“High-impact weather events exact a huge toll on society every year. There is much current interest in the potential of sub-seasonal forecasting to fill the gap between weather forecasts on the one hand and seasonal forecasts on the other. Sub-seasonal forecasting implies shorter averaging periods (such as pentads, weeks) than the 3-month periods typically considered by seasonal forecasts, but high-impact weather events are often daily scale and fall below even these averaging periods. An important unifying idea that we will pursue here is the concept of predicting specific high-impact weather events, but in the absence of temporal and spatial specificity. The concept involves two components: weather-within-climate, which defines climate (and its prediction) in terms of the odds of weather events happening sometime within a target period, but without specifying precisely when; and weather-within-regions, which expresses the odds of weather events happening somewhere within a region, but, again, without specifying precisely where. Our past work on weather-within-climate has demonstrated that the even seasonal forecasts can be expressed in such terms: the predictability of daily rainfall frequency is usually higher than seasonal rainfall totals in the tropics. Similarly, just as long-range forecasts of high-impact weather events necessarily lack temporal specificity, our earlier work indicates that there may be gains in predictive skill for high-impact weather events by relaxing the spatial specificity of the forecasts. By analog with the temporal case, it may be more useful to provide estimates of the probability of an extreme event pooled over a region rather than trying to predict the occurrence of an event at a precise location.
We propose to use the National Multi-Model Ensemble (NMME) database of seasonal and forecasts to investigate the predictability of sub-seasonal high-impact weather events, and to develop an MME-based downscaling system for their prediction at continental scale. Two downscaling methodologies will be used, namely IRI’s Climate Predictability Tool (CPT), based on canonical correlation analysis, together with a hidden Markov model (HMM) toolkit based on generalized linear models and non-homogeneous HMMs. The former will be applied to a hierarchical set of daily high-impact weather statistics constructed from station- or high-resolution gridded observational data, using NMME models as predictor datasets. These CPT experiments will be complemented by a NMME+HMM approach in which we will generate stochastic daily sequences of rainfall at local scale, and then aggregate these both temporally and spatially to assess the predictability of spatio-temporal aggregates. By comparing the HMM sequences from the NMME with similar sequences based on reanalyses we will also be able to diagnose the ability of the NMME to predict realistic sequences of weather, and thereby identify possible systematic errors that may limit the models’ ability to predict high-impact weather events. This exploratory pilot study will deliver new NMME prediction products through evaluations of the predictability of both heavy rainfall and drought events in the NMME models.”