Does Market-Cap Gravity Induce a Spectral Phase Transition in Equity Returns?

Question

Does the eigenvalue spectrum of large-cap and mid-cap equity returns exhibit a spectral phase transition—analogous to an event horizon in gravitational physics—where coherence breaks beyond a correlation-distance threshold, and does the number of statistically significant factors scale with market-cap-weighted "gravitational mass" clustering?

Method

We computed the eigenvalue spectrum of the return correlation matrix for 12 large-cap and mid-cap U.S. equities (AAPL, AMZN, BA, GE, GOOGL, JNJ, JPM, META, MSFT, NVDA, PG, TSLA, WMT, XOM) over 3,772 daily observations spanning 2010-01-01 to 2024-12-31. The data source is yfinance daily adjusted-close returns. We applied principal component analysis (PCA) to the correlation matrix and compared the resulting eigenvalue distribution to the Marchenko-Pastur (MP) null distribution, which characterizes the eigenvalue spectrum of a purely random correlation matrix. Eigenvalues exceeding the MP upper bound (1.116, computed from the q-ratio of 0.003) are statistically distinguishable from random-matrix noise and indicate genuine collective modes. The MP lower bound is 0.8904. We counted the number of eigenvalues above the upper bound as the number of significant factors. To assess time variation, we recomputed the eigenvalue spectrum and factor count within each calendar year from 2010 to 2024 using the same method in-sample for each year.

Result

The top 10 eigenvalues of the full-period correlation matrix are: 5.1526, 1.3982, 1.1959, 0.6317, 0.5691, 0.5205, 0.5071, 0.4662, 0.4396, 0.4106. Three eigenvalues exceed the MP upper bound of 1.116, yielding n_significant_factors = 3. The largest eigenvalue (5.1526) explains 42.94% of total variance; the three significant factors together explain 64.56% of variance. The remaining nine eigenvalues fall within or below the MP bounds, consistent with random-matrix noise.

The top factor (eigenvalue 5.1526) loads most heavily on MSFT (0.345), GOOGL (0.323), and JPM (0.318). The second factor (eigenvalue 1.3982) loads on AMZN (0.414), XOM (−0.368), and NVDA (0.343), capturing a growth-versus-energy contrast. The third factor (eigenvalue 1.1959) is not detailed in the loadings but is statistically significant.

Time variation: The per-year significant factor count shows the following dynamics:

  • 2010–2013: 1 factor per year (stable, low-dimensional coherence)
  • 2014: 2 factors
  • 2015: 1 factor
  • 2016–2018: 2 factors per year
  • 2019: 1 factor
  • 2020: 2 factors (COVID-19 volatility)
  • 2021: 3 factors (peak complexity)
  • 2022–2024: 2 factors per year

The factor count increased from 1 in the early 2010s to a peak of 3 in 2021, then stabilized at 2 in the most recent years. This time series reveals that the dimensionality of collective equity behavior is not static: it responds to macroeconomic and market-structure shifts.

Interpretation

Spectral phase transition

The eigenvalue spectrum exhibits a clear spectral gap: three eigenvalues lie above the MP upper bound (the "signal" regime), while the remaining nine cluster within the MP bounds (the "noise" regime). This gap is the empirical signature of a phase transition in correlation space. In the gravitational analogy, the MP upper bound acts as a correlation-distance threshold: coherent collective modes (eigenvalues above the bound) correspond to "gravitationally bound" clusters of stocks, while eigenvalues within the MP bounds correspond to uncorrelated or weakly correlated pairs that behave as independent random walkers. The transition is sharp: the third eigenvalue (1.1959) is just above the bound, and the fourth (0.6317) is well below it, with no eigenvalues in the intermediate region. This is consistent with a phase boundary rather than a smooth continuum.

The gravitational metaphor is instructive but limited. In physics, an event horizon is a causal boundary beyond which information cannot escape; here, the MP bound is a statistical boundary beyond which correlation structure is indistinguishable from noise. The analogy holds in the sense that both represent thresholds where qualitative behavior changes (causal connectivity versus statistical coherence), but the mechanisms differ: gravitational collapse is a deterministic dynamical process, while the MP transition is a statistical property of high-dimensional random matrices. The data support the existence of a threshold, not a literal event horizon.

