Forecasting natural disaster frequencies using nonstationary count time series models

Because of the climate change, the frequency of natural disaster might be evolving. This could mean an increasing expected risk, and/or more and more uncertainties. In this paper, we identify three potentially suitable models, that are the nonstationary INGARCH(1, 1), the nonstationary INAR(p), and the state-space model of Harvey and Fernandes (J Bus Econ Stat 7(4):407–417, 1989). We derive properties of their long-run behavior, discuss their link and differences, and assess their suitability for Canadian climate event data. We show that first, the Harvey–Fernandes model or INGARCH(1, 1) model often provides better fit and better prediction performance. Second, the long-run prediction of these models can differ substantially, highlighting model uncertainty.