Accurate forecasting of potato prices is a fundamental task for market regulation and supply chain stability. However, price series are often volatile, nonlinear, and affected by abrupt fluctuations, which challenge traditional statistical and deep learning models. This study introduces a directional and ordinal consistency (DOC) framework that enhances conventional regression objectives by integrating directional classification and ordinal ranking consistency within a unified multi-task design. The approach simultaneously constrains the model to maintain correct movement direction and preserve the local order structure of price sequences, improving both pointwise accuracy and temporal coherence. Experiments on a high-resolution provincial potato price dataset show that the proposed method consistently achieves superior forecasting performance and generates smoother, more interpretable temporal trajectories compared with recurrent, convolutional, and transformer-based baselines. The results indicate that incorporating behavioral consistency priors provides an effective principle for stable and reliable potato price forecasting under highly dynamic market conditions.
Full publication URL
