SUMMARYWeb Service Choreography Description Language (WS-CDL) is a mainstream standard for the description of peer-to-peer collaborations for the participants of service composition. To predict the dependability of composite service processes specified in WS-CDL allows service designers and user to decided whether the process meets the nonfunction requirements of trustworthiness, and to choose the process with better dependability to those with identical function. Unfortunately, very limited research attention is paid to dependability of WS-CDL. In this paper, we propose a model-driven approach for dependability prediction of composite service built on WS-CDL. The main innovation of this research includes a complete translation from WS-CDL to generalized stochastic Petri nets (GSPN) and a dependability (using processnormal-completion-probability as the metric of dependability of service composition) calculation algorithm based on GSPN. We also validate the accuracy of the approach in the experimental study by showing 95% confidence intervals obtained from experimental dependability results that cover the corresponding theoretical prediction values.