Ensemble-of-trees algorithms have emerged to the forefront of machine learning due to their ability to generate high forecasting accuracy for a wide array of regression and classification problems. Classic ensemble methodologies such as random forests (RF) and stochastic gradient boosting (SGB) rely on algorithmic procedures to generate fits to data. In contrast, more recent ensemble techniques such as Bayesian Additive Regression Trees (BART) and Dynamic Trees (DT) focus on an underlying Bayesian probability model to generate the fits. These new probability model-based approaches show much promise versus their algorithmic counterparts, but also offer substantial room for improvement. The first part of this thesis focuses on methodological advances for ensemble-of-trees techniques with an emphasis on the more recent Bayesian approaches. In particular, we focus on extensions of BART in four distinct ways. First, we develop a more robust implementation of BART for both research and application. We then develop a principled approach to variable selection for BART as well as the ability to naturally incorporate prior information on important covariates into the algorithm. Next, we propose a method for handling missing data that relies on the recursive structure of decision trees and does not require imputation. Last, we relax the assumption of homoskedasticity in the BART model to allow for parametric modeling of heteroskedasticity. The second part of this thesis returns to the classic algorithmic approaches in the context of classification problems with asymmetric costs of forecasting errors. First we consider the performance of RF and SGB more broadly and demonstrate its superiority to logistic regression for applications in criminology with asymmetric costs. Next, we use RF to forecast unplanned hospital readmissions upon patient discharge with asymmetric costs taken into account. Finally, we explore the construction of stable decision trees for forecasts of violence during probation hearings in court systems.