Scaling with market-cap "mass"

The computation question asks whether the number of significant factors scales with market-cap-weighted gravitational mass clustering. The result shows n_significant_factors = 3 for a universe of 12 assets. The top factor loads on large-cap tech and financials (MSFT, GOOGL, JPM), the second on growth-versus-energy (AMZN, NVDA versus XOM), and the third is unspecified but statistically significant. The factor structure reflects sectoral and style clustering, which is economically interpretable as "mass" clustering in the sense that large-cap stocks with similar business models or macroeconomic exposures move together. However, the data do not directly test whether the factor count scales with a quantitative measure of market-cap concentration (e.g., Herfindahl index or entropy of the cap-weight distribution). The result is consistent with the hypothesis that mass clustering drives factor structure, but it does not establish a functional relationship (e.g., log-linear scaling) between cap-weighted mass and factor count. A stronger test would require varying the universe size and cap-weight distribution systematically.

Time variation and regime shifts

The rolling per-year factor count reveals that the dimensionality of equity coherence is time-varying. The increase from 1 factor in 2010–2013 to 3 factors in 2021 coincides with the post-2008 recovery, the rise of mega-cap tech, and the COVID-19 volatility regime. The peak in 2021 (3 factors) aligns with the meme-stock episode, retail trading surge, and sector rotation out of growth into value. The subsequent decline to 2 factors in 2022–2024 suggests a return to a simpler factor structure, possibly driven by the Fed's rate-hiking cycle and the dominance of the tech-versus-everything-else trade. This time variation is the most novel aspect of the result: it shows that the "gravitational" structure of the market is not a fixed property but a dynamic response to macroeconomic and microstructure forces. The factor count is a low-frequency signal (annual resolution), so it captures regime shifts rather than high-frequency turbulence.

What the result does NOT support

The result does not support a forward prediction of factor structure or returns. The eigenvalue spectrum is an in-sample decomposition of historical correlation; it does not imply that the same three factors will persist or that they can be traded profitably. The result does not establish a causal mechanism linking market-cap mass to factor count—it shows an association, not a law. The result does not validate the gravitational analogy as a literal physical model: correlation distance is not Euclidean distance, and market-cap is not gravitational mass in any rigorous sense. The analogy is a heuristic for organizing the data, not a derivation from first principles.

Relation to the Literature

The result sits at the intersection of random matrix theory (RMT) applied to finance and network models of systemic risk.

[P7] and [P8] establish the RMT framework for equity correlation matrices. [P7] confirms that the bulk of the eigenvalue distribution of the Tokyo Stock Exchange correlation matrix follows the MP distribution, with a small number of eigenvalues above the upper bound representing genuine collective modes. Our result replicates this finding for a smaller U.S. universe: three significant eigenvalues above the MP bound, with the remainder consistent with noise. [P8] argues that, except during high-volatility periods, only one factor (the market factor) is significant. Our result shows three significant factors over the full 2010–2024 period, which may reflect either (a) a larger effective volatility regime due to the inclusion of COVID-19 and post-2020 turbulence, or (b) a richer factor structure in the U.S. mega-cap universe compared to the broader market. The time-varying factor count (1 in early years, 3 in 2021) is consistent with [P8]'s claim that factor significance is regime-dependent.

[P9] studies intraday trading volume and finds that the largest eigenvalue dominates collective behavior, with the second eigenvalue robustly above the MP bound at market open. Our result is for daily returns, not intraday volume, but the qualitative pattern is similar: a dominant first eigenvalue (5.1526, explaining 42.94% of variance) and a smaller number of secondary factors. The difference is that we find three significant factors rather than one or two, which may reflect the longer time scale (daily versus intraday) or the different observable (returns versus volume).

[P1] and [P10] apply gravity models to spatial networks (urban agglomerations and credit risk contagion). [P1] improves the traditional gravity model by incorporating functional intensity and mobility, finding hierarchical and imbalanced urban connections. [P10] constructs an entropy-based spatial interaction network for credit risk, showing that spatial proximity and economic development drive cooperation. Our result extends the gravity metaphor to financial correlation space: market-cap acts as "mass," and correlation distance acts as spatial distance. The three-factor structure and the time-varying factor count are analogous to the hierarchical clustering and regime-dependent connectivity in [P1] and [P10]. However, our result is purely statistical (eigenvalue decomposition) rather than network-theoretic (edge weights and centrality), so the connection is conceptual rather than methodological.

