When Volatility Travels: Mapping Global Spillovers with Rough Multivariate Models

TL;DR for operators

Volatility does not politely stay where it starts. A shock in one index can show up elsewhere, not just as a same-day correlation but as a delayed, asymmetric pattern in future volatility. The paper behind this article proposes a multivariate rough-volatility model that tries to capture that behaviour directly: each market has its own roughness and mean reversion, while pairs of markets have parameters governing contemporaneous dependence and time asymmetry.¹

The practical message is not “throw away your VAR and worship fractional Brownian motion.” Please do not start a new cult before lunch. The message is narrower and more useful: if volatility is rough, slowly mean-reverting, and cross-market covariance decays asymmetrically, then static correlation matrices are missing part of the risk transmission mechanism.

The authors estimate the model on 22 realized-volatility index series from the Oxford-Man Realized Library, covering January 2000 to June 2022, using daily volatility constructed from 5-minute price increments. They report Hurst exponents below $1/2$, strong model fit to auto- and cross-covariance patterns, evidence consistent with slow mean reversion, and a spillover analysis in which total normalized spillovers amount to 86% of the forecast-error variance decomposition of log-realized volatility.

For a risk desk, the useful translation is this:

Volatility travels, but correlation only tells you where it has been

A risk manager looking at a global portfolio usually starts with correlations. That is understandable. Correlations are easy to compute, easy to colour red, and easy to put in a committee deck where everyone pretends the heatmap is an explanation.

But volatility has a more inconvenient habit. It persists, reverts slowly, clusters, and sometimes arrives elsewhere with a delay. The problem is not only that two markets are linked. The problem is how their volatility relationship changes as the lag changes.

That is the difference between a static co-movement map and a transmission mechanism. Static correlation says two markets moved together. A cross-covariance function asks a sharper question: when volatility rises in market $j$, how does market $i$ behave at different future and past lags?

The paper’s key move is to take rough volatility, normally treated as a mostly univariate story, and make it genuinely multivariate. Not “one rough process per index plus a correlation matrix,” which would be the econometric equivalent of putting twenty-two soloists in the same room and calling it an orchestra. The authors instead model log-volatility as a multivariate fractional Ornstein-Uhlenbeck process, where roughness, mean reversion, contemporaneous dependence, and time asymmetry all participate in the same system.

The mechanism: volatility is rough, mean-reverting, and not politely symmetric

The model is built on the fractional Ornstein-Uhlenbeck idea. For each component $i$, log-volatility follows a mean-reverting process driven by fractional Brownian motion:

Here $Y_t^i$ is the log-volatility of market $i$, $\mu_i$ is its long-run mean, $\alpha_i$ is the speed of mean reversion, $\nu_i$ is a diffusion scale, and $H_i$ is the Hurst exponent. When $H_i < 1/2$, the path is “rough”: locally jagged in a way that standard Brownian volatility models do not capture well.

That is the univariate half of the story. The multivariate half is where the paper becomes more interesting.

The driving noise is a multivariate fractional Brownian motion. It includes:

The last parameter is the one many readers would miss. In a plain correlation matrix, the relationship between two assets is symmetric. In this model, the cross-covariance between $i$ and $j$ at lag $k$ does not have to mirror the cross-covariance between $j$ and $i$ at the same lag. That asymmetry depends not only on $\eta_{i,j}$, but also on the two markets’ mean-reversion speeds.

This is the central correction to the easy misconception. Multivariate rough volatility is not just rough volatility with a fancier covariance matrix. It is a model in which volatility transmission has shape: local roughness, persistence, mean reversion, contemporaneous dependence, and asymmetric lag decay.

The paper also shows that when mean reversion is slow, the multivariate fractional OU process locally behaves like a multivariate fractional Brownian motion. In that regime, the cross-covariance becomes approximately linear in a power of the lag, with the power governed by the sum $H_i + H_j$. That small technical observation later becomes a practical diagnostic: if the empirical covariance curves line up against this transformed lag, the model’s mechanism is not merely decorative.

The estimator is built to match covariance shapes, not price paths

The authors estimate the model with a two-step Generalized Method of Moments procedure. The objective is simple in concept: choose the parameters that make the model-implied auto- and cross-covariances resemble the empirical auto- and cross-covariances.

The implementation is less cute. For a system with $N$ components, the model has $N(N+2)$ parameters: three marginal parameters per component, plus pairwise $\rho$ and $\eta$ parameters. For the 22-index empirical system, that becomes 528 parameters and 3,641 moment conditions.

The lag set is deliberately small and mixed:

Short lags are informative about roughness and local behaviour. Longer lags help stabilise estimation and capture broader covariance decay. The authors use a first GMM step with the identity weighting matrix, then a second step using the inverse of the diagonal of a Newey-West estimate. They explicitly avoid a fully optimal weighting matrix because persistent volatility and high dimensionality make that object difficult to estimate reliably, and because optimal GMM can reduce standard errors at the cost of more bias in mean-reversion parameters.

