Machine learning (ML) based models have demonstrated a great potential for streamflow prediction. ML based procedures are relatively easier to apply and are less computationally demanding, especially for applications at regional scales, than traditional physics-based models. Thus, their application for hydrologic predictions have attracted a lot of interest from stakeholders in academia, industry and federal agencies. For streamflow prediction, these models perform very well at capturing streamflow variability. However, they generally fail to accurately predict extreme values (i.e. peak flow) of flood events, which are important to be considered in flood design and for flood warning purposes. To address this, in this work we examine an event-based predictive framework which is solely focused on peak flow prediction and takes into account the characteristics of the flood triggering precipitation, the catchment and antecedent wetness conditions. We compare different ML-based approaches (among them, decision trees and deep neural networks) and demonstrate their relative strengths and limitations. We also use some of these developed models to examine the relative predictive importance of the different variables and its dependence to the hydroclimatic region. Our analysis is based on the CAMeLs Dataset, which provides varying hydrometeorological and land surface characteristics for over 600 catchments across the contiguous US. We carry out various experiments to demonstrate the transferability of the proposed model with particular focus on its accuracy for predicting peak flows in ungauged catchments (i.e. catchments that have not been included in the training dataset). Results show that an event-based ML model provides a predictive tool that can be used in complement to other models, focused on continuous hydrologic simulations, to improve prediction of flood peak magnitudes.