Chronic kidney disease (CKD) is an increasingly common public health issue associated with substantial morbidity and mortality. Risk prediction models provide a useful clinical and research framework for forecasting the probability of adverse events and stratifying patients with CKD according to risk; however, accurate absolute risk prediction requires careful model specification. Competing events that preclude the event of interest (for example, death in studies of end-stage renal disease) must be taken into account. Functional forms of predictor variables and underlying effect modification must be accurately specified; nonlinearity and possible interactions should be evaluated. The potential effect of measurement error should also be considered. Misspecification of any of these components can dramatically affect absolute risk prediction. Evaluation of prognostic models should encompass not only traditional tests of calibration and discrimination, such as the Hosmer-Lemeshow test of 'goodness of fit' and the area under the receiver operating curve, but also newer metrics, such as risk reclassification tables and net reclassification indices. The latter two tests are particularly useful when considering the addition of novel predictors to established models. Finally, models of absolute risk prediction should be internally and externally validated as they typically generalize only to populations with similar baseline characteristics and rates of competing events.