This study contributes to the growing research that uses the news-based measure of U.S. economic policy uncertainty (EPU) suggested in Baker et al. (2016) to predict economic variables out-of-sample. Using simple predictive regressions à la Goyal and Welch (2008), we evaluate the predictive power afforded by various nonlinear transformations of the U.S. EPU measure suggested in Baker et al. (2016) to predict excess returns on the S&P 500 index one-month ahead. Using data from 1985m1 through 2020m12, we find that not all EPU movements are equally important for obtaining point prediction improvements relative to the historical average benchmark at the population level. Particularly, we document that the one-year net EPU increase, defined as EPU increases beyond the peak over the last year, otherwise zero delivers the most consistent pattern of prediction improvements relative to the benchmark. Conversely, other nonlinear specifications as well as the linear models using log-EPU and the first difference of log-EPU do not deliver the same performance. Overall, this study documents that the predictive impact of the U.S. EPU index suggested in Baker et al. (2016) on equity premium is nonlinear in that EPU increases matter only to the extent that they exceed the maximum value over the last twelve months. In other words, there is evidence of threshold nonlinearity. The statistical predictive power afforded by the one-year net EPU increase also translates into economic gains.