Forecasting Cereal Production and Its Implications for Food Security in Somalia: A Comparative Analysis of ARIMA, ETS, and Neural Network Models

Abstract

Cereal production in Somalia is characterized by extreme volatility driven by climate shocks. This study addresses the limitations of traditional agricultural planning by evaluating optimal modeling techniques to predict future output. Utilizing historical time series data from World Bank databases spanning 1961 to 2023, the research conducted a comparative analysis of linear models, specifically ARIMA and ETS, against non-linear Neural Network Autoregressive (NNETAR) algorithms. The forecasting precision of eight distinct models was rigorously validated on unseen test data using Root Mean Square Error (RMSE) and Symmetric Mean Absolute Percentage Error (SMAPE) metrics. The empirical results demonstrated that the NNETAR model significantly outperformed traditional statistical benchmarks, achieving the lowest error rates (SMAPE of 40.42%) by effectively capturing the non-linear structural breaks inherent in the dataset. Conversely, linear models exhibited a systemic bias toward over-forecasting, while the NNETAR 10-year projection (2024-2033). These findings establish Neural Networks as a superior instrument for agricultural planning in volatile regions, validating the shift from traditional econometrics to computational intelligence. Ultimately, the study advocates for integrating AI-driven forecasting into early warning systems to enable policymakers to transition from reactive crisis management to proactive food security strategies.