That trade-off matters. This is not a black-box volatility learner. It is an estimator engineered around a specific object: the covariance structure of log-volatility.

The simulation section is a stress test of the machinery

The Monte Carlo study is not the main empirical claim. It is the estimator’s quality-control department. The authors simulate multivariate fractional OU processes under different parameter settings, then check whether the GMM procedure recovers the known parameters.

The evidence is best read by purpose:

The important result is not perfection. The important result is that the estimator behaves sensibly where the empirical application needs it most, while showing exactly where operators should be careful: slow mean reversion, high dimension, and long-memory-adjacent regimes.

That is useful. A model that tells you where it limps is usually more valuable than one that arrives wearing a cape.

The empirical system is broad enough to expose global structure

The main empirical analysis uses realized volatility from the Oxford-Man Realized Library. The authors start with 31 global stock-index volatility series from January 3, 2000 to June 28, 2022. They retain 22 series that cover the full sample sufficiently well, remove zero-volatility observations, annualise and log-transform volatility, and fill sparse missing values with an AR(5) interpolation. The final system has an average length of 5,616 observations per series.

The estimated univariate parameters tell a familiar rough-volatility story. The Hurst exponents are all below $1/2$, with most liquid indices averaging around 0.17 and less liquid indices somewhat higher. That supports rough behaviour, although the authors note that their estimates are slightly larger than some prior rough-volatility estimates because their GMM procedure uses covariance decay over lags up to 50 days rather than only very short-lag behaviour.

Mean reversion is mostly estimated around 0.6, with higher values for AORD, RUT, BVSP, and especially KSE. This combination gives a useful picture: volatility is locally rough, broadly persistent, but still mean-reverting.

The paper is careful on long memory. Because the estimated Hurst exponents imply no long memory at the univariate level and short-range interdependence at the multivariate level, the fitted model itself is not a long-memory volatility system. But the authors also do not claim to rule out long-range dependence empirically, because their GMM moments use lags only up to 50 days. That boundary is not a footnote. It defines what the model has actually been asked to explain.

The cross-covariance fit is where the model earns its keep

The empirical payoff comes from the cross-covariance curves.

The authors highlight two pairs: FTSE-SPX and FTSE-SSEC. These are chosen as contrasting cases based on the magnitude of $\eta_{i,j}$. FTSE-SPX has $\eta_{i,j}=-0.02$, with $\alpha_i=0.62$ and $\alpha_j=0.61$, suggesting near symmetry. FTSE-SSEC has $\eta_{i,j}=-0.11$, with $\alpha_i=0.62$ and $\alpha_j=0.64$, suggesting more visible asymmetry. The fitted theoretical curves track both cases well.

This is where the mechanism-first reading matters. The model is not merely saying that FTSE and SPX are related, or that FTSE and SSEC are less related. It is explaining why the covariance curve around lag zero can be nearly symmetric for one pair and visibly asymmetric for another. The answer is not a single “correlation” number. It is the interaction of $\eta$, $\alpha_i$, $\alpha_j$, and the Hurst exponents.

The slow-mean-reversion evidence reinforces the same point. When the authors plot empirical cross-covariances against a power transformation of the lag governed by $H_i + H_j$, the relationship is approximately linear up to 50 lags for both the symmetric FTSE-SPX case and the asymmetric FTSE-SSEC case. They then estimate using the small-$\alpha$ asymptotic cross-covariance approximation and again obtain a strong fit. Parameter magnitudes shift, but the evaluated cross-covariance functions remain similar.

Interpretation: the near-nonstationary, slow-reversion story is not just attached to the model after the fact. It is visible in the shape of the empirical covariance decay.

Intraday volatility is promising, but it is not the main event

The paper also applies the model to intraday spot-volatility dynamics for SPX, DJI, and IXIC. The authors estimate spot volatility from 1-minute prices using a Gaussian-kernel estimator, deseasonalise the intraday U-shape with a flexible Fourier form, retain 15-minute intervals, annualise the resulting series, and fit the same model class.

The reported intraday estimates are strikingly rough: $H=0.07$ for SPX, $H=0.07$ for DJI, and $H=0.06$ for IXIC. Pairwise $\rho$ estimates are very high, at 0.98 for SPX-DJI, 0.96 for SPX-IXIC, and 0.93 for DJI-IXIC, while $\eta$ is estimated as zero across the three pairs. The authors interpret the smaller Hurst exponents as consistent with prior evidence that spot volatility is rougher than realized volatility, partly because realized volatility is smoothed by integration.

This section is best read as an exploratory extension. It shows that the mfOU model can fit a small spot-volatility system and that high-frequency estimation does not obviously break the framework. It does not yet show a global intraday volatility spillover network. The authors themselves point to intraday seasonality, spot-volatility estimation, and the lack of aligned trading hours across markets as practical obstacles.

