Multivariate Correction and Statistical Disaggregation for Seasonal Multi-Model Ensemble Forecast Applications

Seasonal multi-model ensemble forecasts are quickly becoming a more routinely used product. The North American Multi-Model Ensemble (NMME), as an example, represents a concerted effort to generate a standardized archive of MME forecasts that adhere to an operational forecast schedule. These forecasts are being used and evaluated for generating seasonal outlooks by the NOAA Climate Prediction Center. More than seasonal outlooks, however, these data are also potentially suitable for use by the impact/application (e.g. hydrology, agriculture) modeling communities. However, their usage is not direct as the models are typically run at coarse resolution (e.g. ~100 km) and may only be archived at monthly resolution. The well-known need for some form of downscaling — either dynamical or statistical — is thus required for seasonal multi-model ensemble forecasting. Further, biases are prevalent in many of the seasonal forecasts. To address biases many MME systems include protocols for simulations over a period 20-30 years prior to real-time. These hindcasts provide an opportunity to characterize and remove these biases. Traditionally, the downscaling community has focused on development of techniques to address long-term climate projections. The bias correction and statistical disaggregation (BCSD) approach, in particular, relies on local bias correction and bootstrap resampling from a historical dataset to provide downscaled scenarios. This technique can also be applied to seasonal forecasts. However, the local bias correction does not take advantage of the wealth of literature on multivariate bias corrections for seasonal forecasting that attempts to address bias in spatial structures. As an example, local bias correction implemented via quantile-quantile mapping (or other rank-preserving approaches) is unable to improve the probabilistic skill (e.g. ranked probability skill score) of MME forecasts. The focus of this presentation will be on the development of a new multivariate bias correction (e.g. CCA/RDA-based) and statistical disaggregation (MCSD) methodology to generate daily, downscaled seasonal forecasts for application modeling. The MCSD technique will be compared to and contrasted with the BCSD approach for use in seasonal forecasting. Results will be presented for downscaling over the East Africa region, but the discussion will stress the benefits and limitations of MCSD technique itself. Emphasis will placed on domain/predictor selection, limited (hindcast) sampling, and overfitting.