Accurate estimation of risk to residential structures from hurricane winds is critical for emergency planning and post-event recovery. Fragility curves are widely used for assessing wind damage risk at the county and census tract levels in models such as HAZUS-MH. Large-scale evaluation of the predictive accuracy of these models has been hampered by the lack of detailed damage data. This research work has three aims: (1) to evaluate the predictive accuracy of fragility-curve based models at the census tract level using a comprehensive damage dataset for Harris County residences collected after Hurricane Ike (2008), (2) to demonstrate the need for geographically refined models of wind damage risk at spatial scales of one-kilometer square blocks and to analyze the performance of fragility-curve models at that level, and (3) to explore the sources of errors made by fragility-curve based models at the census tract and one-kilometer square block level using a statistical machine learning model constructed from twenty-one potential explanatory variables. Our results provide new insights for building the next generation of fragility-curve models for accurately predicting hurricane wind damage risk to residential structures at the spatial scale of one-kilometer square blocks.