Abstract
Precise long-term rainfall prediction is important for agricultural planning, climate resilience, and reducing disaster risk, particularly for countries like Nigeria with diverse regimes of rainfall. In this research, the potential of machine learning (ML) and statistical models to predict monthly univariate rainfall in 24 Nigerian stationswas evaluated. Model training employed historical rainfall data (1960–1999), while validation was carried out for 11 years (2000–2010). SARIMA (p; d; q) (P; D; Q)s models were used in Minitab®, R, and Python, and the most important parameters (p; d; q; P; D; Q) were tuned manually and by using auto.arima(). ML models such as feedforward neural networks, adaptive neuro-fuzzy inference systems, support vector regression and random forest were utilized in MATLAB®and R with hyperparameter-tuned models. Model performancewas evaluated in using statistics such as root mean square error (RMSE) and coefficient of determination (r2). SARIMA performed best in areas where rainfall variability was minimal. Nguru (12.03°N), the area with the lowest average monthly rainfall (35.71 mm), showed the highest SARIMA estimation with RMSE of as low as 7.84mm and r2 of as high as 0.85. ML models underperformed in capturing seasonal dynamics. For instance, SVR failed to model temporal trends effectively, while random forest produced nearly constant outputs across all years. Adjustments to SARIMA parameters (e.g., setting seasonal differencing D = 0 or Q = 1) were essential in reducing unrealistic forecasts. The findings demonstrate that SARIMA, with proper tuning, is better suited for univariate rainfall forecasting in Nigeria than non-customized ML models. Forecast reliability strongly correlates with regional rainfall characteristics and model sensitivity to seasonality.