To accurately assess the soil water-holding capacity of urban green spaces, the topsoil from park green spaces in six cities of Henan Province was selected as the research object. Physicochemical indicators such as dry bulk density (DBD), saturated water rate (SWR), particle composition, and saturated hydraulic conductivity (K_s) were determined through field sampling. Multiple linear regression and random forest models were constructed to predict K_s, and a comprehensive evaluation of water-holding capacity was conducted by combining the entropy weight method, CRITIC method, and improved game theory-based combined weighting method. The results showed that the soil silt content in the study area was the dominant particle fraction, and significant spatial variability was observed in K_s. The random forest model exhibited the optimal prediction accuracy (R=0.979, RMSE=3.077 for the test set), while the multiple linear regression model (R=0.936, RMSE=16.40) also outperformed the classic pedo-transfer functions considerably. The combined weight results indicated that DBD (weight: 28.69%) was the primary factor affecting water-holding capacity, followed by sand content (weight: 16.90%) and SWR (weight: 15.34%). The regional water-holding capacity scores revealed that wetland and natural park areas had relatively high scores, whereas urban squares and roadside green spaces had lower scores. The prediction models and evaluation methods established in this study can provide a scientific basis for optimizing the hydrological functions and implementing hierarchical management of urban green space soils.