Comparison of Seasonal Potential Predictability of Precipitation

Abstract

Abstract Three methods for estimating potential seasonal predictability of precipitation from a single realization of daily data are assessed. The estimation methods include a first-order Markov chain model proposed by Katz (KZ), and an analysis of covariance (ANOCOVA) method and a bootstrap method proposed by the authors. The assessment is based on Monte Carlo experiments, ensemble atmospheric general circulation model (AGCM) simulations, and observation-based data. For AGCM time series, ANOCOVA produces the most accurate estimates of weather noise variance, despite the fact that it makes the most unrealistic assumptions about precipitation (in particular, it assumes precipitation is generated by a Gaussian autoregressive model). The KZ method significantly underestimates noise variance unless the autocorrelation of precipitation amounts on consecutive wet days is taken into account. Both AGCM and observation-based data reveal that the fraction of potentially predictable variance is greatest in the tropics, smallest in the extratropics, and undergoes a strong seasonal variation. The three methods give consistent estimates of potential predictability for 67% of the globe.