Oil and equity fear gauges emerge as top warning signals of global market turmoil

When financial panic strikes, it rarely stays contained. The 2008 global financial crisis, the Brexit referendum, the COVID-19 pandemic, and the inflation and geopolitical shocks of the early 2020s all produced the same signature: volatility surging in unison across countries, sectors, and asset classes. Economists call this synchronized phenomenon global common volatility, or COVOL, a measure introduced by Robert Engle and Susana Campos-Martins that captures the shared component of volatility rippling through international markets simultaneously. But while researchers have documented how these synchronized shocks spread, a fundamental question has remained surprisingly open: which sources of uncertainty actually accompany the build-up of systemic volatility? A new study by Jesús Molina-Muñoz and Andrea Valeria Torres-Martinez of Tecnológico de Monterrey, published in Discover Artificial Intelligence, offers one of the most detailed answers yet, and it comes from an unexpected direction: explainable artificial intelligence.

The study tackles two gaps at once. Empirically, although systemic risk and financial uncertainty have each been studied extensively, little evidence existed on which specific types of uncertainty act as predictors of synchronized volatility across global markets. Methodologically, machine learning models have become powerful tools for capturing the nonlinear, interaction-heavy dynamics of financial risk, but they often operate as inscrutable black boxes, a problem regulators and academics alike have flagged as unacceptable when the stakes are financial stability. The authors’ solution was to combine flexible machine learning with not one but three complementary explainable AI techniques: SHAP values, Accumulated Local Effects (ALE) plots, and Local Interpretable Model-agnostic Explanations (LIME). Each tool illuminates a different facet of the model’s behavior, and together they provide a far richer interpretation than any single method could deliver.

The empirical foundation is a daily dataset spanning June 2008 to August 2025, covering 4,487 observations and encompassing crises of strikingly different origins: the global financial crisis, the European sovereign debt crisis, the 2014–2016 oil price collapse, the pandemic, and the geopolitical turbulence of recent years. The dependent variable, COVOL, is constructed through a sophisticated multi-step procedure in which a multivariate volatility model produces jointly standardized residuals across a broad panel of international assets, and a one-factor multiplicative specification then extracts a common latent factor, the global COVOL itself, that drives correlated volatility spikes across all assets. Crucially, unlike the VIX, which reflects expectations about future American equity volatility, COVOL captures realized volatility shocks that have already swept through multiple markets at once.

Against this target, the researchers arrayed a broad battery of uncertainty proxies spanning five channels: commodity and energy markets, equity markets, currency and monetary conditions, inflation expectations, and policy and geopolitical events. The set includes option-implied volatility indices for gold (GVZ), crude oil (OVX), Treasury bonds (MOVE), the broad US equity market (VIX), and technology stocks (VXN), alongside the US Dollar Index, the Bloomberg Dollar Spot Index, five-year break-even and expected inflation measures, the Economic Policy Uncertainty index, the Monetary Policy Uncertainty index, the Geopolitical Risk index, and Bekaert’s Risk Aversion Index. Rather than pre-selecting a handful of favorites, the authors deliberately let all of these compete within a single framework, so their relative associations with systemic volatility could be honestly compared.

On the modeling side, the team trained five tree-based algorithms: XGBoost, LightGBM, CatBoost, Random Forest, and a single Decision Tree as a lower-bound benchmark. Tree ensembles were a deliberate choice for three reasons. First, uncertainty does not transmit to systemic volatility in a smooth, proportional way; its effects are muted in calm periods and intensify sharply once markets enter stress regimes, and tree learners recover such thresholds and interactions endogenously from the data. Second, the uncertainty proxies are strongly correlated with one another, and tree methods remain stable under this redundancy, allowing all proxies to enter simultaneously. Third, and most importantly, the goal was interpretation rather than forecasting, so predictive accuracy had to be paired with a dedicated explainability layer. Hyperparameters were tuned with randomized search and five-fold time-series cross-validation, and all input features were standardized with a robust scaler to limit the influence of extreme values.

