Accurate production prediction of fractured horizontal wells is important for productivity evaluation and development-plan adjustment in tight oil reservoirs. However, newly producing wells often have short production histories, while production behavior varies significantly among wells due to reservoir heterogeneity, stimulation differences, and production operations. This study proposes a similarity-guided sequential transfer learning framework for short-history production prediction. Candidate source wells are selected by jointly considering inter-well distance, static geological-engineering similarity, and dynamic production similarity. Their production dynamics are then sequentially transferred into a target-well gated recurrent unit (GRU) model along optimized transfer paths, rather than directly mixing multi-well samples. The framework was evaluated using production data from 40 hydraulically fractured horizontal wells in a tight oil reservoir block of the Daqing Oilfield and compared with the Arps decline model, tree-based models, back-propagation neural network (BPNN), and conventional GRU. Results show that the proposed method improves prediction stability under limited target-well histories, with 30 target wells achieving high-quality predictions and only two classified as low-quality predictions. Abnormal-well diagnosis indicates that persistent prediction bias mainly results from low source-target transferability. Data-efficiency analysis shows that approximately 20 months of production history provides initial engineering usability, while 25 months or longer leads to more stable prediction. The proposed method provides a practical tool for new-well production forecasting and dynamic adjustment of tight oil development plans.