Optimal prediction threshold and spatial characteristics of burglary crime across cities

Burglary, as a prevalent and detrimental crime type, poses a major threat to public safety and property security. Accurate prediction of burglary occurrence is therefore critical. Although deep learning has achieved notable progress in crime prediction, the influence of varying urban spatial characteristics on predictive performance and resource efficiency remains underexplored. This study analyzes burglary prediction in eight representative cities from China, the United States, and Canada, uniformly employing the model based on convolutional neural networks and long short-term memory networks (CNN–LSTM) under a consistent spatio–temporal scale. A comprehensive evaluation framework based on precision, hit rate, prediction accuracy index (PAI), and prediction effectiveness index (PEI) was established, focussing on the validity of PEI and variations in optimal prediction thresholds across cities. Furthermore, standard deviation ellipses, Moran index, and kernel density analysis were applied to quantify spatial characteristics and explore their associations with PEI and resource allocation. The results indicate that cities with more concentrated spatial distributions and stronger spatial autocorrelation exhibit superior predictive efficiency and resource utilization. This study enriches the analytical scope of crime prediction efficiency and supports a shift from ‘accuracy–oriented’ to ‘efficiency–oriented’ modelling for intelligent allocation of urban public safety resources.