Item

Abstract

Operational forecasts using direct output from numerical weather prediction models exhibit poor skill over northern tropical Africa compared to simple climatology-based forecasts. A recent study found potential in using Spearman’s rank correlations of gridded rainfall estimates from TRMM to predict July-September tropical African rainfall. Using the satellite-based gridded GPM-IMERG product from 2001-2019, we build on this approach using the Coefficient of Predictive Ability (CPA) developed for improved variable selection for statistical models and expanding up to 3-day lags over tropical Africa and the Atlantic Ocean. High CPAs straddling the zonally oriented rainbelt are attributed to large-scale drivers causing coherent spatio-temporal anomalies. Low CPAs over the rainbelt centre indicate the lack of dominant forcing, high stochasticity, or both. The coherent-linear-propagation factor (coh) introduced at every grid point quantifies the coherence of the identified rainfall by summarising the extent to which lagged CPAs reflect propagation with constant phase speed and direction. High coh over the Sahel suggests African easterly waves’ dominance. Stochastic precipitation driven by small-scale processes causes low coh over the rainbelt. We train a statistical model using the rainfall linked to the identified CPAs and compare it with three benchmarks: a climatology-based forecast, the ECMWF 1-day ensemble prediction-system forecast and its statistically post-processed output. The statistical model is outperformed only by the post-processed output and only in the western Sahel and central Africa. However, the Diebold-Mariano test for forecast significance suggests no significant differences between the statistical and post-processed forecasts making the former a cheaper alternative over the analysis region.