As the volume of police recording data continue to improve so fast that the interest in predicting future crime appear. This had led to a growing interest, particularly amongst criminologists, with prediction, however the preoccupation with methods such as data mining or machine learning has been almost exclusively focused on the dynamics of space-time forecasting as opposed to techniques that might better identify likely repeat offenders. This paper seeks to address this gap in the literature by taking advantage of a sample of offender data provided by the Beijing Municipal Public Security Bureau (BMPSB) to test the potential utility of a machine learning approach for identifying repeat offenders. Two separate crime categories, pick-pocketing theft and residential burglary, and five variables: age, gender, education level, registration place and drug history, are processed and a binary logistic regression approach is used to test their suitability for accurately predicting repeat offenders. Results from this simple experiment suggest these five, simple individual-level attributes accurately predict roughly 70 per cent of repeat offenders. Implications for policy and practice are discussed.