We estimate Boosted Regression Trees (BRT) on a sample of monthly data that extends back to 1889 to recover the predictive value of disaggregated news-based uncertainty indexes for U.S recessions. We control for widely-studied standard predictors and use out-of-sample metrics to assess forecast performance. We find that war-related uncertainty is among the top five predictors of recessions at three different forecast horizons (3, 6, and 12 months). The predictive value of war-related uncertainty has fallen in the second half of the 20th century. Uncertainty regarding the state of securities markets has gained in relative importance. The probability of a recession is a nonlinear function of war-related and securities-markets uncertainty. Receiver-operating characteristic curves show that uncertainty improves out-of-sample forecast performance at the longer forecast horizons. A dynamic version of the BRT approach sheds light on the importance of various lags of government-related uncertainty for recession forecasting at the long forecast horizon.