In the predictive analytics world, ensemble modeling strategy is often pursued to improve model accuracy and robustness. Ensemble modeling is the process of creating multiple models and incorporating them into a single scoring algorithm. The value of ensemble modeling for enhancing predictive accuracy and increasing model stability is widely recognized. However, it is an art that is not easily mastered. Through our predictive modeling practice within the telecommunication industry, we have found that in general, heterogenous ensemble modeling produces better results than homogeneous built-in ensembles. We have also found that depending on the size of the modeling target, homogeneous ensemble modeling may not always be the best choice, compared to a single model. In this paper, we present empirical ensemble strategies and suggest best practices pertaining to modeling techniques and target sizes.