The recent emergence of cryptocurrencies has added another layer of complexity in the fight towards financial crime. Cryptocurrencies require no central authority and offer pseudo-anonymity to its users, allowing criminals to disguise themselves among legitimate users. On the other hand, the openness of data fuels the investigator's toolkit to conduct forensic examinations. This study focuses on the detection of illicit activities (e.g., scams, financing terrorism, and Ponzi schemes) on cryptocurrency infrastructures, both at an account and transaction level. Previous work has identified that class imbalance and the dynamic environment created by the evolving techniques deployed by criminals to avoid detection are widespread in this domain. In our study, we propose Adaptive Stacked eXtreme Gradient Boosting (ASXGB), an adaptation of eXtreme Gradient Boosting (XGBoost), to better handle dynamic environments and present a comparative analysis of various offline decision tree-based ensembles and heuristic-based data-sampling techniques. Our results show that: (i) offline decision treebased gradient boosting algorithms outperform state-of-the-art Random Forest (RF) results at both an account and transaction level, (ii) the data-sampling approach NCL-SMOTE further improves recall at a transaction level, and (iii) our proposed ASXGB successfully reduced the impact of concept drift while further improving recall at a transaction level.