💰 quantitative finance

A Spectral Generalisation of the Variance Ratio: Eigenstructure of Long-Horizon Portfolio Covariance and a Multi-Memory Factor Model of U.S. Equity Returns

This paper introduces a spectral generalization of the variance ratio to identify a robust five-factor model that decomposes long-horizon equity returns into distinct return and volatility memory channels, revealing a late-1980s regime shift in volatility persistence and demonstrating that the cross-sectional drivers of return momentum are economically distinct from those governing volatility persistence.

Original authors: Anders G Frøseth

Published 2026-07-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Anders G Frøseth

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the stock market not as a single, chaotic ocean, but as a complex orchestra playing a symphony. For decades, economists have tried to understand the music by listening to just one instrument (a single stock) or by measuring the total volume of the whole orchestra. This paper proposes a new way to listen: by identifying the specific "notes" (patterns) the orchestra plays and understanding how those notes change over time.

Here is a breakdown of the paper's findings using simple analogies:

1. The Problem: Listening to the Wrong Way

Imagine you are trying to understand how a crowd of people moves.

Old Method: You pick one person and watch them for a year. You might think, "Oh, they walk in a straight line," or "They stop and start." But this ignores how the whole crowd moves together.
The Paper's Method: Instead of watching one person, the authors look at the entire crowd's movement patterns simultaneously. They use a mathematical tool called "eigenstructure" to find the hidden "dance moves" that the whole group does together. They call these the "principal directions" or the orchestra's "main melodies."

2. The Discovery: The Market Has Two "Channels"

The authors found that the market's memory works through two distinct channels, like two different radio stations broadcasting at the same time:

The "Price" Channel (Linear): This is about the actual stock prices going up and down.
The "Volatility" Channel: This is about how wild or calm the price swings are (the fear and excitement).

The paper shows that these two channels have different "memories." Just because a price trend lasts a long time doesn't mean the fear (volatility) lasts the same amount of time. They are like two different instruments playing different rhythms.

3. The "Five-Factor" Orchestra

The authors discovered that the entire U.S. stock market can be explained by just five main "musicians" (factors) playing together. These five factors explain almost everything about how stocks behave over long periods (years, not just days):

The Persistent Player: A musician who keeps playing a steady, long-lasting tune (momentum).
The Antipersistent Player: A musician who plays a tune that constantly corrects itself (mean reversion—when things go up too much, they tend to come back down).
The "Multifractal" Conductor: This is the most complex musician. They create a "cascade" of noise that gets louder and more complex over time. This explains why market crashes or booms can feel like they have a life of their own that stretches over years.
The Volatility-Only Player: A musician who only plays the "fear" notes, not the price notes.
The "Burst" Player: A musician who plays short, intense bursts of noise that fade away quickly.

The Big Finding: This exact same "five-musician" band explains the market whether you look at:

Different industries (like Tech vs. Healthcare).
Different types of companies (Small vs. Large, Cheap vs. Expensive).
Different time periods (before 1998 vs. after 1998).
Even the European market (a "replica" of the US findings).

4. The "Great Shift" of the Late 1980s

One of the most exciting findings is a "regime transition." The authors found that the market didn't just slowly change; it underwent a structural shift in the late 1980s.

Before the Shift: The "fear" (volatility) in the market was like a short-term storm. It would rage for a bit and then calm down relatively quickly (about 2 years).
After the Shift: The "fear" became a deep, slow-moving ocean current. The longest-lasting waves of volatility stretched out to 4 years.

Think of it like a river changing its flow. Before the late 80s, the water moved in quick, shallow ripples. After the late 80s, the river developed deep, slow-moving currents that took years to pass. The authors pinpointed this change using a "rolling window" technique, showing that the shift happened gradually but firmly between 1985 and 1995, not at the arbitrary 1998 date often used by other researchers.

5. The "Beta Inversion" Surprise

Finally, the authors tested a common assumption: that the same companies driving price trends (like momentum) are the same ones driving volatility trends.

