NextGen: A Next-Generation System for Calibrating, Ensembling and Verifying Regional Seasonal and Subseasonal Forecasts

Successful climate services often involve the use of tailored regional climate forecasts at one or multiple timescales. The way those forecasts are implemented is not always straightforward, and depends on several different factors, like which variables, models and calibration methods to use, how to produce the ensemble and tailoring, or even how to present them to the decision makers. Here, NextGen, a systematic general objective approach for designing, calibrating, building ensembles, and verifying objective climate forecasts is presented and discussed. NextGen involves the identification of decision-relevant variables by the stakeholders, and the analysis of the physical mechanisms, sources of predictability and suitable candidate predictors (in models and observations) for those key relevant variables. In those cases when prediction skill is deemed high enough, NextGen helps select the best dynamical models for the region of interest through a process-based evaluation, and automates the generation and verification of tailored multi-model, statistically calibrated predictions at seasonal and sub-seasonal timescales, at regional, national or sub-national level.

The system takes advantage of the expertise of forecasters and local scientists at the country's national meteorological service and universities, to maximize predictive skill and the potential to tailor the forecasts. Rather than focusing on probabilities of above normal, normal and below normal categories of total rainfall or mean temperature, NextGen also provides probabilities of exceeding (or not) particular thresholds of interest in the decision-making process, thus enabling users to forecast with the same system both mean and extreme values. Furthermore, NextGen is a general, flexible approach that helps produce probabilistic forecasts not only of the total amount of rainfall expected during the next season or next few weeks, but also rainfall characteristics, like how precipitation will be distributed: frequency of rainy/dry days in the target season, onset, demise and duration of the rainy season(s). Other variables of interest can also be provided by the system if quality data is available and the variables exhibit enough predictability.

Several examples are discussed, from the forecast generation phase using a publicly available Python interface of the International Research Institute for Climate and Society's Climate Predictability Tool (PyCPT, https://github.com/agmunoz/PyCPT), to climate-service-oriented products like a Forecast-based Finance pilot for drought, and predictions of potential risk of transmission of mosquito-borne diseases.