There is also a small textual wrinkle: the paper’s dataset overview and the intraday subsection do not state the same starting year for the SPX/DJI/IXIC 1-minute sample. That does not undermine the mechanism, but it is another reason to treat the intraday evidence as illustrative rather than as the article’s main empirical payload.

Spillovers convert covariance asymmetry into a network view

The spillover analysis is the business-facing part of the paper, but it should not be read too casually.

The authors draw on the Diebold-Yilmaz framework, where spillovers are based on forecast-error variance decomposition. In plain English: for a future volatility forecast error in market $i$, how much of that error is attributable to innovations in market $j$?

To compute this inside their framework, the authors move to a causal and discretised version of the mfOU process. This matters. The spillover calculation is not simply “take the unconstrained model and print network centrality.” The causal version restricts the process to depend on past innovations and imposes an additional constraint linking $\eta_{i,j}$ to $H_i$, $H_j$, and $\rho_{i,j}$. Under the adapted discretisation, the resulting normalized variance shares become independent of forecast horizon, history, and discretisation step.

That independence is elegant, but also a boundary. It makes the spillover table clean. It also means the reported spillovers are structural full-sample quantities under the causal formulation, not a rolling crisis-by-crisis dashboard.

The headline result is large: total spillovers amount to 86% of the normalized forecast-error variance decomposition of log-realized volatilities over January 2000 to June 2022.

The directional findings are more useful than the headline:

For operators, the distinction between “received” and “transmitted” is valuable. A market can be risky because it absorbs global shocks. Another can be risky because it broadcasts them. Those are different problems. The first affects exposure management. The second affects scenario design.

The business value is diagnostic leverage, not a magic forecast

The paper’s direct contribution is methodological and empirical. It proposes a continuous-time multivariate rough-volatility model, estimates it jointly, validates the estimator, fits it to a global realized-volatility system, and derives spillover indices.

The business inference is more practical: volatility systems should be monitored as asymmetric networks, not just as clusters of correlated assets.

That matters in at least four workflows.

First, portfolio stress testing. A static correlation matrix can tell a portfolio manager that assets are linked. A volatility spillover model can help identify which markets are more likely to act as volatility transmitters. Scenario design becomes less arbitrary when shocks originate from empirically central transmitters rather than from whoever shouted loudest in the morning meeting.

Second, hedging. If a market is primarily a receiver, hedging its own volatility in isolation may miss the upstream source. If it is a transmitter, hedging decisions may need to account for broader portfolio knock-on effects.

Third, volatility forecasting. The authors do not run the rolling forecasting exercise, but the scale of the estimated spillovers suggests that univariate forecasting may leave useful information on the table. A market’s own volatility history is not the whole state vector.

Fourth, model governance. The framework separates several phenomena that are often collapsed into “correlation went up”: roughness, mean reversion, contemporaneous dependence, and time asymmetry. That separation gives risk teams more diagnostic vocabulary. Sometimes the issue is not that markets are more correlated. Sometimes the issue is that volatility shocks are decaying differently across lags.

The limits are specific, and they matter

This paper is useful because it does not pretend the hard parts are gone.

The estimator relies on local identification and suitable starting values. The authors use initial values from univariate estimators and prior multivariate correlation-parameter estimators, then refine with joint GMM. That is reasonable, but it means the method is not a “press button, receive truth” pipeline.

The empirical model uses selected lags up to 50 trading days. That is enough to study short- and medium-lag covariance decay, but not enough to settle all questions about long-range dependence. The authors are explicit that roughness and long memory can coexist, and their estimates should not be read as a universal rejection of long memory in volatility.

The Oxford-Man realized-volatility data require filtering and interpolation. The final 22-index system is broad and global, but it excludes indices with prolonged missing periods. Sparse missing values are filled with AR(5) interpolation. That is sensible housekeeping, not divine revelation.

The spillover analysis is full-sample. It says something about the fitted system over 2000-2022, not about whether SPX, DJI, and FTSE remain the same transmitters in every crisis, policy regime, liquidity shock, or post-pandemic structural break.

The intraday analysis is deliberately small. It supports the model’s potential at higher frequency, but it does not solve global intraday alignment, market microstructure noise, or large-scale spot-volatility spillover estimation.

Finally, the paper proposes dynamic forecasting and rolling-window spillover analysis as future work. That is the correct next step for business deployment. A static full-sample network is useful for understanding the mechanism. A rolling network is what operators would eventually want on a screen.

What to take from this without overbuying it

The best reading of this paper is not that it replaces existing volatility models. The best reading is that it sharpens the object those models should try to capture.

Volatility is rough within markets. It mean-reverts, often slowly. Across markets, it can move with contemporaneous dependence and asymmetric lag structure. Once those features are modelled together, spillovers become more than a storytelling device. They become estimable consequences of a covariance mechanism.

That is the quiet business value. Not a louder forecast. Not a shinier dashboard. A better map of how volatility travels.

And in markets, a better map is rarely enough to prevent the storm. But it does help you stop calling every storm “unexpected,” which is at least a start.

Cognaptus: Automate the Present, Incubate the Future.