A0683
Title: Bayesian model selection for analyzing predictor-dependent directional data
Authors: Ingrid Guevara - Pontificia Universidad Católica de Chile (Chile) [presenting]
Vanda Inacio - University of Edinburgh (United Kingdom)
Luis Gutierrez - Pontificia Universidad Catolica de Chile (Chile)
Abstract: Developing statistical models for directional data is increasingly important given the growing need to analyze peak hours in 24-hour services, such as Seoul's bike rental system. Motivated by the aim of identifying factors that influence demand across distinct hours, a model is introduced based on a linear dependent Dirichlet process mixture of projected normal distributions. The approach flexibly captures asymmetric and multimodal densities, while incorporating discrete spike-and-slab priors to facilitate model selection and allow model averaging to account for model uncertainty. The simulation study shows that, across various scenarios, our model successfully recovers the true functional form of the conditional density while reliably selecting the correct model structure, improving accuracy as the sample size increases. The application of the method to the data from the Seoul bike-sharing system successfully unveils that weather conditions significantly impact bike demand fluctuations across distinct hours. The approach also allows predicting peak rental times, revealing that, for instance, on a typical summer day, bike demand decreases between 8 am and 4 pm, while in winter, it drops during the early morning.
