Detecting fraud is a problem characterized by class-imbalance. The events of interest, the true fraudulent cases, are outnumbered by the legitimate ones. Rare outcomes are hard to predict accurately, and standard classifiers, designed to maximize overall prediction accuracy, tend to produce rules ignoring the minority class. Further, the information on each case can be extensive, adding high-dimensionality to the problem of imbalance. This thesis explores three boosting algorithms proposed by Blagus and Lusa (2017) for high-dimensional prediction of rare events: Bal-St-GrBoost, DS-St-GrBoost and StEnsemble. These algorithms confront the imbalance problem by undersampling the majority class, and the high-dimensionality by subsampling. While the authors were motivated by clinical research, here, the three modified boosting algorithms are applied to data that more representative of the fraud detection problem. Their performance is compared to the standard stochastic boosting algorithm (St-GrBoost), using both simulated and real data on VAT-fraud, with varying degree of imbalance and high-dimensionality. In both cases, the covariates include a mixture of continuous and discrete variables. The Gaussian copula is used to simulate data with different marginal distributions. The best algorithm for fraud detection in the high-dimensional setting, consequently, depends on what need the classifier should fulfil. If classification is of primary interest and identifying the most fraudulent cases important, then St-Ensemble is the best choice. When an accurate ranking of new cases has the highest priority, then better results are obtained using the standard procedure, St-GrBoost.