Nonparametric density estimation by switching

According to standard MDL and Bayesian model selection, we should (roughly) prefer the model that minimises overall prediction error. But if the goal is to predict well, it may well depend on the sample size which model is most useful to predict the next outcome. By re-interpreting the Bayesian prediction strategies associated with the models as "experts", we can use the various algorithms for "expert tracking" to improve model selection for prediction without introducing a substantial computational overhead.