An adaptive method for bandwidth selection in circular kernel density estimation

Abstract Kernel density estimations of circular data are an effective type of nonparametric estimation. The performance of these estimations depends significantly on a smoothing parameter referred to as bandwidth. Selecting suitable bandwidths for these types of estimation pose fundamental challenges, therefore fixed bandwidth selectors are often the initial choice. The study investigates common bandwidth selection methods and proposes novel methods which adopt the idea from the linear case. The attention is also paid to variable bandwidth selection. Using simulations which incorporate a range of circular distributions that exhibit multimodality, peakedness and skewness, the proposed methods were evaluated and then compared with other bandwidth selectors to determine their potential advantages. Two real datasets, one containing animal movements and the other wind direction data, were applied to illustrate the utility of the proposed methods.