This study evaluates techniques to identify high-quality teachers. Since tenure restricts dismissals of experienced teachers, schools must predict productivity and dismiss those expected to perform ineffectively prior to tenure receipt. Many states rely on evaluation scores to guide these personnel decisions without considering other dimensions of teacher performance. I use predictive models to rank teachers based on expected value-added and summative ratings. I then simulate revised personnel decisions and test for changes in average retained teacher performance. In this exercise, I adjust two factors that impact the quality of the predictions: the number of predictors and the length of the pretenure period. Both factors impact the precision of the predictions, though extended pretenure periods also negatively impact selection into teaching. I estimate optimal weights on each performance measure to maximize measures of teacher quality using a range of utility parameters. These improvements are a product of using additional information (value-added) rather than advanced algorithms, as OLS regressions and advanced machine learning techniques produce similar gains. In comparison, prediction models that extend the pretenure period beyond one year do not provide enough additional information to significantly improve average retained teacher performance unless dismissal rates increase dramatically.