Research on predictive modeling methods has been remarkable in recent years,  and the models that we actuaries use are advancing day by day. Actuaries are interested not only in prediction  accuracy but also in evaluating prediction errors and interpreting what the source of the error is. A lot of previous research about error decomposition into parameter error and process error has been conducted for traditional actuarial modeling, such as GLM. However, there is not enough research about error decomposition for various predictive modeling method including machine learning. In this study, we, Data Science Related Basic Research Working Group of The Institute of Actuaries of Japan (IAJ), propose a new framework decomposing prediction errors into process errors, parameter errors, and other errors, which is widely applicable to lots of predictive modeling methods.