Abstract Seasonal forecasts play a crucial role in delivering early warnings to various sectors, particularly the agricultural sector. The Greater Horn of Africa region depends on rainfed agriculture, hence the need for accurate forecasts. This study uses Global Climate Models (GCMs) and satellite precipitation observations to assess the predictability of observed precipitation by deploying traditional machine learning algorithms and deep learning models. We compare the predictability of long and short rainy seasons in the region. The results highlight the challenges of forecasting the long rains season, with traditional machine learning algorithms showing low feature importance. In contrast, short rains can be predicted and achieved with high accuracy using both traditional machine learning models and deep learning architectures, particularly Long Short-Term Memory (LSTM) networks. In this study, we used ten Global Climate Models as input features for seasonal climatological forecasts, with a single output feature derived from observations of the Global Precipitation Climatology Center (GPCC) over a 30-year period (1990-2019). We measured the level of explained variance of this set of GCMs. Regardless of the method, high explainable variability was achieved in short rains, and the European Centre for Medium-Range Weather Forecasts (ECMWF) was the best predictor in the region for long rains. On the other hand, the National Aeronautics and Space Administration (NASA) was the most significant contributor to the predictions for short rains.
Hafashimana et al. (2025) studied this question.