Abstract: Tight sandstone gas reservoirs are an important target for oil and gas exploration. However, their complex pore structures, low permeability, and high bound water content pose significant challenges for well logging evaluation. Although nuclear magnetic resonance (NMR) logging can effectively distinguish irreducible water from movable water, it is not available in all wells. On the other hand, bound water saturation and permeability estimation from conventional log data is not sufficiently accurate and requires further study. This study proposes a stacking-based ensemble learning method for estimating bound water saturation and permeability in tight sandstone reservoirs. The proposed ensemble learning framework integrates petrophysical models with data-driven strategies, and uses bound water saturation from NMR logs as the target and conventional logs together with rock-physics-derived parameters as inputs. It employs Random Forest, Extreme Gradient Boosting, and Gradient Boosting as base learners, with Random Forest as the meta-learner. The initial predictions from multiple base learners are used as features to train the meta-learner through cross validation, enabling bound water saturation prediction from conventional well logs. Permeability of tight sandstone gas reservoirs is then calculated using the Timur-Coates equation regressed from rock physics experimental data combined with the predicted bound water saturation. Application of this method to the HG Member in Block C of the XH sag in the East China Sea Shelf Basin achieves mean absolute errors of 5.1% on the validation set, 4.87% on multi-well data, and 6.1% on the new Well C5 for bound water saturation, with a mean logarithmic error below 0.3 for permeability. The proposed method significantly improves the accuracy of bound water saturation and permeability estimation in tight sandstone reservoirs and demonstrates strong adaptability and applicability for broader use.