During the last few decades, emerging machine learning techniques such as ensemble methods have become increasingly popular in every domain as they can handle a lot of uncertainty present in the dataset. The present study focuses on the prediction of resilient modulus (MR) of soil from a large experimental dataset using an ensemble method, i.e., random forest regressor (parallel ensemble) and gradient boosting (sequential ensemble), and compares the results with the most widely used multi-linear regression (MLR) and decision tree regressor (DTR) methods. The published laboratory test data on fine-grained soils from Ohio, USA, was used to examine the models. The percent fines, Atterberg limits, optimum moisture content, degree of saturation, unconfined compressive strength, and stress state of the soil sample were considered as influencing input variables in developing the models and predicting the resilient modulus of the sample. The results showed that an MLR model did not perform well in terms of accuracy as most of the input parameters do not hold a linear relationship with MR, whereas DTR suffers from high variance. On the other hand, the ensemble-based random forest regressor and eXtreme Gradient Boosting methods performed significantly well in terms of improved prediction accuracy and reduced variance as compared to the base models. Furthermore, sensitivity analysis was performed for the best-fitted model and results showed that the degree of saturation was the most influencing parameter affecting the resilient modulus of soil, followed by the stress state and plasticity index of the soil sample.