Revisiting test score forecasts and teacher value-added

We develop a framework for comparing forecasts in teacher value-added models. We apply it to interpret the teacher-switching test of Chetty, Friedman, and Rockoff (2014), showing that the teacher-switching coefficient cannot be mechanically interpreted as a test of forecast unbiasedness: the correct null may differ from one, and a coefficient of one is insufficient for forecast unbiasedness. We also analyze forecast choice for ranking teachers and estimating teachers' effects on long-term outcomes. Although forecasts including current-year data may contain contemporaneous classroom shocks, they contain more teacher signal than leave-year-out forecasts, which improves rankings and may increase precision in long-term-outcome regressions.