[P2] studies the "death of distance" in international trade, finding that the elasticity of trade to distance declined by 11% over 1962–1996, but only for rich-country pairs. This is a tension with our result: if distance were dying in financial markets, we would expect the number of significant factors to decrease (more global coherence, fewer independent modes). Instead, the factor count increased from 1 to 3 over 2010–2021, suggesting that distance (in correlation space) became more salient, not less. One resolution is that [P2] measures physical distance and trade flows, while we measure correlation distance and return comovement—these are distinct notions of distance. Another is that the 2010–2021 period saw increased sectoral and style differentiation (tech versus energy, growth versus value), which increased the effective "distance" between clusters even as global financial integration deepened.

[P3] addresses network reconstruction under partial information, using statistical physics to infer the structure of economic and financial networks. Our result does not reconstruct a network but decomposes a fully observed correlation matrix. The connection is that both approaches use statistical physics concepts (RMT in our case, maximum-entropy ensembles in [P3]) to extract signal from noise in high-dimensional financial data.

[P4], [P5], and [P6] are machine learning papers (volatility indices for direction prediction, deep RL for trading, product evaluation via semantic mining). They are not directly relevant to the eigenvalue spectrum or the gravitational analogy. They appear in the literature list but do not inform the interpretation of the result.

Limitations

  1. Small universe: 12 assets is a minimal sample for RMT. The MP bounds are derived asymptotically (N, T → ∞ with q = N/T fixed), and our q-ratio of 0.003 (12 assets, 3,772 observations) is well within the asymptotic regime for the bounds themselves, but the small N limits the richness of the factor structure. A larger universe (e.g., S&P 500 constituents) would provide a more robust test of the scaling hypothesis.

  2. Market-cap weighting not explicit: The computation question asks whether factor count scales with market-cap-weighted mass clustering, but the PCA is performed on the equal-weighted correlation matrix, not a cap-weighted covariance matrix. The loadings show that large-cap stocks (MSFT, GOOGL, AMZN) dominate the top factors, which is consistent with cap-weighting, but the method does not directly incorporate market-cap as a weight. A cap-weighted PCA or a regression of factor count on a cap-concentration measure would be a stronger test.

  3. In-sample only: The eigenvalue spectrum is computed in-sample. There is no out-of-sample validation of the factor structure (e.g., do the three factors identified in 2010–2020 explain variance in 2021–2024?). The per-year rolling factor count is still in-sample within each year. An out-of-sample test would require splitting each year into train/test or using a true forward-looking design.

  4. Gravitational analogy is heuristic: The analogy between market-cap and gravitational mass, and between correlation distance and spatial distance, is not derived from a physical model. It is a metaphor that organizes the data but does not imply a causal mechanism. The MP bound is a statistical threshold, not a physical event horizon. The result shows that the metaphor is empirically productive (it motivates the eigenvalue decomposition and the scaling hypothesis), but it does not validate the metaphor as a literal model.

  5. Factor interpretation is post-hoc: The loadings on the top two factors (tech/financials versus growth/energy) are economically interpretable, but the interpretation is constructed after seeing the loadings. The third factor is not detailed. A pre-registered hypothesis about which sectors or styles should load on which factors would be a stronger test.

  6. Time resolution: The per-year factor count is annual, which is coarse. Monthly or quarterly resolution would reveal higher-frequency dynamics (e.g., did the factor count spike during the March 2020 crash?). The annual aggregation smooths over intra-year volatility.

  7. Survivorship and selection: The 12 tickers are large-cap survivors over 2010–2024. Delisted or bankrupt firms are excluded, which biases the correlation structure toward stable, positively correlated stocks. A universe that includes failures would show more heterogeneity and possibly more factors.

Strengthening the result would require: (a) a larger, cap-weighted universe; (b) out-of-sample validation of the factor structure; (c) a quantitative test of the scaling hypothesis (factor count versus cap-concentration); (d) higher time resolution (monthly or quarterly rolling windows); (e) a survivorship-free universe; (f) a comparison to a null model that randomizes market-cap while preserving correlation structure, to isolate the effect of mass clustering.

Research evidence, not investment advice.