Prognostic models for disease occurrence, tumor progression and survival are abundant for most types of cancers. Physicians and cancer patients are utilizing these models to make informed treatment decisions and corresponding arrangements. However, not all cancer prognostic models are built and validated rigorously. Some are more useful and reliable than others. In this review, we briefly introduce some popular machine learning methods for constructing cancer prognostic models, and discuss pros and cons of each. We also introduce the commonly used discrimination and calibration metrics for assessing predictive performance and validating the prognostic models. In the end, we outline several challenges of using prognostic models in the real world for clinical decision-making support, and propose related suggestions.