Inference for Heavy-Tailed Max-Renewal Processes

Max-renewal processes, or Continuous Time Random Maxima, assume that events arrive according to a renewal process, and track the running maximum of the magnitudes of events up to time $t$. In many complex systems of interest, notably earthquakes, trades and neuron voltages, inter-arrival times exhibit heavy-tailed distributions. The dynamics of events then exhibits memory, which affects the rate at which events occur: rates are highly variable in some time intervals, while other intervals have long quiescent periods, a behaviour which has been dubbed bursty in the physics literature. This article provides a statistical model for the exceedances $X(ell)$ and interexceedance times $T(ell)$ of events whose magnitude exceeds a given threshold $ell$. We derive limit theorems for the distribution of $X(ell)$ and $T(ell)$ as $ell$ approaches extreme values. The standard Peaks Over Threshold inference approach in extreme value theory is based on the fact that $X(ell)$ is approximately generalized Pareto distributed, and models the threshold crossing times as a standard Poisson process. We show that for waiting times with infinite mean, the threshold crossing times approach a fractional Poisson process in the limit of high thresholds. The inter-arrival times of threshold crossings scale with $p^{1/beta}$, where $p$ is the threshold crossing probability and $beta in(0,1)$ is the tail parameter for the waiting times. We provide graphical means of estimating model parameters, and show that these methods provide useful results on simulated and real-world datasets.