Testing Local Self-Similarity in Univariate Heavy-Tailed Data
Rakhee Dinubhai Patel
Advisor: Frederic Paik Schoenberg
The Pareto distribution, or power-law distribution, has long been used to model phenomena in many fields, including wildfire sizes, earthquake seismic moments and stock price changes. Recent observations have brought the fit of the Pareto into question, however, particularly in the upper tail where it often overestimates the frequency of the largest events. This research proposes a graphical local self- similarity test specifically designed to assess whether a Pareto distribution fits better than a tapered Pareto or another alternative. Unlike some model selection methods, this graphical test provides the advantage of highlighting where the model fits well and where it breaks down. Specifically, for data that seem to be better modeled by the tapered Pareto or other alternatives, the test assesses the degree of local self-similarity at each value where the test statistic is computed. The basic properties of the graphical test and its implementation are discussed, and applications of the test to wildfire, seismological and financial data are considered.