Multisite precipitation generation using a latent autoregressive model

Stochastic simulation of precipitation at multiple sites in a region can be accomplished using a latent multivariate autoregressive Gaussian process as the driver of both precipitation occurrence and precipitation amounts. This idea was proposed by Bardossy and Plate (1992) and has been used in several subsequent studies. We investigate this modeling framework in more detail, considering both the single-site and the multisite cases. Among other things, we demonstrate how model parameters can be conveniently estimated using the method of maximum likelihood for censored data. We explore various theoretical properties of the model, including its connection to traditional Markov chain models frequently used to model precipitation occurrence. A case study based on data from Manitoba, Canada, is used to demonstrate the range of statistics that can be reproduced by the model and to highlight potential limitations.