A targeted methodology to study the West African monsoon (WAM) rainfall variability is considered where monthly rainfall is averaged over 10°W-10°E to take into account the latitudinal migration and temporal distribution of the WAM summer rainfall. Two observational rainfall data sets and a large number of quasi-operational forecast systems, among them two systems from the European Seasonal to Interannual Prediction initiative and six systems from the North American Multi-model Ensemble project, are used in this research. The two leading modes of the WAM rainfall variability, namely, the Guinean and Sahelian regimes, are estimated by applying principal component analysis (PCA) on the longitudinally averaged precipitation. The PCA is performed upon the observations and each forecast system and lead time separately. A statistical model based on simple linear regression using sea surface temperature indices as predictors is considered both as a benchmark and an additional forecast system. The combination of the dynamical forecast systems and the statistical model is performed using different methods of combination. It is shown that most forecast systems capture the main features associated with the Guinean regime, that is, rainfall located mainly south of 10°N and the northward migration of rainfall over the season. On the other hand, only a fraction of the forecast systems capture the characteristics of the rainfall signal north of 10°N associated with the Sahelian regime. A simple statistical model proves to be of great value and outperforms most state-of-the-art dynamical forecast systems when predicting the principal components associated with the Guinean and Sahelian regimes. Combining all forecast systems do not lead to improved forecasts when compared to the best single forecast system, the European Centre for Medium-Range Weather Forecasts System 4 (S4). In fact, S4 is far better than any forecast system when predicting the variability of the WAM rainfall regimes several months ahead. This suggests that in some special occasions like this one, a multimodel approach is not necessarily better than an especially skillful model.