Modeling nonlinearities with mixtures-of-experts of time series models

We discuss a class of nonlinear models based onmixtures-of-experts of regressions of exponential family timeseries models, where the covariates include functions of lags ofthe dependent variable as well as external covariates. Thediscussion covers results on model identifiability, stochasticstability, parameter estimation via maximum likelihood estimation,and model selection via standard information criteria.Applications using real and simulated data are presented toillustrate how mixtures-of-experts of time series models can beemployed both for data description, where the usual mixturestructure based on an unobserved latent variable may beparticularly important, as well as for prediction, where only themixtures-of-experts flexibility matters.