Studies evaluating the effects of traffic safety measures are often done for the purpose of predicting the effects of future applications of the measures. The predictive value of evaluation studies is unknown. Some general arguments for and against attributing a general predictive value to the results of evaluation studies are discussed. Predictability is shown to depend on many factors. Meta-analyses of evidence from evaluation studies can be used as a basis for testing the predictive performance of such studies. The predictive performance of studies that have evaluated the safety effects of road lighting and traffic separation is tested. Predictive performance is found to depend mainly on whether the results of evaluation studies are stable over time or exhibit a trend. In the latter case, predictions based on evidence accumulated before the trend became apparent can be very erroneous. It is shown that increasing the amount of evidence that predictions are based on does not necessarily make the predictions more accurate. More research does not always improve predictive performance.