Student (educational) loans are highly vulnerable to default risk and thus guaranteed by governments. We show that collateral-free educational loans are a case for the application of Machine Learning models to predict default factors with greater accuracy, helping banks in risk management and the government in designing economic policies of interest suspension and credit guarantees. We argue that heterogeneous ensembles constructed using stacking or a Hill Climb Ensemble approach are most suited for imbalanced data set since the interaction between diverse features would create non-linearities that are impossible to model using a single algorithm. Borrower/student social background emerges as an important feature explaining the loan defaults, warranting a correction in the public policy in designing the educational loan schemes for under privileged borrowers. Our paper also shows that Machine learning models are not systematically biased against underprivileged borrowers and do not lead banks to refuse credit. Ours is the first study to apply Statistical, Machine learning and Deep learning Models on a data set of student loans.