Potato price forecasting poses significant challenges due to its inherent non-stationarity, abrupt trend shifts, and complex seasonal fluctuations. Existing time series forecasting models often rely on point-wise loss functions such as mean squared error (MSE), which fail to capture the underlying structural dynamics essential for accurate and robust predictions in agricultural markets. To address this gap, we propose TDLoss, a novel triplet decomposition loss function that enhances model supervision by jointly learning from three key perspectives: trend evolution, short-term changes, and frequency-domain patterns. Specifically, TDLoss encourages consistency in long-term trends through a smoothed moving average, guides the prediction of change magnitude using a directional error formulation, and preserves periodic structures via a spectral alignment mechanism. These components are adaptively combined using a learnable weighting scheme, enabling the loss function to focus on the most informative structural cues during training. We demonstrate the effectiveness of TDLoss on a real-world, multi-region potato price dataset, integrating it with multiple neural forecasting backbones. Experimental results show that TDLoss consistently improves forecasting accuracy and robustness across different models and prediction horizons. The proposed loss function is lightweight, generalizable, and particularly well-suited for structure-aware time series forecasting in volatile agricultural domains.
Full publication URL
