Probabilistic assessment of the practical predictability of extreme wet and dry years in the southwestern US in observed and CMIP6 climates

Abstract:

Despite receiving considerable attention, early and reliable prediction of extreme precipitation totals over the southwestern US (SWUS) remains a challenge, with great implications for the region’s economy. To aid with policy and decision making, here we explore the predictability of extreme dry/wet and normal precipitation years with a lead time of one month. Specifically, for each year in the last five decades, we use copula-based statistical models and estimate the entire predictive distribution of precipitation, conditioning on various combinations of sea surface temperature (SST) indices. Prediction is performed in a five-fold cross validation setting to limit overfitting, using both observations and state-of-the-art climate model simulation outputs. We design statistical metrics to assess the null hypothesis that there is no predictive skill. The theoretical and asymptotic values of these metrics (under the null hypothesis) are provided, and preasymptotic distributions are obtained via Monte Carlo. We find that copula models are more reliable when predicting wet years than dry years, while normal years are virtually not predictable. Our results also highlight that, with respect to its relevance to SWUS precipitation, climate models on average overestimate the importance of El Niño-Southern Oscillation (ENSO) and undermine the importance of non-ENSO SST variability.