Dominoes or weather? Separating self-excitation from latent-regime clustering, in three forms

Abstract

Clustered crisis events admit two very different explanations. In the dominoes reading each event raises the hazard of the next: self-excitation, endogenous contagion. In the weather reading a slowly varying latent regime makes some periods broadly dangerous: exogenous modulation with no event-to-event causation. The two produce nearly identical clustering, and the policy readings they invite are opposite. This paper races a Hawkes self-excitation arm against a Markov-switching multifractal Cox arm with an exact filter, in three forms, and reports where the race can be decided at all.

The central negative is a decidability result: over much of the sample-size and branching-ratio range the two mechanisms are not separable, and a Hawkes fit to a regime-switching truth infers a spurious near-critical branching ratio. On the corpus at hand the mechanism question is undecidable at the available N — which is a finding about what the data can support, and is reported as such rather than resolved by choosing the more interesting arm.