The Hypothesis: If a company is a "price leader," it should also be a "fear leader."
The Reality: The authors found the opposite. The companies that drive long-term price trends are completely different from the companies that drive long-term volatility trends. It's as if the "price orchestra" and the "fear orchestra" are playing in different rooms with different musicians.

Summary

This paper argues that the stock market is not a random mess. It is a highly structured system with five core patterns that repeat across different countries and time periods. However, the "rules of the game" for how long market fear lasts changed permanently in the late 1980s, shifting from short-term spikes to long-term, multi-year waves. Most importantly, the forces that move prices are distinct from the forces that move volatility; they are two separate stories happening at the same time.

Technical Summary: A Spectral Generalisation of the Variance Ratio

Problem Statement
The aggregation of asset returns across time horizons ( $H$ ) fundamentally alters the second-moment structure of the cross-section. While the univariate variance ratio test of Lo and MacKinlay (1988) assesses whether $H$ -period returns aggregate cleanly for a single series, it discards the cross-sectional information contained in the eigenstructure of the covariance matrix. Existing multivariate extensions, such as the matrix-valued variance ratio of Hong et al. (2017), preserve more structure but reduce it to scalar functionals (trace, determinant) for inference, failing to index long-horizon dynamics by eigenmode. Furthermore, equity returns exhibit documented departures from independent, identically distributed (i.i.d.) Gaussian behavior, including power-law tails, long-range dependence in absolute returns, and multi-scale self-similarity in volatility consistent with multifractal cascades. The paper addresses the need for a framework that decomposes these joint deviations into structured, eigenstructure-indexed components across both linear (return) and volatility channels.

Methodology
The authors propose a multivariate, eigenmode-indexed generalisation of the variance ratio test. The framework relies on two cross-sectional statistics defined for a symmetric matrix-valued time series $\Sigma_t$ and its $H$ -period aggregate $\Sigma^{(H)}_t$ :

Per-Eigenmode Variance Ratio ( $\kappa$ ):
$\kappa_i(H) := \frac{\lambda_i(\Sigma^{(H)}_t)}{H \cdot \lambda_i(\Sigma_t)}$
This measures the autocorrelation of the $i$ -th principal direction across aggregation horizons. Under i.i.d. returns, $\kappa_i(H) = 1$ for all $i, H$ .
Eigenvector-Overlap Matrix ( $O$ ):
$O_{ij}(H) := |\langle v_i(\Sigma^{(H)}_t), v_j(\Sigma_t) \rangle|^2$
This measures the geometric rotation of the eigenbasis as the horizon extends. Under i.i.d. returns, $O_{ij}(H) = \delta_{ij}$ .

The paper derives closed-form predictions for these statistics under parametric models. For the linear channel, it establishes that an AR( $p$ ) process decomposes $\kappa$ into a weighted sum of AR(1) variance ratios indexed by characteristic roots. For the eigenvector overlap, it derives a first-order perturbation result showing that rotation depends on the eigenvalue gap and a horizon-dependent saturation factor.

To model the empirical deviations, the authors construct a five-factor multi-memory model applied jointly to the linear channel (covariance of log-returns) and the volatility channel (covariance of squared returns). The factors are:

$F_P$ (Persistent): Fractional Brownian motion (fBm) with Hurst exponent $H_P > 0.5$ .
$F_A$ (Antipersistent): fBm with $H_A < 0.5$ .
$F_M$ (Multifractal): An ARFIMA(1, $d_M$ , 0) process in the linear channel coupled with a Markov-Switching Multifractal (MSM) cascade in the volatility channel.
$F_V$ (Volatility Companion): A pure-volatility factor sharing the MSM cascade of $F_M$ but contributing no linear structure.
$F_{Vt}$ (Transitory Volatility): An ARFIMA process governing the volatility-of-volatility, capturing transient clustering.

The model is estimated via joint least squares (LS) on the empirical $(\kappa^{lin}, \kappa^{vol})$ matrices using a moving-block bootstrap (1,000 replicates) and L-BFGS-B optimization.

