Volatility connectedness across energy, metals, agricultural and financial markets has become central to portfolio management and systemic-risk surveillance. The financialisation of commodities and their common exposure to global macroeconomic shocks have tightened cross-asset volatility linkages over the past two decades, making the intensity and direction of transmission a first-order concern for investors and regulators alike. Yet the dominant Diebold–Yılmaz framework remains mean-based and therefore silent on the tails of the conditional distribution—precisely the region in which crises materialise. Three gaps motivate this study. First, existing quantile-connectedness applications in the commodity–financial domain seldom pair quantile estimation with penalised, data-driven model selection—which filters spurious spillovers into a sparse, interpretable network and stabilises estimation in the thin tails of rolling windows—and typically resolve the trade-off between regularisation and sparse tail information in an ad hoc manner. Second, the literature routinely equates a market's variance-decomposition share with its systemic importance, conflating linear, symmetric centrality with the direction in which information actually flows. Third, the link between connectedness and macroeconomic uncertainty is almost always estimated at the mean, implicitly assuming that uncertainty affects connectedness uniformly across market states.
We study daily realised volatilities for eleven futures markets—WTI crude oil, natural gas, gold, silver, copper, wheat, corn, soybeans, the S&P 500, the Nikkei 225 and the 10-year U.S. Treasury note—constructed from 5-minute intraday data over April 2010 to June 2023, a balanced panel of 3,061 trading days. Connectedness is estimated with a kernel-weighted, cross-validated LASSO-penalised quantile VAR at five quantiles (0.05 to 0.95), adapting the local-conditioning logic of Ando, Greenwood-Nimmo and Shin (2022), from which we compute the generalised forecast-error variance decomposition at each quantile. We complement this with a three-layer causal network—variance shares, LASSO-implied Granger causality, and Kraskov–Stögbauer–Grassberger transfer entropy—that separates statistical centrality from directed information flow, and we relate rolling connectedness to four macroeconomic uncertainty indices (VIX, OVX, EPU and GPR) through quantile regression. Bai–Perron tests date shifts in the connectedness regime.
Three results emerge. First, connectedness is strongly tail-dependent and asymmetric: total connectedness remains in the mid-fifty-percent range across the central quantiles but rises sharply to roughly 75% in the upper tail, and the quantile Granger-causality network becomes complete at the ninety-fifth percentile, so that in stressed states essentially every market transmits shocks to every other. Integration thus intensifies exactly when diversification is most needed, and a stable structural backbone of causal linkages persists across all quantiles even as the identity of the dominant transmitters shifts with the market state. Second, the mean-based picture of which markets drive the system is misleading. Under the variance decomposition the S&P 500 appears as a central hub, yet transfer entropy shows that it emits no significant outgoing information while WTI crude oil is the dominant source; variance-share centrality and directed information flow point in opposite directions—a centrality–causality divergence in which a variance-decomposition hub (equity) is not the true origin of directed shocks, while the dominant information source (crude oil) is masked in the variance shares. Third, oil-specific uncertainty (OVX) dominates aggregate equity-market uncertainty (VIX) as a driver of connectedness, subsuming the latter's explanatory content in a manner our diagnostics confirm is economic rather than a collinearity artifact (condition number 2.70), with its influence amplified in the upper tail. A structural break in connectedness at January 2019—well before the onset of COVID-19—indicates that the integration regime had already shifted prior to the pandemic.
The paper contributes a penalised quantile-connectedness estimator that genuinely conditions on tail states; evidence that variance-decomposition hubs need not be the true sources of systemic shocks, cautioning against a widespread inferential shortcut; and the identification of oil-specific uncertainty as the dominant, tail-amplified macroeconomic driver of connectedness. These findings carry direct implications for hedging, margining and systemic-risk monitoring in an increasingly integrated commodity–financial system.

JEL classification: C32; C58; G15; Q02; Q43.