Dynamic kernel models

This paper introduces the family of Dynamic Kernel models. These models approximate the predictive density function of a time series through a weighted average of kernel densities possessing a dynamic bandwidth. A general specification is presented and several particular models are studied in detail. We propose an Mestimator for model parameters and derive its asymptotic properties under a misspecified setting. A consistent density estimator is also introduced. Monte Carlo results show that the new models effectively track the time-varying distribution of several data generating processes. Dynamic Kernel models outperform extant kernel-based approaches in tracking the predictive distribution of GDP growth.