Temporal feature engineering for agricultural commodity price forecasting in Nigeria: evaluating machine learning, deep learning and time-series approaches

Abstract

This study investigates the effectiveness of machine learning, deep learning and traditional time-series approaches for predicting agricultural commodity prices in Nigeria, focussing on improving forecasting reliability through temporal feature engineering and time-aware evaluation using the WFP Nigeria dataset. The study uses 60,566 observations across Nigerian commodity markets (2002–2025). Linear Regression, Random Forest, ARIMA, and LSTM models were evaluated using lag and rolling-average temporal features, with walk-forward validation preserving temporal integrity for realistic assessment. Machine learning models outperformed traditional statistical approaches. Random Forest (R2 = 0.497) exceeded Linear Regression (R2 = 0.465), while ARIMA yielded a negative R2 (−0.259). LSTM achieved its strongest performance for maize (R2 = 0.640) and sorghum (R2 = 0.468), while several highly volatile commodities produced negative R2 values despite extended training and Early Stopping. These results highlight the role of commodity-specific characteristics in price forecasting, with walk-forward validation revealing considerable temporal performance variation. The study relies primarily on historical price data and does not incorporate exogenous variables such as weather, inflation, transportation costs or policy interventions. This study contributes to the agricultural forecasting literature by integrating temporal feature engineering, chronological validation, walk-forward evaluation, and comparative modelling within a unified framework using a Nigerian commodity price dataset, providing commodity-level evidence on the varying effectiveness of machine learning, deep learning and traditional statistical approaches under realistic forecasting conditions.