Abstract: Fluid prediction in tight sandstone gas reservoirs poses a significant challenge in reservoir evaluation, which conventional methods struggle to address effectively. Constrained by limited samples and poor data quality, achieving accurate intelligent prediction using machine learning also remains difficult. This paper proposes a logging-based fluid prediction method for tight sandstone gas reservoirs using the RTF-MR-Stacking model. First, an RTF model is designed for data augmentation; then, an improved undersampling algorithm, MAHAKIL, combined with average inter-class potential, is employed to optimize the samples; finally, an enhanced Stacking module is utilized for fluid identification. Experimental results show that the proposed method achieves a prediction accuracy of 93%, effectively mitigating the impact of limited data samples on reservoir fluid prediction in the study area. Its key evaluation metrics (F1-score, G-mean, and AUC) outperform those of mainstream machine learning models and similar Stacking models. Validation results demonstrate that the RTF-MR-Stacking-based fluid prediction model exhibits good applicability in the study area, effectively addressing the challenges of tight sandstone reservoir prediction under conditions of scarce geological data. It provides high-quality technical support and theoretical guidance for productivity prediction in old and blind wells, while also offering a new perspective for traditional fluid evaluation in tight sandstone reservoirs.