In this study, a Bayesian Network (BN) model of serial arson cases was constructed, and the validity of the model in offender profiling was investigated: the accuracy of the constructed model to infer ex-convict status of theft and employment of serial arsonists. BN is a probability model that describes causal structure of events as chain networks of conditional probability; which is capable of predicting the possibility of uncertain events. The detailed procedure and results in this study were the following. Firstly, a BN model was constructed from the training data (518 data of serial arsonists and cases) using a K2 search-algorithm and information criteria, MDL (minimum description length). The constructed model indicated that ex-convict status of theft was related to reporting a fire after arson and crime scenes such as parking lots. Moreover, it suggested that employment was related to car use. Secondly, the validity of the model was examined using the validation data (assumed 30 unsolved arsons). The results of model estimation showed that the accuracy of inferring an ex-convict status of theft was 80%, but the accuracy of inferring employment was 50%. To make more exact inferences, a database needs to be constructed from accurate information and tested repeatedly using various search-algorithm and information criteria.