Developing and Validating a Machine Learning Framework for Sub-Seasonal Rainfall Forecasting Over Nigeria Beyond Two Weeks

The sub-seasonal to seasonal (S2S) forecasting gap, spanning two to six weeks, represents one of the most scientifically challenging frontiers in operational meteorology and a critical unmet need for agricultural and disaster risk management in Nigeria. This study develops and validates an original machine learning framework for sub-seasonal rainfall forecasting over Nigeria at lead times of two to six weeks, filling a documented gap in NiMet's operational forecast portfolio. The framework combines large-scale teleconnection indices including the Madden-Julian Oscillation (MJO), ENSO, and Atlantic SST anomalies with reanalysis-derived atmospheric predictors in an ensemble of machine learning models including Random Forest, Gradient Boosting, and a deep learning Long Short-Term Memory architecture. Training was performed on 30 years of ERA5 reanalysis data, with model outputs evaluated against NiMet station observations using a rigorous nested cross-validation framework to prevent data leakage. Original contributions include derivation of a Nigeria-specific MJO rainfall teleconnection climatology demonstrating that MJO phases 2 to 3 and 7 to 8 are associated with statistically significant rainfall anomalies over specific geopolitical zones, and the development of a probabilistic ensemble output system generating tercile probability forecasts at the state level. Results demonstrate that the ML framework achieves Brier Skill Scores of 0.18 to 0.31 at three-week lead time, substantially outperforming both ECMWF S2S model direct output and climatological forecasts. The operationalization pathway for integration into NiMet's forecast production system is specified. Keywords: sub-seasonal forecasting, machine learning, MJO, Nigeria, Brier skill score.