional in turbulent regimes. The phase transition is a change in dimensionality, not a loss of structure. A more complete test of the gravity hypothesis would require explicit distance metrics, mass-weighting, and out-of-sample validation, none of which are present in this computation. The result establishes the presence and time variation of significant factors but does not test their predictive power or their mechanistic connection to a gravity model.

Research evidence, not investment advice.