Title: Selective linear segmentation for detecting relevant parameter changes
Authors: Arnaud Dufays - Namur University (Belgium) [presenting]
Elysee Houndetoungan - Laval University (Canada)
Abstract: Change-point processes are one flexible approach to model long time series. We propose a method to uncover which model parameter changes when a change-point is detected. When the number of break points is small, an exhaustive search based on a consistent criterion is used to select the best set of parameters that change over time. In the other situation, we use a penalized likelihood approach to reduce the number of models to consider, and we prove that the penalty function will lead to a consistent selection of the true model. Estimation in such a case is carried out via the deterministic annealing expectation-minimisation algorithm. Interestingly, the method accounts for model selection uncertainty and provides a probability of selecting a specific set of covariates. Monte Carlo simulations highlight that the method works well in small and large samples for many time series models. An application on hedge funds returns shows how we can exploit the framework.