Comparisons of traditional and novel stochastic models for the generation of daily precipitation occurrences

SummaryFour stochastic models, namely, the alternating renewal process using geometric distribution, the second-order Markov chain model, the Dictionary approach, and the Probabilistic Word Matrix have been assessed to evaluate their efficacy for the generation of daily precipitation sequences. Comparisons involved three guiding principles capable of assessing the ability of models to preserve the short-term temporal-dependency in data through the concepts of autocorrelation, average mutual information and the Hurst exponent, the ability of models to preserve the persistency within the homogenous dry/wet weather states through the analysis of dry/wet-spell lengths between the observed and generated data, and the ability to assess the goodness-of-fit of models through the likelihood estimates and information theoretic criteria i.e., Akaike and Bayesian. The observed daily precipitation records for the past 30 years from 10 Canadian meteorological stations were utilized for comparative analyses of the four models. In general, the Markov chain model performed well. The other models were found to be competitive from one another depending upon the scope and purpose of the comparison. Although the Markov chain model has a certain advantage in the generation of daily precipitation occurrences, the structural flexibility offered by the Dictionary approach in modeling the varied segment lengths of heterogeneous weather states provides a distinct and powerful advantage in the generation of precipitation sequences.