Methods, computer program products, and systems are presented for service cost prediction using machine learning. A claims risk model is formed for predicting a per member per month cost based on variables trained in early prediction models and variables of claim record data from data sources. A training dataset for the claims risk model is modified based on a preconfigured cap value and two distinctive datasets are generated, which trains the claims risk model into a capped submodel and an outlier submodel, respectively. Each submodel makes prediction of the per member per month cost and a slice model interpolates a cost predicted by the capped submodel and another predicted by the outlier submodel. The splice model redistributes the outcomes of the capped submodel and the outlier submodel.