Preventative pipe maintenance is a key activity in pipe asset management. Central to such activity is a precise pipe failure (burst/leakage) prediction. Here a statistical pipe failure prediction approach is proposed based on the massive data including pipe physical property, environmental factor, operational condition, historical failure records, and etc. Considering the biased training cases, survival analysis model is adopted to avoid over-fitting. The effectiveness of such an approach over several machine learning algorithms is proven in an Asia city with 4 pipe physical indicators (material type, age, diameter, and length) considered over a given region in the past 10 years. Compared with a heuristic approach, there is 58 times improvement in targeting precision. It also shows that there still a significant improvement opportunity by incorporating more factors. © 2011 IEEE.