The agricultural market is characterized by unpredictable price fluctuations influenced by diverse uncontrollable factors, often leading to a mismatch between production and income, thus creating a vicious cycle. Accurate forecasting of cucumber prices is paramount to enhancing the agricultural industry’s development and safeguarding the interests of all stakeholders in the cucumber a supply chain. In this study, we introduce a novel multi-task learning model that leverages Long Short-Term Memory (LSTM) neural networks and self-attention mechanisms to jointly predict cucumber prices and production. This model is designed to capture the intricate relationship between these two variables, enhancing feature extraction through the incorporation of self-attention. By adopting a multi-task learning approach, we aim to gain a deeper understanding of the interplay between cucumber prices and production. Our experimental results confirm the superiority of the proposed model in terms of predictive accuracy and generalization ability when compared to traditional methods. A comprehensive robustness analysis further demonstrates the model’s stability across various scenarios. Specifically, the self-attention multi-task LSTM model exhibits smaller average absolute and root mean square errors than traditional neural network models. This research contributes a valuable tool for agricultural decision-making, empowering farmers and policymakers with more precise predictions of cucumber prices and production. Thus, in turn, it facilitates better adaptation to market fluctuations and promotes overall agricultural productivity.