A Controlled Empirical Evaluation of Input Design and Temporal Architectures for Multi-Horizon Forecasting of Onion Wholesale Prices

Abstract

Onion wholesale prices in South Korea exhibit substantial volatility due to seasonal production concentration, storage-dependent supply adjustments, and weather sensitivity, making accurate multi-horizon forecasting important for agricultural supply management. While previous research has predominantly focused on model development, comparatively few studies have systematically evaluated input design alongside representative forecasting architectures under a unified experimental protocol. This study presents a controlled empirical comparison for daily onion wholesale price forecasting using historical price, calendar, weather, and market-arrival data from 2000 to 2025. We evaluate three input window lengths (180, 365, and 730 days), five feature sets, and eight representative forecasting architectures across three forecast horizons (30, 90, and 180 days) under a walk-forward protocol incorporating three rolling-origin evaluation folds (2023–2025), fold-specific validation tuning, and multi-seed retraining. Utilizing repeated-measures blocks defined by model and fold, a statistically significant window-length effect was identified only at the 30-day horizon, where the 365-day window yielded significantly lower forecasting error than the 730-day window (Friedman test with Holm-corrected Wilcoxon signed-rank post-hoc tests); no statistically significant overall feature-set effect was observed at any horizon. Furthermore, Holm-corrected pairwise Diebold–Mariano tests based on validation-selected configurations revealed no statistically significant differences among the eight architectures at any of the three forecast horizons, despite considerable variation in numerical performance metrics. These empirical findings indicate that under a controlled, repeated-measures evaluation protocol, only a limited subset of observed numerical differences in input design are statistically distinguishable, while architectural performance differences remain statistically indistinguishable in multi-horizon onion price forecasting.