On the inherent predictability of precipitation across the United States

Predictability of climate variables is known to be limited to few days up to few weeks due to the inherent chaotic nature and resulting sensitivity to initial conditions. However, such generalization of limited predictability is cautioned because of the highly nonlinear nature and the known influence of localized causal factors on many climate variables. Additionally, even though an improvement in predictability is expected with coarsening in spatial and temporal resolutions, the extent and rate of this expected improvement is still unexplored. This study investigates the spatial distribution of predictability of daily precipitation across the USA. The emphasis is on determining the rate of increase in predictability with spatio-temporal averaging, by defining three predictability statistics (maximum predictability, predictive error, and predictive instability) based on the nonlinear finite-time Lyapunov exponent. From our analyses, we find that predictability increases monotonically with temporal averaging, while spatial averaging has minimal influence, pointing to the possible spatially invariant nature of precipitation dynamics. Modeling the precipitation dynamics at relatively coarser scales of 1° × 1° and higher temporal scales of 5–10 days could markedly improve the predictability statistics. Significant changes in the predictability characteristics of daily precipitation across large areas of the USA and associated non-stationarity are also identified over the 1948–2006 period. This is consistent with sudden changes in the overall nature of precipitation over time, which include a reduction in non-rainy days, an increase in signal-to-noise ratio, an increase in average precipitation events, and an increase in extremes.