Tree-Based Models

Tree-based models have increased in popularity as an alternative to traditional regression models. Building such models involves algorithms that repeatedly partition the region of the explanatory variables to create non-overlapping nodes for predictions. Tree-based models divide the predictor space into a number of non-overlapping regions and use the mean or the mode of the region as the prediction. As a result, tree-based models can be used to predict continuous or categorical responses. This chapter utilizes representative policies selected by Latin hypercube sampling to train the model. It introduces tree-based models and illustrated their use in predicting fair market values for a large portfolio of variable annuity contracts. Tree-based models have some advantages over other models.