Title: Statistical learning of cyclic autocorrelation functions with application to streamflow data modelling
Authors: Samuel Perreault - University of Toronto (Canada) [presenting]
Abstract: A cyclostationnary process is considered and the usual Kendall autocorrelation is modified to account for the dynamic, yet periodic nature of the process under study. This is done by letting the autocorrelation be a function of the usual lag parameter, as well as a periodic time index indicating, e.g., the time of the year. We then describe a learning algorithm whose purpose is to smooth the empirical autocorrelation along both the temporal (cyclic time index) and lag dimensions. We apply it to Canadian daily streamflow data and show how it can be used for testing certain hypotheses, such as the presence of negative dependence within a seasonal cycle.