### A Flexible Bayesian Model for Clustering Seasonal Time Series With Linear and Circular Components

Statistics Seminar

29th September 2017, 3:30 pm – 4:30 pm

Main Maths Building, SM3

In this work we present a new model for clustering spatial time data, i.e. a bivariate time series with one linear and one circular observation. We propose a mixture model where the mixing probabilities are time specific and are assumed to follow a Logistic-Normal distribution. We introduce dependence betweens the vectors of mixing probabilities by means of the Gaussian processes representation of the Logistic-Normal distributions. Then through correlation functions defined over a mixed circular-linear domain, we can evaluate seasonal effects. The seasonal periods are considered as latent variables and then estimated along with the other model parameters. We show how to implement the model in a Bayesian framework. We estimate the model on an animal movement dataset and we compare our results with the ones obtained with an hidden Markov model, showing a great improvement in terms of interpretability and DIC (Deviance Information Criterion.)

*Organiser*: Song Liu

## Comments are closed.