When volatility cannot separate economic shocks

The identification problem

An oil-price increase can accompany a supply disruption, stronger global demand or rising demand specific to oil. Separating these explanations requires more than observing prices. Economists use structural vector autoregressions, or SVARs, to model relationships among variables and estimate how they respond over time to underlying shocks.

One identification strategy exploits changes in volatility. If different shocks become more or less variable by different proportions between two periods, those changes can help distinguish their effects. The model assumes that the underlying economic relationships remain stable while shock variances change.

But this strategy has a precise failure point: two or more shocks may share the same proportional variance change. In that case, volatility information alone cannot uniquely separate those shocks, even if economic volatility clearly changes overall.

In their preprint, first submitted in March 2024 and revised in November 2026, Bacchiocchi and colleagues characterize what researchers can still learn in this situation.

From one answer to an admissible set

The mathematics involves an eigenvalue decomposition constructed from the covariance matrices of model errors before and after the volatility break. Its eigenvalues represent the relative changes in structural shock variances.

Distinct eigenvalues can identify individual shocks, provided their ordering and signs are resolved. Repeated eigenvalues leave multiple admissible ways to separate the associated shocks. A shock corresponding to a distinct eigenvalue can remain individually identified even when other shocks are not.

The authors derive the resulting identified sets for impulse responses: the responses consistent with the model, the observable covariance information and the maintained restrictions. This distinguishes uncertainty caused by limited data from ambiguity that remains because the model does not uniquely determine the answer.

Additional assumptions can narrow these sets. Zero restrictions rule out specified relationships, while sign restrictions constrain the direction of responses. Under the paper’s stated rank and restriction conditions, suitable zero restrictions can restore unique identification. Sign restrictions can instead leave a narrower range of admissible responses.

Making Bayesian uncertainty more explicit

The authors also adapt an existing robust Bayesian framework. In a set-identified model, a conventional Bayesian analysis requires prior beliefs about distinctions that the data cannot resolve. Those beliefs can continue to influence conclusions even as more observations become available.

Their procedure retains a chosen posterior distribution for the reduced-form parameters—the statistical relationships estimated from the data—while allowing conditional prior beliefs over admissible structural decompositions to vary. It reports bounds on posterior means and robust credible regions that satisfy a posterior credibility requirement across this class.

This is not protection against every modeling assumption or every prior choice. The reduced-form posterior remains fixed within the robustness analysis.

The paper also develops Bayesian comparisons of models with equal or distinct variance shifts, including joint equality among several shifts. Model averaging can incorporate uncertainty between identification scenarios. However, the main algorithm’s reported bounds are conditional on the selected scenario rather than automatically including that model uncertainty.

A historical oil-market illustration

The application uses monthly observations from January 1973 through December 2007 on global crude oil production growth, global real economic activity and the real oil price. The model includes 12 monthly lags and assumes a volatility break in October 1987.

The authors find strong evidence of changing shock variances, but only modest support for treating two relative variance shifts as equal. With equal prior probabilities across five candidate models, the equal-pair model receives approximately 60.45% posterior probability, compared with 39.55% for the all-distinct model. These probabilities do not establish equality as a fact.

Imposing that equality leaves one shock, interpreted as oil-specific demand, point-identified. The authors use static and dynamic sign restrictions to set-identify oil supply and aggregate demand shocks. These assumptions include requiring an oil supply disruption to raise the real oil price on impact and for twelve months afterward. The resulting price response peaks after about six months, broadly matching the benchmark pattern discussed in the paper.

What the evidence does not establish

Simulation experiments show that accurate selection of the identification model does not guarantee nominal coverage. Some reported 90% robust credible regions cover the true identified set in only about 87% of replications even at the largest sample size. These coverage calculations are conditional on correct model selection and valid inference.

The authors explicitly do not establish frequentist coverage guarantees for their procedure. They interpret the empirical regions in Bayesian terms. Their implementation also assumes two volatility regimes and a known break date; unknown break timing and broader regime structures remain topics for further work.