Key Contributions and Results
The framework is applied to four U.S. equity panels (Fama-French 49-industry full sample and pre/post-1998 splits; Fama-French 100 size $\times$ book-to-market sort) and a non-U.S. replication (Europe 25). The analysis yields three primary findings:

Universal Five-Factor Decomposition: The same five-factor model recovers seven stylised facts of long-horizon equity dynamics simultaneously across all panels:
- Market-mode variance-ratio anomaly: A rise-then-fall pattern in the linear channel at the market mode.
- Factor momentum: Monotone-rising variance ratios in sub-leading eigenmodes (ranks 2–3).
- Long-horizon mean reversion: Steeply decreasing variance ratios in deep eigenmodes (ranks 24+).
- Short-range volatility clustering: Sharp initial rise in volatility variance ratios.
- Multi-scale volatility long memory: A massive accumulation of variance ratio in the market-mode volatility channel (reaching $\approx 48$ at 5 years), consistent with a multifractal cascade.
- Transitory volatility: Episodic volatility bursts in deep eigenmodes.
- Cross-sectional concentration: The volatility spectrum concentrates onto a single dominant mode as the horizon lengthens, whereas the linear spectrum broadens.
  The fractional Brownian exponents ( $H_P \approx 0.52\text{--}0.57$ , $H_A \approx 0.17\text{--}0.27$ ) remain stable across industry sorts, size-value sorts, and temporal sub-samples, indicating a universal eigenstructure-level specification.
Regime Transition in Volatility Memory: The paper identifies a structural break in U.S. equity volatility memory occurring in the late 1980s, rather than at the static 1998 split-half boundary.
- A rolling-window analysis (28-year windows, 2-year strides) localizes the transition. The weight of the MSM cascade on the market-mode volatility channel ( $w^{vol}_{MSM}$ ) rises monotonically from a median of 0.30 (centered 1985) to 0.93 (centered 1991) and saturates at 1.00 from 1995 onward.
- The 90% confidence bands for the pre-transition (1983 center) and post-transition (1999 center) windows are strictly non-overlapping.
- The "cross-over time scale" of the cascade (duration of the lowest-frequency component, $1/\gamma_1$ ) lengthens from approximately 2.2 years in the pre-1998 period to 4.0 years in the post-1998 period. The static "first half" panel (1969–1997) exhibits weak identification because it mixes these two regimes.
Cross-Channel $\beta$ -Inversion and Distinct Loadings: The authors test the hypothesis that the multifractal factor's linear and volatility signatures are governed by a single shared cross-asset loading vector ( $\beta_M$ ).
- A $\beta$ -inversion test recovers the squared loadings $|\beta_M[a]|^2$ independently from the linear and volatility channels.
- The results reject the shared-loading hypothesis: the Pearson correlation between the recovered loadings is significantly negative on three of the four panels (e.g., $-0.771$ on the full sample).
- This implies that the cross-sectional characteristics driving return-momentum patterns are economically distinct from those driving volatility-persistence patterns; a single loading cannot explain both channels.

Significance
The paper claims to provide the first eigenstructure-indexed decomposition of the variance ratio, distinguishing between temporal autocorrelation along stable principal directions and cross-sectional rotation of those directions. By separating the linear and volatility channels, the framework reveals that the "market mode" is not a monolithic entity but a composite of distinct memory mechanisms: a rise-then-fall linear profile driven by antipersistent and ARFIMA components, and a pure multi-scale volatility cascade.

The identification of a regime transition in the late 1980s suggests that the long-memory structure of volatility is not static but has evolved, extending the lowest-frequency cascade component to multi-year horizons in the modern era. Finally, the rejection of the shared-loading hypothesis challenges models that assume a single set of risk factors drives both return and volatility dynamics, suggesting instead that industry and size/book-to-market characteristics predict these channels independently. The framework offers a robust, parsimonious (five-factor) description of long-horizon equity dynamics that holds across different asset universes and time periods.