Price volatility in agricultural commodities impacts food security, farmer income, and market stability, with potatoes in India providing a key example due to their dietary and economic importance. Traditional models such as autoregressive integrated moving average (ARIMA) offer useful benchmarks but struggle to capture nonlinear dependencies and redundancy in multi-centre data. To address this gap, this study proposes a forecasting framework that integrates machine learning models with binary metaheuristic feature selection. Daily potato price data from Indian districts were analyzed using nine binary optimizers, including the Binary Puma Optimizer–Stochastic Fractal Search (bPOSFS), Binary Puma Optimizer (bPO), and Binary Stochastic Fractal Search (bSFS). Feature subsets were optimized using a fitness function balancing prediction error and subset compactness, and performance was assessed with MSE, RMSE, MAE, $${R}^{{2}}$$, RRMSE, NSE, WI, and correlation coefficient, supported by ANOVA and Wilcoxon tests. Baseline models performed modestly (ARIMA: RMSE = 0.1476, $${R}^{{2}} {= 0.3883}$$), while metaheuristic feature selection achieved substantial gains. bPOSFS outperformed all competitors, yielding the lowest average error (0.5438) and best fitness (0.6099), with improvements statistically significant ($${p < 0.0001}$$ for ANOVA, $${p = 0.002}$$ for Wilcoxon). These findings confirm that metaheuristic-driven feature selection enhances agricultural price forecasting and provides interpretable, robust tools for policymakers, traders, and farmers.
Full publication URL
