Recent evidence suggests that many economic time series are subject to structural breaks. In the presence of breaks, including historical data prior to the most recent break to estimate a forecasting model will lead to prediction errors that are biased but also may have a smaller variance. This paper examines the trade-off between the bias and variance of forecast errors and proposes a new set of reversed Cusum procedures to determine the window size that minimizes mean squared forecast error. This window size varies over time and depends on the size of the break, the distance to the break and the squared correlation coefficient between predicted and realized values. The forecasting performances of several procedures for determination of window size are compared in a simulation experiment and in a recursive prediction exercise using data on US stock returns. We find evidence that out-of-sample forecasting performance can be improved by explicitly accounting for breaks and adopting the proposed method for optimally determining the window size