The performance results were decisive. XGBoost achieved the lowest errors on every metric, with a mean squared error of 0.004497 and an out-of-sample R-squared of 0.963, meaning the model explained more than 96 percent of the variance in COVOL on held-out data. The predicted series closely tracked the observed dynamics, capturing both general volatility patterns and short-lived spikes. Interestingly, traditional econometric benchmarks, an autoregressive model with exogenous regressors, a GARCH specification, and a linear regression, achieved accuracy comparable to or sometimes better than some of the ensemble models, a finding the authors attribute to COVOL’s smooth, persistent nature as an aggregate measure, which leaves relatively little asset-specific short-term variation for machine learning to exploit. But the benchmarking exercise served a different purpose here: identifying the best-performing model to carry forward into the explainability analysis that forms the study’s core contribution.

That analysis produced a clear and consistent hierarchy. Across SHAP, ALE, and LIME, four predictors dominated: oil market uncertainty (OVX), broad equity market volatility (VIX), technology-sector volatility (VXN), and the global Risk Aversion Index (RAI). Oil market uncertainty ranked first, a result the authors connect to the central macro-financial role of energy, since oil price uncertainty feeds into inflation expectations, production costs, risk premia, and cross-market volatility, particularly amid the overlapping geopolitical and climate-related shocks of recent years. The SHAP dependence plots revealed striking nonlinearities: the contribution of OVX stays small at low levels and turns increasingly positive as oil volatility rises, an effect amplified when the dollar is strong, consistent with the reserve-currency role of the US dollar in global commodity markets. The VIX contribution strengthens when geopolitical risk is elevated, while the risk aversion effect is strongest at low-to-moderate levels and flattens at extremes, especially when gold market volatility is high, hinting at interactions between investor sentiment and safe-haven demand.

The ALE plots confirmed that these effects are positive and stable across the distributions: oil uncertainty shows a monotonic, steepening relationship with COVOL, while even moderate increases in equity and technology-sector volatility measurably raise predicted systemic volatility. LIME then added the event-level texture that global methods miss. For the September 2008 collapse, equity market volatility dominated the prediction, followed by risk aversion and oil uncertainty, mirroring the crisis’s origin in the financial system itself. During the 2015 oil downturn and China-related turbulence, technology-sector and oil uncertainty took the lead. And for the March 2020 COVID-19 shock, expected inflation emerged as the largest contributor, followed by risk aversion and oil market volatility, reflecting the pandemic’s disruption of global production, energy demand, and inflation expectations, all coinciding with an oil price war that drove crude volatility to extreme levels.

The practical implications are immediate. Because a small set of observable, market-based indicators, led by OVX, VIX, and VXN, captures most of the systematic variation in global common volatility, monitoring these indices can serve as an early-warning signal for the build-up of synchronized volatility across markets. For risk managers, this argues for centering hedging strategies and stress tests on energy-market and equity-market uncertainty. For investors, tracking these channels can help anticipate the periods when diversification fails, since synchronized volatility erodes its benefits precisely when they are needed most. For regulators and macroprudential authorities, the prominence of oil-market uncertainty suggests that disturbances originating outside the core financial system deserve a place in systemic-risk surveillance frameworks alongside traditional financial-stability metrics. The authors are careful to note the caveats: the findings represent interpretable associations rather than causal effects, several proxies are US-centric while COVOL is global, and the 2008–2025 sample may not generalize to all market regimes. Yet the convergence of three independent explainability methods on the same handful of predictors lends the conclusion real credibility, and it demonstrates something larger: that explainable AI can open the black box of financial machine learning and turn it into a transparent window on the anatomy of global market fear.

Subject of Research: Explainable machine learning analysis of uncertainty predictors of global common volatility in financial markets

Article Title: Uncertainty and global common volatility: an explainable artificial intelligence approach

Article References: Molina-Muñoz, J., & Torres-Martinez, A. V. (2026). Uncertainty and global common volatility: an explainable artificial intelligence approach. Discover Artificial Intelligence, 6(1), Article 1410. https://doi.org/10.1007/s44163-026-02421-7

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02421-7

Keywords: global common volatility, COVOL, explainable AI, SHAP values, XGBoost, systemic risk, oil market volatility, VIX, risk aversion, machine learning, financial stability, uncertainty