Insurance loss data often cannot be well modeled by a single distribution. Mixture of models are often applied in insurance loss modeling. The Expectation-Maximization (EM) algorithm is used for parameter estimation in mixture of models. Feature engineering and variable selection are challenging for mixture of models due to several component models involving. Overfitting is also a concern when predicting future loss. To address those issues, we propose an Expectation-Boosting (EB) algorithm, which replaces the maximization step in the EM algorithm by a gradient boosting decision tree. The boosting is overfitting-sensitive, and it performs automated feature engineering, model fitting and variable selection simultaneously. The EB algorithm fully explores the predictive power of covariate space. We illustrate those advantages using two simulated data and a real insurance loss data.