<p>Accurate forecasting of agricultural commodity prices is vital for enhancing market transparency, guiding farmers’ production and marketing decisions, and informing effective policy interventions. Yet, the non-linear, non-stationary, and multi-scale nature of price time series limits the effectiveness of conventional forecasting approaches. This study proposes a novel hybrid framework that combines complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) and the PatchTST model to deliver robust and scalable agricultural price predictions. CEEMDAN decomposes the raw price series into a set of stationary intrinsic mode functions (IMFs), isolating meaningful oscillatory patterns and suppressing noise. These IMFs are then jointly forecast using a single PatchTST architecture, which exploits channel-independent attention and patch-based sequence learning to efficiently capture both fine-grained fluctuations and longer-term temporal dependencies. The effectiveness of the proposed framework is evaluated using daily potato price data over three forecasting horizons: 1-day, 10-day, and 20-day, corresponding to short-, medium-, and long-term predictions. Results show that CEEMDAN–PatchTST consistently outperforms both standard deep learning models and their CEEMDAN-based variants at all horizons. Notably, at the 20-day horizon, it achieves a mean absolute percentage error (MAPE) of just 2.86% and an <i>R</i>² exceeding 0.97, demonstrating its robustness even under greater forecast uncertainty. These findings underscore the complementary strengths of CEEMDAN and PatchTST in addressing the complex dynamics of agricultural price series, establishing the framework as a reliable, accurate, and computationally efficient solution for commodity price forecasting. Beyond agriculture, the methodology shows promise for broader applications to other domains characterised by noisy, non-linear, and multi-scale time series data.</p>