Can teleconnections help to improve the seasonal prediction over the Southern African Development Community Region?

The limited skill of seasonal climate predictions in some regions of the Southern African Development Community (SADC) restricts their potential application to the development of climate services. This study explores the feasibility of improving the quality of these predictions by using the observed relationship between Essential Climate Variables (ECVs) and large-scale teleconnection indices, namely Niño3.4, the Atlantic Niño and the Indian Ocean Dipole. The underlying hypothesis is that, for certain areas, the empirical observed teleconnections could improve the predictions offered by the seasonal forecasting systems. This is achieved by implementing linear regression models between each index and ECV, for up to 12 months into the past, and for each season and grid point. After obtaining the index-derived estimates of the variables, the correlation coefficients and fair Ranked Probability Skill Scores (fRPSS) are computed and compared to those of the ECMWF SEAS5 (SEAS5) predictions for different lead times. The results show that 10–25 % of the entire domain exhibits improved correlations for the index-derived precipitation in all seasons. In the case of temperature, though, higher correlations could be observed only in six seasons (and solely for Niño3.4). Regarding fRPSS, up to 7 % of the entire area shows an improvement when using Niño3.4 to estimate temperature (in four seasons). Conversely, for precipitation there is no detected enhancement. In future work, it would be worth investigating whether a combined multi-index regression can further raise the observed increase in performance. [Display omitted] •Constrained predictive accuracy of seasonal climate predictions prevents the adoption of climate services.•Internal model uncertainties could be overcome through teleconnections between ECVs and large-scale indices.•Up to 25% of the entire Southern Africa exhibits improved accuracy.•Teleconnection-based regression models hold the potential to transcend the limitations.