The 2007 financial crisis highlighted the severe risks posed by counterparty defaults in financial markets. Assessing and addressing counterparty credit risk has consequently been a focal point of new regulations introduced in the wake of the crisis. The Central Clearing Counterparty (CCP) is at the heart of the solution, an entity dedicated to managing and mitigating counterparty risk in a market. CPPs manage risk by collecting collateral, referred to as margin, from the participants trading on the market. Appropriately sizing the margin is of utmost importance for the CCP to maintain the integrity of its operation and, by extension, protect the participants in the market. Most contemporary margin methodologies require significant resources which precludes frequent margin updates. In light of this issue, our work examines the capability of replicating the popular margin methodology Historical Simulation Value at Risk using machine-learning-based methods envisioning that an adequate such model could be used as a complement to the traditional model, providing real-time margin estimations. The experiment concerns portfolios containing stocks, bonds, and options and uses static market data and scenarios. We conclude that neither of the ensemble methods are sufficiently accurate, while both of the neural network-based models show moderate promise, warranting further development.