Malaria remains a leading climate-sensitive disease in West Africa, where temperature and rainfall variations strongly influence mosquito dynamics and transmission intensity. Senegal and Nigeria, representing the Sahel-Sudanian and rainforest-savanna ecological gradients, are among the most vulnerable zones. However, few climate–malaria forecasting studies combine systematic benchmarking, uncertainty quantification and model explainability within a single framework, or contrast two epidemiologically distinct settings. This study addresses that gap with three objectives: to benchmark a deep recurrent architecture against alternative machine-learning models in two contrasting settings; to quantify predictive uncertainty and validate the framework against independent surveillance data; and to identify the climatic predictors and lags driving the predictions. The BiLSTM implemented here captures the long-term temporal dependencies inherent in climate-driven transmission dynamics. Historical climate data (1985–2014) from the CMIP6 historical experiment were combined with CMIP6 multi-model future projections (2015–2080) under the SSP585 scenario to assess the climatic determinants of malaria incidence. The BiLSTM was calibrated using rainfall and temperature anomalies and benchmarked against six machine-learning models (standard LSTM, GRU, ANN, SVR, Random Forest and Decision Tree), which it matched or exceeded, and externally validated using independent national malaria surveillance data from Senegal and Nigeria. Results show that eastern Senegal experienced a +1.5 °C warming and 20% rainfall decline, while northern Nigeria warmed by +1.2 °C with a 15% rainfall increase. Under SSP585, additional warming of +2–3 °C is projected. The proposed BiLSTM model achieved high predictive accuracy in reproducing the VECTRI-simulated malaria incidence (R2 between 0.93 and 0.98 across the training, validation and test partitions of both countries). Spatial projections indicate persistent hotspots in southeastern Senegal (Kedougou) and a northward expansion of risk zones in Nigeria by mid-century under SSP585. When the same framework is instead trained and evaluated on observed malaria surveillance data with observed NASA POWER climate in an autoregressive, one-step-ahead setting, it yields a more realistic estimate of operational forecasting skill (test R2 = 0.72 for Senegal and 0.77 for Nigeria); the near-perfect accuracy obtained against the climate-generated target should therefore be interpreted as an upper bound rather than real-world performance.