Hybrid Machine Learning Models for Optimized Potato Price Prediction

Accurate prediction of agricultural commodity prices, such as potatoes, is crucial for enhancing market efficiency, supporting supply chain decisions, and ensuring economic stability in the agricultural sector. This study proposes an enhanced machine learning framework for potato price prediction using Light Gradient Boosting Regression (LGBR), optimized through two metaheuristic algorithms: the Stochastic Paint Optimizer (SPO) and the Population-based Vortex Search Algorithm (PVSA). The hybrid models LGSP (LGBR+SPO) and LGPB (LGBR+PVSA) were developed to reduce prediction error and improve generalization. Experimental results demonstrate that the optimized models outperform the baseline LGBR model. Specifically, LGPB achieved the lowest training mean squared error (MSE) of 3.33E + 03, though it increased to 6.30E + 03 in validation, indicating a potential overfitting issue. LGSP achieved moderate performance with a training MSE of 5.35E+03 and validation MSE of 7.77E+03. In contrast, the baseline LGBR model had the highest MSE values in both training (1.13E+04) and validation (1.34E+04), reflecting weaker predictive accuracy. Uncertainty measures (U95) followed a similar trend. The findings confirm that metaheuristic optimization can significantly improve regression performance in price forecasting tasks. However, challenges in model generalization highlight the need for further tuning and diverse datasets.