Statistical Approaches for Applying Member Weighting Techniques in Seasonal Weather Forecasting

This work introduces a statistical post-processing approach designed to enhance the predictive performance of global seasonal forecast systems, developed within the framework of a forecasting assignment led by the Spanish State Meteorological Agency (AEMET).

 

The study is carried out over the Iberian Peninsula (IP), located in the Northern Hemisphere's mid-latitudes. As happens with midlatitudes, this area shows low predictability on seasonal timescales due to high internal variability and low signal-to-noise ratio. However, some works have found a window of opportunity for North Atlantic Oscillation (NAO) predictability, which, in turn, can lead to improvements in the skill of the forecasts for climatic parameters (Baker, 2018) (Sánchez García et al., 2019) (Trigo, 2004). Following this example, the NAO and other teleconnection patterns are used through the methodology, and the potential for improvement, by weighting the ensemble accordingly, is explored

 

Two statistical approaches are applied: Empirical Orthogonal Functions (EOFs) and Partial Least Squares (PLS) regression. Both methodologies are employed on various global seasonal models, first without applying a weighted member technique and then applying an ensemble member weighting approach to the whole ensemble. The member-weighted technique uses the prediction of the best-performing models for the analysed variability patterns. These patterns, derived from ERA5 reanalysis data, are treated as "perfect forecasts" of dominant circulation modes.  Deterministic and probabilistic verification metrics are used - according to the definitions provided in the WMO forecast guidance (WMO, 2018) - to evaluate the potential for improvement. 

 

Results indicate that some variability patterns demonstrate skill for predicting months in advance, suggesting the potential to enhance seasonal forecasting through model weighting methods incorporating variability mode information. However, additional improvements could still be found: the potential for improvement is conditioned by the definition of variability patterns used.