There has been a surge in models for clinical diagnosis and prognosis. In a systematic review of clinical prediction models, those using machine learning did not show systematic performance benefit over statistical models. Although diagnosis and prognosis use the same type of models and the same criteria to assess model performance, there is an essential difference between both. Diagnosis refers to a person’s disease condition at the time of observation.  Often there is an unambiguous true value, as reflected in the term gold standard (or ground truth). Prognosis concerns the future; the time component plays an essential role. Time can add additional characteristics to the data, such as censoring and the presence of competing risks. There is an intrinsic limit to prognosis in systems that have nonlinear dynamic components, which can induce chaotic behaviour. Under chaotic behaviour, the value of the outcome can be highly sensitive to small changes in conditions that are part of the mechanisms leading to the outcome. It is impossible to measure all these conditions. This component of unexplained variation cannot be eliminated and therefore enlarging the data set or building more flexible models may not show improvement in performance.